1

Volume

FOXES TEAM

Reference for Xnumbers.xla

Numeric Calculus
in Excel

REFERENCE FOR XNUMBERS.XLA

Numeric Calculus
in EXCEL

Oct 2007

Xnumbers Tutorial

Index
About this Tutorial ..................................................................................................................... 8
Array functions........................................................................................................................... 9
What is an array-function? ....................................................................................................... 9
How to insert an array function ................................................................................................ 9
How to get the help on line...................................................................................................... 12
Xnumbers installation.............................................................................................................. 13
How to install .......................................................................................................................... 13
How to uninstall ...................................................................................................................... 14
Installation troubles ................................................................................................................ 14
Multiprecision Floating Point Arithmetic ............................................................................... 15
Why using extended precision numbers? .............................................................................. 15
Multiprecision methods........................................................................................................... 17
How to store long number ...................................................................................................... 17
Functions .................................................................................................................................. 19
General Description................................................................................................................ 19
Using Xnumbers functions ..................................................................................................... 19
Using extended numbers in Excel.......................................................................................... 20
Functions Handbook .............................................................................................................. 21
Precision................................................................................................................................. 22
Formatting Result ................................................................................................................... 22
Arithmetic Functions............................................................................................................... 23
Addition............................................................................................................................................. 23
Subtraction........................................................................................................................................ 23
Multiplication ..................................................................................................................................... 24
Division ............................................................................................................................................. 24
Inverse .............................................................................................................................................. 24
Integer Division ................................................................................................................................. 24
Integer Remainder ............................................................................................................................ 25
Sum .................................................................................................................................................. 27
Product ............................................................................................................................................. 27
Raise to power.................................................................................................................................. 27
Square Root...................................................................................................................................... 28
Nth- Root ........................................................................................................................................... 28
Absolute............................................................................................................................................ 28
Change sign...................................................................................................................................... 28
Integer part ....................................................................................................................................... 29
Decimal part...................................................................................................................................... 29
Truncating......................................................................................................................................... 29
Rounding .......................................................................................................................................... 29
Relative Rounding ............................................................................................................................ 30

Extended Numbers manipulation ........................................................................................... 31

Digits count ....................................................................................................................................... 31
Significant Digits count...................................................................................................................... 31
Numbers comparison........................................................................................................................ 31
Extended number check ................................................................................................................... 32
Format extended number.................................................................................................................. 32
Check digits ...................................................................................................................................... 32
SortRange......................................................................................................................................... 33
Digits sum ......................................................................................................................................... 33
Vector flip.......................................................................................................................................... 33
Scientific format ................................................................................................................................ 34
Split scientific format ......................................................................................................................... 34
Converting multiprecision into double ............................................................................................... 34
Macros X-Edit ................................................................................................................................... 35
Macro X-Converter ........................................................................................................................... 36

Statistical Functions ............................................................................................................... 39

Factorial ............................................................................................................................................ 39
Factorial with double-step ................................................................................................................. 39

Xnumbers Tutorial
Combinations.................................................................................................................................... 39
Permutations..................................................................................................................................... 40
Arithmetic Mean................................................................................................................................ 40
Geometric Mean ............................................................................................................................... 40
Quadratic Mean ................................................................................................................................ 40
Standard Deviation ........................................................................................................................... 40
Variance............................................................................................................................................ 41
Probability distributions ..................................................................................................................... 41
Univariate Statistic ............................................................................................................................ 46
Linear Regression Coefficients ......................................................................................................... 47
Linear Regression - Standard Deviation of Estimates ...................................................................... 49
Linear Regression Formulas............................................................................................................. 50
Linear Regression Covariance Matrix............................................................................................... 51
Linear Regression Statistics ............................................................................................................. 52
Linear Regression Evaluation ........................................................................................................... 53
Polynomial Regression - Coefficient ................................................................................................. 54
Polynomial Regression - Standard Deviation of Estimates ............................................................... 54
Polynomial Regression Statistics...................................................................................................... 54
Macro - Regression .......................................................................................................................... 55
Linear Regression with Robust Method ............................................................................................ 59
Linear Regression Min-Max .............................................................................................................. 60
NIST Certification Test...................................................................................................................... 61

Transcendental Functions ...................................................................................................... 62

Logarithm natural (Napiers) ............................................................................................................. 62
Logarithm for any base ..................................................................................................................... 62
Exponential ....................................................................................................................................... 62
Exponential for any base .................................................................................................................. 62
Constant e ....................................................................................................................................... 63
Constant Ln(2) .................................................................................................................................. 63
Constant Ln(10) ................................................................................................................................ 63
Hyperbolic Sine................................................................................................................................. 63
Hyperbolic ArSine ............................................................................................................................. 63
Hyperbolic Cosine............................................................................................................................. 64
Hyperbolic ArCosine ......................................................................................................................... 64
Hyperbolic Tangent........................................................................................................................... 64
Hyperbolic ArTangent ....................................................................................................................... 64
Euler constant ................................................................................................................................ 64

Trigonometric Functions......................................................................................................... 65
Sin .................................................................................................................................................... 65
Cos ................................................................................................................................................... 65
Tan ................................................................................................................................................... 66
Arcsine.............................................................................................................................................. 66
Arccosine .......................................................................................................................................... 66
Arctan ............................................................................................................................................... 66
Constant ........................................................................................................................................ 66
Complement of right angle................................................................................................................ 67

Polynomial Rootfinder ............................................................................................................ 68

Input parameters............................................................................................................................... 69
Output ............................................................................................................................................... 70
How to use rootfinder macros ........................................................................................................... 71
Integer roots...................................................................................................................................... 74
Multiple roots .................................................................................................................................... 76

Polynomial Functions ............................................................................................................. 77

Polynomial evaluation ....................................................................................................................... 77
Polynomial derivatives ...................................................................................................................... 78
Polynomial coefficients ..................................................................................................................... 79
Polynomial writing............................................................................................................................. 80
Polynomial addition........................................................................................................................... 80
Polynomial multiplication................................................................................................................... 80
Polynomial subtraction...................................................................................................................... 81
Polynomial division quotient.............................................................................................................. 81
Polynomial division remainder .......................................................................................................... 81
Polynomial shift................................................................................................................................. 85
Polynomial center ............................................................................................................................. 85
Polynomial roots radius..................................................................................................................... 86
Polynomial building from roots.......................................................................................................... 87
Polynomial building with multi-precision............................................................................................ 89
Polynomial Solving ........................................................................................................................... 90
Integer polynomial ............................................................................................................................ 90

Xnumbers Tutorial
Polynomial System of 2nd degree...................................................................................................... 91
Bivariate Polynomial ......................................................................................................................... 92
Orthogonal Polynomials evaluation .................................................................................................. 97
Weight of Orhogonal Polynomials..................................................................................................... 99
Zeros of Orthogonal Polynomials...................................................................................................... 99
Coefficients of Orthogonal Polynomials .......................................................................................... 100

Complex Arithmetic and Functions ...................................................................................... 101

How to insert a complex number .................................................................................................... 101
Complex Addition............................................................................................................................ 102
Complex Subtraction ...................................................................................................................... 102
Complex Multiplication .................................................................................................................... 102
Complex Division ............................................................................................................................ 102
Polar Conversion ............................................................................................................................ 103
Rectangular Conversion ................................................................................................................. 103
Complex absolute ........................................................................................................................... 103
Complex power............................................................................................................................... 104
Complex Roots ............................................................................................................................... 104
Complex Log................................................................................................................................... 105
Complex Exp .................................................................................................................................. 105
Complex inverse ............................................................................................................................. 105
Complex negative ........................................................................................................................... 105
Complex conjugate ......................................................................................................................... 105
Complex Sin ................................................................................................................................... 106
Complex Cos .................................................................................................................................. 106
Complex Tangent ........................................................................................................................... 106
Complex ArcCos............................................................................................................................. 106
Complex ArcSin .............................................................................................................................. 106
Complex ArcTan ............................................................................................................................. 106
Complex Hyperbolic Sine................................................................................................................ 106
Complex Hyperbolic Cosine............................................................................................................ 107
Complex Hyperbolic Tan................................................................................................................. 107
Complex Inverse Hyperbolic Cos.................................................................................................... 107
Complex Inverse Hyperbolic Sin..................................................................................................... 107
Complex Inverse Hyperbolic Tan.................................................................................................... 107
Complex digamma .......................................................................................................................... 107
Complex Exponential Integral ......................................................................................................... 107
Complex Error Function .................................................................................................................. 108
Complex Complementary Error Function........................................................................................ 108
Complex Gamma Function ............................................................................................................. 108
Complex Logarithm Gamma Function ............................................................................................ 108
Complex Zeta Function................................................................................................................... 108
Complex Quadratic Equation .......................................................................................................... 109

Number Theory .................................................................................................................... 110

Maximum Common Divisor............................................................................................................. 110
Minimum Common Multiple ............................................................................................................ 110
Rational Fraction approximation ..................................................................................................... 111
Continued Fraction ......................................................................................................................... 112
Continued Fraction of Square Root ................................................................................................ 113
Check Prime ................................................................................................................................... 113
Next Prime ...................................................................................................................................... 113
Modular Addition............................................................................................................................. 114
Modular Subtraction........................................................................................................................ 114
Modular Multiplication ..................................................................................................................... 114
Modular Division ............................................................................................................................. 114
Modular Power................................................................................................................................ 114
Perfect Square................................................................................................................................ 116
Check odd/even.............................................................................................................................. 116
Check Integer ................................................................................................................................. 116
Macro - Factorize ............................................................................................................................ 117
Batch Factorization with Msieve ..................................................................................................... 118
Factorization function...................................................................................................................... 120
Macro - Prime Numbers Generator................................................................................................. 121
Prime Test ...................................................................................................................................... 121
Diophantine Equation ..................................................................................................................... 123
Brouncker-Pell Equation ................................................................................................................. 124
Euler's Totient function ................................................................................................................... 125
Integer relation................................................................................................................................ 126
Macro Integer Relation Finder ........................................................................................................ 130

Linear Algebra Functions ..................................................................................................... 133

Matrix Addition ................................................................................................................................ 133

Xnumbers Tutorial
Matrix Subtraction........................................................................................................................... 133
Matrix Multiplication ........................................................................................................................ 133
Matrix Inverse ................................................................................................................................. 133
Matrix Determinant ......................................................................................................................... 133
Matrix Modulus ............................................................................................................................... 134
Scalar Product ................................................................................................................................ 134
Similarity Transform ........................................................................................................................ 134
Matrix Power................................................................................................................................... 134
Matrix LU decomposition ................................................................................................................ 135
Matrix LLT decomposition ............................................................................................................... 135
Vector Product ................................................................................................................................ 136
Solve Linear Equation System........................................................................................................ 136
Square Delta Extrapolation ............................................................................................................. 137
Macro for Multiprecision Matrix Operations..................................................................................... 139

Integrals & Series ................................................................................................................. 141

Discrete Fourier Transform ............................................................................................................. 141
Discrete Fourier Inverse Transform ................................................................................................ 143
Discrete Fourier Spectrum .............................................................................................................. 144
Inverse Discrete Fourier Spectrum ................................................................................................. 144
2D Discrete Fourier Transform ....................................................................................................... 145
2D Inverse Discrete Fourier Transform........................................................................................... 145
Macro DFT (Discrete Fourier Transform)........................................................................................ 146
Macro Sampler ............................................................................................................................... 149
Integral function .............................................................................................................................. 151
Zeros of integral function ................................................................................................................ 154
Function Integration (Romberg method) ......................................................................................... 155
Function Integration (Double Exponential method)......................................................................... 156
Function Integration (mixed method) .............................................................................................. 158
Complex Function Integration (Romberg method) .......................................................................... 160
Data Integration (Newton-Cotes) .................................................................................................... 161
Function Integration (Newton-Cotes) .............................................................................................. 163
Integration: symbolic and numeric approaches............................................................................... 165
Integration of oscillating functions (Filon formulas) ......................................................................... 166
Integration of oscillating functions (Fourier transform) .................................................................... 168
Infinite Integration of oscillating functions ....................................................................................... 169
Double Integral ............................................................................................................................... 172
Macro for Double Integration .......................................................................................................... 172
Macro for Triple Integration............................................................................................................. 175
Double integration function ............................................................................................................. 178
Infinite integral ................................................................................................................................ 181
Double Data integration .................................................................................................................. 183
Series Evaluation............................................................................................................................ 186
Series acceleration with ^2 ........................................................................................................... 187
Complex Series Evaluation............................................................................................................. 188
Double Series ................................................................................................................................. 189
Trigonometric series ....................................................................................................................... 190
Trigonometric double serie ............................................................................................................. 192
Discrete Convolution....................................................................................................................... 193

Interpolation.......................................................................................................................... 195
Polynomial interpolation.................................................................................................................. 195
Interpolation schemas..................................................................................................................... 197
Interpolation with continued fraction ............................................................................................... 199
Interpolation with Cubic Spline........................................................................................................ 201
Cubic Spline 2nd derivatives........................................................................................................... 201
Cubic Spline Coefficients ................................................................................................................ 203
2D Mesh Interpolation..................................................................................................................... 204
Macro Mesh Fill .............................................................................................................................. 205

Differential Equations ........................................................................................................... 211

ODE Runge-Kutta 4........................................................................................................................ 211
ODE Multi-Steps ............................................................................................................................. 215
Multi-step coefficents tables............................................................................................................ 216
Predictor- Corrector ........................................................................................................................ 218
PECE algorithm of 2nd order ........................................................................................................... 218
ODE Predictor-Corrector 4.............................................................................................................. 221
ODE Implicit Predictor-Corrector .................................................................................................... 223
Differential Systems ........................................................................................................................ 226
OD Linear System .......................................................................................................................... 226
High order linear ODE..................................................................................................................... 227
ODE for integral function solving .................................................................................................... 230

Xnumbers Tutorial
Macro ODE Solver .......................................................................................................................... 232
Macro ODE - Slope Grid ................................................................................................................. 233

Nonlinear Equations ............................................................................................................. 235

Bisection ......................................................................................................................................... 235
Secant............................................................................................................................................. 236

Derivatives............................................................................................................................ 239
First Derivative................................................................................................................................ 239
Second Derivative........................................................................................................................... 240
Gradient .......................................................................................................................................... 240
Jacobian matrix............................................................................................................................... 241
Hessian matrix ................................................................................................................................ 241
Non-linear equation solving with derivatives ................................................................................... 243

Conversions ......................................................................................................................... 245

Decibel............................................................................................................................................ 245
Base conversion ............................................................................................................................. 245
Multiprecision Base Conversion...................................................................................................... 246
Log Relative Error........................................................................................................................... 247

Special Functions ................................................................................................................. 249

Error Function Erf(x) ...................................................................................................................... 249
Exponential integral Ei(x) ................................................................................................................ 249
Exponential integral En(x)............................................................................................................... 249
Euler-Mascheroni Constant ......................................................................................................... 250
Gamma function (x) ...................................................................................................................... 250
Log Gamma function ...................................................................................................................... 251
Gamma quotient ............................................................................................................................. 251
Gamma F-factor.............................................................................................................................. 252
Digamma function........................................................................................................................... 252
Beta function................................................................................................................................... 252
Incomplete Gamma function ........................................................................................................... 253
Incomplete Beta function ................................................................................................................ 253
Combinations function .................................................................................................................... 253
Bessel functions of integer order ................................................................................................... 254
Cosine Integral Ci(x) ....................................................................................................................... 254
Sine Integral Si(x) ........................................................................................................................... 254
Fresnel sine Integral ....................................................................................................................... 255
Fresnel cosine Integral.................................................................................................................... 255
Fibonacci numbers ......................................................................................................................... 255
Hypergeometric function ................................................................................................................. 256
Zeta function (s) ........................................................................................................................... 256
Airy functions .................................................................................................................................. 257
Elliptic Integrals............................................................................................................................... 257
Kummer confluent hypergeometric functions.................................................................................. 258
Integral of sine-cosine power .......................................................................................................... 259
Spherical Bessel functions of integer order..................................................................................... 260

Formulas Evaluation............................................................................................................. 261

Multiprecision Expression Evaluation ............................................................................................. 261
Complex Expression Evaluation ..................................................................................................... 265
Multiprecision Excel Formula Evaluation ........................................................................................ 266
Math expression strings .................................................................................................................. 268
List of basic functions and operators .............................................................................................. 271

Function Optimization........................................................................................................... 274

Macros for optimization on site ....................................................................................................... 274
Example 1 - Rosenbrock's parabolic valley .................................................................................... 276
Example 2 - Constrained minimization ........................................................................................... 277
Example 3 - Nonlinear Regression with Absolute Sum................................................................... 279
Example 4 - Optimization of Integral function ................................................................................. 280

How to call Xnumbers functions from VBA ......................................................................... 281

References & Resources ....................................................................................................... 284
Analytical index ...................................................................................................................... 287

Xnumbers Tutorial

WHITE PAGE

Xnumbers Tutorial

Chapter

1
About this tutorial
About this Tutorial
This document is the reference guide for all functions
and macros contained in the Xnumbers addin. It is a
printable version of the help-on-line, with a larger
collection of examples.

XNUMBERS.XLA is an Excel addin containing useful functions for numerical

calculus in standard and multiprecision floating point arithmetic up to 200
significant digits.
The main purpose of this document is to be a reference guide for the functions of this package,
showing how to work with multiprecision arithmetic in Excel. The most part of the material
contained in this document comes from the Xnumbers help-on-line. You may print it in order to
have a handle paper manual. Of course it speaks about math and numerical calculus but this is
not a math book. You rarely find here theorems and demonstrations. You can find, on the
contrary, many explaining examples.
Special thanks to everyone that have kindly collaborated.

Leonardo Volpi

Xnumbers Tutorial

Array functions
What is an array-function?
A function that returns multiple values is called "array-function". Xnumbers contains many of
these functions. Each function returning a matrix or a vector is an array functions. Function
performing matrix operations such as inversion, multiplication, sum, etc. are examples of arrayfunctions. Also complex numbers are arrays of two cells. On the contrary, in the real domain,
the logarithm, the exponential, the trigonometric functions, etc. are scalar functions because
they return only one value.
In a worksheet, an array-function always returns a (n x m) rectangular range
of cells. To enter it, you must select this range, enter the function as usually
and and give the keys sequence CTRL+SHIFT+ENTER. Keep down both
keys CTRL and SHIFT (do not care the order) and then press ENTER.

How to insert an array function

The following example explains, step-by-step, how to insert an array-function

The System Solution

Assume to have to solve a 3x3 linear system. The solution is a vector of 3 values.
Ax = b
Where:

4
b = 2
3

1 1 1
A = 1 2 2
1 3 4

The function SYSLIN returns the vector solution x, that is an (3 x 1) array.

To see the three values you must select
before the area where you want to insert
these values.
Now insert the function either from menu or
by the icon

Select the area of the matrix A "A5:C7" and

the constant vector b "E5:E7"

Xnumbers Tutorial
Now - attention! - give the "magic" keys sequence CTRL+SHIFT+ENTER
That is:

Press and keep down the CTRL and SHIFT keys

Press the ENTER key

All the values will fill the cells that you have selected.

Note that Excel shows the function around two braces { }. These symbols mean that the
function return an array (you cannot insert them by hand).
An array function has several constrains. Any cell of the array cannot be modified or deleted.
To modify or delete an array function you must selected before the entire array cells.

Adding two matrices

The CTRL+SHIFT+ENTER rule is valid for any function and/or operation returning a matrix or a
vector
Example - Adding two matrices

1 2 1 0
2 1 + 0 1

We can use directly the addition operator "+". We can do this following these steps.
1)

Enter the matrices into the spreadsheet.

2)

Select the B8:C9 empty cells so that a 2 2 range is highlighted.

3)

Write a formula that adds the two ranges. Either write =B4:C5+E4:F5 Do not press
<Enter>. At this point the spreadsheet should look something like the figure below. Note
that the entire range B8:C9 is selected.

10

Xnumbers Tutorial
4)

Press and hold down <CTRL> + <SHIFT>

5)

Press <ENTER>.

If you have correctly followed the procedure, the spreadsheet should now look something like
this

This trick can also work for matrix subtraction and for the scalar-matrix multiplication, but not for
the matrix-matrix multiplication.
Let's see this example that shows how to calculate the linear combination of two vectors

Functions returning optional values

Some function, such as for example the definite integral of a real function f(x), can return one
single value and optional extra data (iterations, error estimation, etc...)
If you do not want to see this additional information simply select one cell and insert the
function with the standard procedure. On the contrary, if you want to see also the extra
information, you must select the extra needed cells and insert it as an array-function

11

Xnumbers Tutorial

How to get the help on line

Xnumbers provides the help on line that can be recalled in the same way of any other Excel
function. When you have selected the function that you need, press the F1 key or click on the
guide hyperlink

There is also another way to get the help-on-line. It is from the Xnumbers Function Handbook

Select the function that you want and press the Help button
You can also recall the help guide from the function wizard window

Of course you can open the help on-line from the Xnumber menu

or directly by double clicking on the Xnumbers.hlp file

12

Xnumbers Tutorial

Xnumbers installation
How to install
This addin for Excel XP is composed by the following files:
Addin file

Help file

Handbook file

(It contains the Excel macros

and functions)

(It contains the help notes)

(It contains the macros and

functiosn description list for the
Xnumbers Handbook)

This installation is entirely contained in the folder that you specify.

Put all these files in a same directory as you like.
Open Excel and follow the usually operations for the addin installation:
1) Select <addins...> from <tools> menu,
2) Excel shows the Addins Manager,
3) Search for the file xnumbers.xla,
4) Press OK,

After the first installation, Xnumbers.xla will be add to the Addin Manager
By this tool, you can load or unload the addins that you want, simply switching on/off the checkboxes.
At the starting, the checked addins will be automatically loaded
If you want to stop the automatic loading of xnumbers.xla simply deselect the check box before
closing Excel
If all goes right you should see the welcome
popup of Xnumbers. This appears only when
you activate the check box of the Addin
Manager. When Excel automatically loads
Xnumbers, this popup is hidden

.
13

Xnumbers Tutorial

How to uninstall
If you want to uninstall this package, simply delete its folder. Once you have cancelled the
Xnumbers.xla file, to remove the corresponding entry in the Addin Manager, follow these steps:
1) Open Excel
2) Select <Addins...> from the <Tools> menu.
3) Once in the Addins Manager, click on the Xnumbers entry
4) Excel will inform you that the addin is missing and ask you if you want to remove it from
the list. Give "yes".

Installation troubles
If you do not see the Xnumbers icon or the welcome popup the installation has gone wrong and
probably it is due to same configuration parameter of your Windows / Excel environment
The default security setting of Excel XP 2002/2003 doesn't allowe to install Xnumbers as is. To
prevent errors it is necessary to open the Excel menu option:
Tools > Options > Security > Macro Security, and set: "Average Level security" and "Trust
Access to Visual Basic Project".
Do not worry about it. Xnumbers contains only safe macros; they do not alter your PC in any
way. If you want to remove Xnumbers simply delete all the files contained in the package.
In addition we have to point out that, nowadays, nobody use to hide virus, spy-software, trojan
and related things in a macro Excel: this things belong to the prehistory of informatics!
Nowadays, E-mail and Internet are by far the most important media.

14

Xnumbers Tutorial

Multiprecision Floating Point Arithmetic

Any computer having hardware at 32-bit can perform arithmetic operations with 15 significant
digits1. The only way to overcome this finite fixed precision is to adopt special software that
extends the accuracy of the native arithmetic

Why using extended precision numbers?

First of all, for example, to compute the following operation:
90000000002341 x
8067 =
-------------------726030000018884847 .
Any student, with a little work, can do it. Excel, as any 32-bit machine, cannot! It always gives
the (approximate) result 726030000018885000 , with a difference of +153.
But do not ask Excel for the difference. It replies 0!
The second, deeper, example regards numeric analysis.
Suppose we have to find the roots of the following 9th degree polynomial.
n

Coefficients

P( x) = ai x i

a9
a8
a7
a6
a5
a4
a3
a2
a1
a0

i =0

There are excellent algorithms for finding a

numerical solution of this problem. We can
use the Newton-Raphson method: starting
from x = 32 and operating with 15 significant
digits (the maximum for Excel), we have:

xn
32
31,71962617
31,44427498
31,186622
31,02434159
30,99711858
31,00020851
30,99030069
31,01228318
31,00220065

P(x) (15 digit)

120
43,77734375
15,69921875
4,78125
0,65625
-0,07421875
0,23828125
-0,52734375
0,2421875
-0,03515625

P' (x) (15 digit)

428
158,9873047
60,93164063
29,46289063
24,10644531
24,01953125
24,04980469
23,98925781
24,02050781
23,99023438

1
-279
34606
-2504614
116565491
-3617705301
74873877954
-996476661206
7738306354988
-26715751812360

-P/P'
0,280373832
0,275351191
0,257652979
0,162280411
0,02722301
-0,003089933
0,009907825
-0,021982495
0,01008253
-0,00146544

|xn-x|
1
0,7196262
0,444275
0,186622
0,0243416
0,0028814
0,0002085
0,0096993
0,0122832
0,0022007

As we can see, the iteration approaches the solution x = 31 but the error |xn - x| remains too
high. Why? Multiple roots? No, because P'(x) >> 0. Algorithm failed? Of course not. This
method is very well tested. The only explanation is the finite precision of the computing. In fact,
repeating the calculus of P(x) and P'(x) with 25 significant digits, we find the excellent
convergence of this method.

The basic structure is the IEEE-754 which allows for a 64 bit standard floating point double precision number. The
later has a mantissa of 53 bits (one is the implied bit) which becomes equivalent to 15.4 digits. Excel reports a
maximum of 15 digits. For a good, accurate note see "Excel Computation and Display Issue" by David. A. Heiser,
http://www.daheiser.info/excel/frontpage.html

15

Xnumbers Tutorial
xn

P(x) (25 digit)

P' (x) (25 digit)

-P/P'

|xn-x|

32

120

428

0,28037383

31,71962617

43,71020049043

158,979858019937

0,27494175

0,719626

31,44468442

15,71277333004

61,059647049872

0,25733482

0,444684

31,1873496

4,83334748037

29,483621556222

0,1639333

0,18735

31,02341629

0,56263326884

24,082301045236

0,02336294

0,023416

31,00005336

0,00128056327

24,000000427051

5,3357E-05

5,34E-05

31

0,00000000053

23,999999999984

2,2083E-11

1,54E-11

31

0,00000000004

23,999999999995

1,6667E-12

6,66E-12

The graph below resumes the effect of computation with 15 and 25 significant digits.
|Xn -X|
1
0,01

15 digits

0,0001
1E-06
1E-08

25 digits
1E-10
1E-12
1

Iteration

10

The application field of multi-precision computation is wide. Especially it is very useful for
numeric algorithms testing. In the above example, we had not doubt about the NewtonRaphson method, but what about the new algorithm that you are studying? This package helps
you in this work.

16

Xnumbers Tutorial

Multiprecision methods
Several methods exist for simulating variable multi-precision floating point arithmetic. The basic
concept consists of breaking down a long number into two or more sub-numbers, and repeating
cyclic operations with them. The ways in which long numbers are stored vary from one method
to another. The two most popular methods use the "string" conversion and the "packing"

How to store long number

String Extended Numbers
In this method, long numbers are stored as vectors of characters, each representing a digit in
base 256. Input numbers are converted from decimal to 256 base and vice versa for output. All
internal computations are in 256 base. this requires only 16 bit for storing and a 32 bit
accumulator for computing. Here is an example of how to convert the number 456789 into
string

(456789)10 (6, 248, 85)256

String = chr(6)&chr(248)&chr(85)
This method is very fast, and efficient algorithms for the input-output conversion have been
realized. A good explanation of this method can be found in "NUMERICAL RECIPES in C - The
Art of Scientific Computing", Cambridge University Press, 1992, pp. 920-928. In this excellent
work you can also find efficient routines and functions to implement an arbitrary-precision
arithmetic.
Perhaps the most critical factor of this method is the debug and test activity. It will be true that
the computer does not care about the base representation of numbers, but programmers
usually do it. During debugging, programmers examine lots and lots of intermediate results,
and they must always translate them from base 256 to 10. For this kind of programs, the
debugging and tuning activity usually takes 80 - 90% of the total develop time.
Packet Extended Numbers
This method avoids converting the base representation of long numbers and stores them as
vectors of integers. This is adopted in all FORTRAN77 routines of "MPFUN: A MULTIPLE
PRECISION FLOATING POINT COMPUTATION PACKAGE" by NASA Ames Research
Center. For further details we remand to the refined work of David H. Bailey published in
"TRANSACTIONS ON MATHEMATICAL SOFTWARE", Vol. 19, No. 3, SEPTEMBER, 1993,
pp. 286-317.
Of course this add-in does not have the performance of the mainframe package (16 million
digits) but the method is substantially the same. Long numbers are divided into packets of 6 or
7 digits.
For example, the number 601105112456789 in packet form of 6 digits becomes the following
integer vector:
456789
105112
601

As we can see, the sub-packet numbers are in decimal base and the original long number is
perfectly recognizable. This a great advantage for the future debugging operation.
An example of arithmetic operation - the multiplication A x B = C - between two packet numbers
is shown in the following:
A
456789
105112
601

B
654321
X

The schema below illustrates the algorithm adopted:

17

Xnumbers Tutorial
carry

+
298886 +
68777 +
393 +
0

x
105112 x
601 x
0 x
456789

C'
=
=
=
=

654321
654321
654321
654321

298886635269
68777287838
393315698
393

=>
=>
=>
=>

635269
287838
315697
393

The numbers in the accumulator C' are split into two numbers. The last 6 digits are stored in C,
the remaining left digits are copied into the carry register of the next row.
As we can see, the maximum number of digits is reached in accumulator C'. In the other
vectors, the numbers require only six digits at most. The maximum number of digits for a single
packet depends of the hardware accumulator. Normally, for a 32-system, is 6 digits.. This is
equivalent to conversion from a decimal to a 10^6 representation base. This value is not critical
at all. Values from 4 to 7 affect the computation speed of about 30 %. But it does not affect the
precision of the results in any case.

18

Xnumbers Tutorial

Functions
General Description
Xnumbers is an Excel addin (xla) that performs multi-precision floating point arithmetic.
Perhaps the first package providing functions for Excel with precision from 15 up to 200
significant digits. It is compatible with Excel XP and consists of a set of more than 270 functions
for arithmetic, complex, trigonometric, logarithmic, exponential and matrix calculus covering the
following main subjects.
The basic arithmetical functions: addition, multiplication, and division were developed at the
first. They form the basic kernel for all other functions.
All functions perform multiprecision floating point computations for up to 200 significant digits.
You can set a precision level separately for each function by an optional parameter. By default,
all functions use the precision of 30 digits, but the numerical precision can easily be regulated
continually from 1 to 200 significant digits. In advance some useful constants like , Log(2),
Log(10) are provided with up to 400 digits.

Using Xnumbers functions

These functions can be used in an Excel worksheet as any other built-in function. After the
installation, look up in the functions library or click on the icon

Upon "user's" category you will find the functions of this package.
From version 2.0 you can manage functions also by the Function Handbook. It starts by the
Xnumbers menu

All the functions for multi-precision computation begin with "x". The example below shows two
basic functions for the addition and subtraction.

As any other functions they can also be nested to build complex expressions. In the example
below we compute x^4 with 30 digits precision

As we can see, Xnumbers is powerful, but it does slow down the computation considerably
Therefore, use the multiprecision x-functions only when they are really necessary.

19

Xnumbers Tutorial

Using extended numbers in Excel

If you try to enter a long number with more than 15 digits in a worksheet cell,
Excel automatically converts it in standard precision eliminating the extra
digits. The only way to preserve the accuracy is to convert the number in a
string. It can be done by prefixing it with the hyphen symbol ' .
This symbol is invisible in a cell but avoid the conversion.
Example: enter in a cell the number 1234567890123456789.
We have inserted the same
number with the hyphen in B2
and without the hyphen in B3.
Excel treats the first number as a
string and the second as a
numbers
Note also the different alignment

We have inserted a long

numbers with full precision as a
string in B2
If we try to multiply the cell B2 for
another number, example for 2,
Excel converts the string into
number before performing the
multiplication. In this way the
originally accuracy is destroyed
The only way to perform arithmetic operations
preserving the precision is to use the
multiprecision functions of the Xnumbers
library.
In that case we use the function xmult
Note from the alignment that the result is still
a string

You can also insert extended numbers directly in the function. Only remember that, for
preserving Excel to convert them, you must insert extended numbers like string, within quote
"...".
2469135780246913578 =xmult( "1234567890123456789" , 2 )

20

Xnumbers Tutorial

Functions Handbook
Xnumbers includes an application for searching and pasting the Xnumbers functions,
cataloged by subject. This application can also submit the Xnumbers macros.
You can activate the Functions Handbook from the menu bar Help > Function manager.

Category: you can filter macros by category (Arithmetic, Statistical, Trigonometric, etc.)
Macro Type: filters by macro Functions, by macro Subroutines, or both
Paste Into: choose the cell you want to paste a function, default is the active cell
Search: searches macros by words or sub-words contained into the name or description. For
example, if you input "div" you list all macros that match words like (div, divisor, division,...)
You can associate more words in AND/OR. Separate words with comma "," for OR, with plus
"+", for AND. For example, if you type "+div +multi" you will get all the rows containing words
like (div, divisor, division,...) and words like (multi, multiprecision,...). On the contrary, if you
type "div, multi", you get all the rows that contain words like (div, divisor, division,...) or also the
words like (multi, multiprecision,...). Remember to choose also the Category and Macro Type.
Example, if you enter the word hyperbolic, setting the Category complex, you find the
hyperbolic functions restricted to the complex category.
Help: recalls the help-on-line for the selected function.
OK: insert the selected function into the worksheet ". This activates the standard Excel function
wizard panel. If the macro selected is a "sub", the OK button activates the macro.

21

Xnumbers Tutorial

Precision
Most functions of this package have an optional parameter - Digit_Max - setting the maximum
number of significant digits for floating point computation, from 1 to 200 (default is 30). The
default can be changed from the menu X-Edit\Defaut Digits
This parameter also determines how the output is automatically formatted. If the result has
fewer integer digits than Digit_Max, then the output is in the plain decimal format
( 123.45, -0.0002364, 4000, etc.), otherwise, if the number of integer digits exceeds the
maximum number of digits allowed (significant or not), the output is automatically converted in
exponential format (1.23456789E+94).
The exponent can reach the extreme values of +/- 2,147,483,.647.
The output format is independent of the input format.
In synthesis, the Digit_Max parameter limits:
The significant digits of internal floating point computation
The maximum number output digits, significant or not.
The default of Digit_Max can be changed from the X-Edit menu . It affects all multiprecision
functions and macros.

Formatting Result
The user can not format an extended number with standard Excel number format tools,
because, it is a string for Excel. You can only change the alignment. To change it you can use
the usual standard Excel format tools.

It is possible to separate the digits of a x-numbers in groups, by the user function xFormat()
and xUnformat() 1.
It work similar at the built-in function Format(x, "#,##0.00")
2,469,135,780,246,913,578 = xformat("2469135780246913578",3)

These functions were original developed by Ton Jeursen for the add-in XNUMBER95, the downgrade version of
XNUMBERS for Excel 5. Because they are very useful for examining long string of number, we have imported them in
this package

22

Xnumbers Tutorial

Arithmetic Functions
Addition
xadd(a, b, [Digit_Max])
Performs the addition of two extended numbers: xadd(a, b) = a + b.

Subtraction
xsub(a, b, [Digit_Max])
Performs the subtraction of two extended numbers: xsub(a, b) = a b.
NB. Do not use the operation xadd(a, -b) if b is an extended number. Excel converts b into
double, then changes its sign, and finally calls the xadd routine. By this time the original
precision of b is lost. If you want to change sign at an extended number and preserve its
precision use the function xneg()

Accuracy lack by subtraction

The subtraction is a critical operation from the point of view of numeric calculus. When the
operands are very near each others, this operation can cause a lack of accuracy. Of course
this can happen for addition when the operands are near and have opposit signs. Lets see this
example. Assume one performs the following subtraction where the first operand has a
precision of 30 significant digits

800000.008209750361424423316366
800000
0.008209750361424423316366

(digits)
30
6
25

The subtraction is exact (no approximation has been entered). But the final result have 25 total
digits, of wich only 22 are significant. 8 significant digits are lost in this subtraction. We cannot
do anything about this phenomenon, except to increase the precision of the operands, when
possible.

23

Xnumbers Tutorial

Multiplication
xmult(a, b, [Digit_Max])
Performs the multiplication of two extended numbers: xmult(a, b) = a x b.
The product can often lead to long extended numbers. If the result has more integer digits than
the ones set by Digit_Max, then the function automatically converts the result into the
exponential format.

Division
xdiv(a, b, [Digit_Max])
Performs the division of two extended numbers: xdiv(a, b) = a / b.
If b = 0 the function returns ?. The division can return long extended numbers even when the
operands are small. In the example below we see the well-known periodic division 122 / 7 =
17,428571 . with 30 significant digits.

Inverse
xinv(x, [Digit_Max])
It returns the inverse of an extended number. If x = 0, the function returns ?.
xinv(x) = 1 / x

Integer Division
xdivint(a, b)
Returns the quotient of the integer division for a>0 , b>0.
If b = 0 the function returns ?.
xdivint(a, b)= q , where: a = b*q + r , with 0 r < b

24

Xnumbers Tutorial

Integer Remainder
xdivrem(a, b)
Returns the remainder of the integer division for a>0 , b>0. If b = 0 the function returns ?.
xdivrem(a, b)= r , where: a = b*q + r , with 0 r < b
This function is also called "a mod b "
How to test multiprecision functions ?
This test is the most important problem in developing multiprecision arithmetic. This activity,
absorbs almoust the 60% of the totally realization effort.
Apart the first immediate random tests, we can use many known formulas and algorithms. The
general selecting criterions are:
1.
2.
3.
4.
5.
6.

Formulas should be iterative

Formulas should have many arithmetic elementary operations
Final results should be easyly verified
Intermediate results should be easily verified
Algorithms should be stable
Efficency is not important

For example, a good arithmetic test is the Newton algoritm to compute the square root of a
number. The iterative formula:

xn +1 =

xn 1 xn xn + 2
+ =
2 xn
2 xn

converges to 21/2, starting from x0 = 1.

We have rearranged the formula in order to increase the number of operations (remember: the
efficency is not important). In this way we can test mutliplication, division and addition.
x0 = 1
x1 = 1.5
x2 = 1.41
x3 = 1.41421
Here is a possible Excel arrangement. We have limited the number of the significant digits to
100 only for the picture dimensions, but there is no difficult to repeat it with the maximum digits.

For each iterate only the blu digits are exacts. We see the progressive convergence. By the
way, we note that this algorithm is also very efficent. The rate of convergence is quadratic. The
number of digits approximately doubles at each iteration (In fact this is just the algorithm used
by the xsqr multiprecision function)
But, as said, for testing, the efficiency has no influence. It is important only that the algorithm
involves the most multiprecision functions as possible.

25

Xnumbers Tutorial
Another algorithm quite suitable for testing multiprecision accuracy is the approximatation by
continuous fraction1.
Initialize
X = 31/2, Y = 1/2, T = 6
Iteration
X = (2 + X)1/2
Y = Y/X
T = 2T
P = Y*T*(5Y6/112 + 3Y4/40 + Y2/6 + 1)
Accuracy: approximately 12 decimal digits every 5 iterations)
Below, step by step, a possible Excel arrangement:

The Digit_Max parameter is in the A1 cell. By this parameter we can modulate the arithmetic
accuracy. We have set 30 digits only for the picture dimensions. But you can try with 60, 100 or
more.
Note that, in order to have a more compact form, we have used the xeval function for
calculating the X and P formulas that are inserted into the cells B3 and E3 rispectively.
Selecting the last row (range A6:F6) and dragging it down, we get the following iteration table

The convergence to is evident.

This version, studied by David Sloan (2003), full of many arithmetic operations, permitted us to detect a very hidden
bug in Xnumbers

26

Xnumbers Tutorial

Sum
xsum(v, [Digit_Max])
This is the extended version of the Excel built-in function SUM. It returns the sum of a vector of
numbers. The argument is a standard range of cells.

vi = v1 + v2 + ...vn

Note that you can not use the standard function SUM, because it recognizes extended
numbers as strings and it excludes them from the calculus.

Product
xprod(v, [Digit_Max])
Returns the product of a vector of numbers.

i vi = v1 v2 ...vn
Note that the result is an extended number
even if all the factors are in standard
precision

Raise to power
xpow(x, n, [Digit_Max])
Returns the integer power od an extended number. xpow(x, n) = x n
xpow("0.39155749636098981077147016011",90) = 1.9904508921478176508981155284E-7
xpow(5,81,60) = 5

81

= 413590306276513837435704346034981426782906055450439453125

xpow(122.5 ,1000) = 122.5 1000 = 1.36800819983562905361091390093E+2088

Note the exponent +2088 of the third result. Such kind of numbers can be managed only with
extended precision functions because they are out side of the standard limits for 32bit double
precision.
For not integer power see the exponential functions xexp and xexpa

27

Xnumbers Tutorial

Square Root
xsqr(x, [Digit_Max])
Returnts the square root of an extended number

2 with 30 and 60 significant digits:

The example below shows how to compute the

xsqr(2) =

xsqr(x) =

1.41421356237309504880168872420969807

xsqr(2, 60) = 1.41421356237309504880168872420969807856967187537694807317667973799

Nth- Root
xroot(x, n, [Digit_Max])
Returns the nth root of an extended number
The root's index must be a positive integer.

xroot(x, n)=

The example below shows how to compute the

xroot(100,9) =

100 with 30 and 60 significant digits:

1.66810053720005875359979114908

xroot(100,9,60) = 1.66810053720005875359979114908865584747919268415239470704499

Absolute
xabs(x)
Returns the absolute value of an extended number xabs(x)= |x|
Do not use the built-in function "abs", as Excel converts x in double, then takes the absolute
value but, at that time, the original precision of x is lost.

Change sign
xneg(x)
Returns the opposite of an extended number: xneg(x) = x
Do not use the operator (minus) for extended numbers. Otherwise Excel converts the
extended numbers into double and changes its sign but, at that time, the original precision is
lost. In the following example the cell B8 contains an extended number with18 digits. If you use
the as in the cell B9, you lose the last 3 digits. The function xneg(), as we can see in the cell
B10, preserves the original precision.

28

Xnumbers Tutorial

Integer part
xint(x)
Returns the integer part of an extended number, that is the greatest integer less than or equal
to x.
Examples:
xint(2.99) =
2
xint(2.14) =
2
xint(-2.14) =
3
xint(-2.99) =
3
xint(12345675.00000001) =
xint(12345675.00000001) =

12345675
12345676

Decimal part
xdec(x)
Returns the decimal part of an extended number
Examples:
xdec(2.99) = 0.99
xdec(-2.14) = - 0.14

Truncating
xtrunc(x, [dec])
Returns the truncated number; the parameter "dec" sets the number of decimals to keep
(default 0). It works like Excel function TRUNC. dec can be negative; in that case x is cut to
the integer number, counting back from the decimal point. See the following examples.
Examples:
number

dec

number truncated

4074861.076055370173
4074861.076055370173
4074861.076055370173
4074861.076055370173
4074861.076055370173
4074861.076055370173

-2
-1
0
1
2
3

4074800
4074860
4074861
4074861
4074861.07
4074861.076

Rounding
=xround(x, [dec])
Rounds an extended number, the parameter "dec" sets the decimal number of is to keep
(default 0). It works like standard round function. dec can be negative, in that case x is
rounded to the integer number, starting to count back from decimal point. See the following
examples.

29

Xnumbers Tutorial
number to round

dec

number rounded

6.2831853071795864769
6.2831853071795864769
6.2831853071795864769
6.2831853071795864769
6.2831853071795864769
100352934.23345
100352934.23345
100352934.23345

0
1
2
3
4
0
-1
-2

6
6.3
6.28
6.283
6.2832
100352934
100352930
100352900

When the number is in exponential format, it is internally converted into decimal before the
rounding.
number to round

Decimal format

Dec

number rounded

1.238521E-17
1.238521E-17
1.238521E-17
1.238521E-17

0.00000000000000001238521
0.00000000000000001238521
0.00000000000000001238521
0.00000000000000001238521

16
17
18
19

0
1.E-17
1.2E-17
1.24E-17

Relative Rounding
=xroundr(x, [dgt])
Returns the relative round of a number. The optional parameter Dec sets the significant digits
to keep. (default = 15)
This function always rounds the decimal place no matter what the exponent is
number to round

dgt

number rounded

1.23423311238765E+44
1.23423311238765E+44
1.23423311238765E+44
1.23423311238765E+44
1.23423311238765E+44
1.23423311238765E+44

15
14
13
12
11
10

1.23423311238765E+44
1.2342331123876E+44
1.234233112388E+44
1.23423311239E+44
1.2342331124E+44
1.234233112E+44

30

Xnumbers Tutorial

Extended Numbers manipulation

Digits count
xdgt(x)
Returns the number of digits, significant or not, of an extended number.
It is useful for counting the digits of long numbers
xdgt(4.47213595499957939281834733746) = 30
xdgt(3.99999999999999999999999990000) = 30
xdgt(100000) = 6
xdgt(5856.51311933374) = 15
xdgt(1.2345678E-12) = 8

Significant Digits count

xdgts(x)
Returns the number of significant digits of a number( trailing zeros are not significant).
xdgts("1240100000") = 5
xdgts(3.99999999999999999999999990000) = 26
xdgts(100000) = 1
xdgt(1.2345678E-12) = 8

Numbers comparison
xcomp(a [b])
Compares two extended numbers. It returns the value y defined by:

1 a > b

y = 0a =b
1 a < b

The number b is optional (default b=0)

If the second argument is omitted, the function returns the sign(a)
xcomp(300, 299)= 1
xcomp(298, 299)= -1
xcomp(300, 300)= 0
if b is missing, the function assumes b = 0 for default and then it returns the sign(a)
xcomp(3.58E-12)= 1
xcomp(0)= 0
xcomp(-0.0023)= -1

31

Xnumbers Tutorial

Extended number check

xIsXnumber(x)
Returns TRUE if x is a true extended number.
That is, x cannot be converted into double precision without losing of significant digits. It
happens if a number has more than 15 significant digits.
xIsXnuber(1641915798169656809371) = TRUE
xIsXnuber(1200000000000000000000) = FALSE

Format extended number

=xFormat(str, [Digit_Sep])
=xUnformat(str)
This function1 separates an extended number in groups of digits by the separator character of
you local system ( e.g. a comma "," for USA, a dot "." for Italy). Parameter "str" is the string
number to format, Digit_Sep sets the group of digits ( default is 3)
The second function removes any separator character from the string
Example (on Italian standard):
x = 1230000012,00002345678
xFormat(x,3) = 1.230.000.012,000.023.456.79
xFormat(x,6) = 1230.000012,000023.45679
Example (on USA standard):
xFormat(x,3)= 1,230,000,012.000,023,456,78
xFormat(x,6)= 1230,000012.000023,45678

Check digits
DigitsAllDiff(number)
This function2 return TRUE if a number has all digits different.
DigitsAllDiff(12345) = TRUE
DigitsAllDiff(123452) = FALSE
Argument can be also a string. Example
DigitsAllDiff(12345ABCDEFGHIM) = TRUE
DigitsAllDiff(ABCDA) = FALSE

These functions were original developed by Ton Jeursen for his add-in XNUMBER95, the downgrade version of
XNUMBERS for Excel 5.
Because it works well and it is very useful for examining long string of number, I have imported it in this package.
2

This function appears by the courtesy of Richard Huxtable

32

Xnumbers Tutorial

SortRange
=SortRange (ArrayToSort, [IndexCol], [Order], [CaseSensitive])
This function returns an array sorted along a specified column
ArrayToSort: is the (N x M ) array to sort
IndexCol: is the index column for sorting (1 default)
Order: can be "A" ascending (default) or "D" descending
CaseSensitive: True (default) or False. It is useful for alphanumeric string sorting
Example: The left table contains same points of a function f(x,y). The right table is ordered from
low to high function values (the 3-th column)

Digits sum
sumDigits(number)
This useful1 function returns the digits sum of an integer number (extended or not)
sumDigits(1234569888674326778876543) = 137

Vector flip
Flip(v)
This function returns a vector in inverse order [a1, a2, a3, a4] [a4, a3, a2, a1]

This function appears by the courtesy of Richard Huxtable

33

Xnumbers Tutorial

Scientific format
xcvexp(mantissa, [exponent])
This function converts a number into scientific format. Useful for extended numbers that, being
string, Excel cannot format.
xcvexp(-6.364758987642234, 934) = -6.364758987642234E+934
xcvexp(1.2334567890122786, ) = 1.2334567890122786E-807

This function is useful also to convert any xnumbers into scientific notation, simply setting
exponent = 0 (default)
xcvexp(12342330100876523, 0) = 1.2342330100876523E+16
xcvexp(0.000023494756398348) = 2.3494756398348E-5

Split scientific format

xsplit(x)
This function returns an array containing the mantissa and exponent of a scientific notation.
If you press Enter this function returns only the mantissa. If you select two cells and give the
CTRL+SHIFT+ENTER sequence, you get both mantissa and exponent

Note that, in the last case, you cannot convert directly into double (for example, using the
VALUE function), even if the number of digits is less than 15. The exponent is too large for the
standard double precision.

Converting multiprecision into double

=xcdbl(x)
This function converts an extended number into standard double precision
It can be nested with other functions and/or array-functions.
Usually the extended numbers are too long for a compact visualization. So, after, the
multiprecision computation, we would like to convert the result in a compact standard precision.
For example, if you invert the matrix
2
7
-4

1
12
0

4
4
8

using the multiprecision xMatInv function, you will get a matrix like the following
0.3076923076923076923076923
-0.2307692307692307692307692
0.1538461538461538461538462

-0.02564102564102564102564103
0.1025641025641025641025641
-0.01282051282051282051282051

34

-0.1410256410256410256410256
0.06410256410256410256410256
0.05448717948717948717948718

Xnumbers Tutorial
If you use the functions xcdbl nested with the multiprecision function, the matrix will be
rounded in standard precision and the output will have a more compact format

Note that xcdbl affects only the output. The internal computing is always performed in
multiprecision.

Macros X-Edit
These simple macros are very useful for manipulating extended numbers in the Excel
worksheet. They perform the following operations:
Format
Unformat
Double Conversion
Round
Relative Round
Mop-Up

Separates groups of digits

Removes the separator character
Converts multiprecision numbers into standard double precision
Rounding multiprecision numbers
Relative rounding multiprecision numbers
Converts small numbers into 0

From this menu you can also change the default Digit_Max parameter
Using these macros is very simple. Select the range where you want to operate and then start
the relative macro. They work only over cells containing only numeric values, extended or
standard. Cells containing function are ignored
Tip. For stripping-out a formula from a cell and leaving its value, you can select the cell and
then click in sequence
(copy + paste values)
Here are same little examples:
Format - group 6
31415926.53589793238462643
19831415926.53589793238462
0.535897932384626433832734

31,415926.535897,932384,62643
19831,415926.535897,932384,62
0.535897,932384,626433,832734

Double Conversion
31415926.53589793238462643
19831415926.53589793238462
0.535897932384626433832734

31415926.5358979
19831415926.5358
0.535897932384626

31415926.536
19831415926.536
0.536

Rounding 3 decimals.
31415926.53589793238462643
19831415926.53589793238462
0.535897932384626433832734

35

Xnumbers Tutorial
Relative rounding - significant digits 15.
4.5399929762484851535591E-5
1.0015629762484851535591E-6
0.539929762484851535591E-12

4.53999297624849E-05
1.00156297624849E-06
5.39929762484852E-13

Mop-Up - Error limit 1E-15.

31415926.53589793238462643
-1.00E-15
5.78E-16
-1.40E-18

31415926.53589793238462643

0
0
0

Note that the function mopup is used overall for improving the readability. The cells having
values greater than the limit are not modified.

Macro X-Converter
This macro1 (*) converts a well formed Excel formula to the equivalent in terms of Xnumber
multiprecision functions (xadd, xmult, etc.).
The advantage over x-evaluate is that the code can be debugged in normal excel using small
values
and in a familiar form. When the spreadsheet works it is ' converted to using nested x-calls for
the precision work. The intention is to tag the conversion onto the copy worksheet function ' so
that you end up with a multi-precision copy of the original.
Its main features are:
converts a range of cells directly on site or in a new worksheet.
skips cells having functions not convertible.
skips array functions except MINVERSE and MMULT that are substituted with the
correspondent x-functions xMatInv and xMatMult.
Errors (if any ) are shown on the panel
The digits parameter can be set in 4 different ways:
1) an integer number i.e. 30
2) a fixed reference of a cell, i.e. $A$1
3) a name of a cell, i.e. "digits"
4) simply nothing. In that case all the x-functions use the internal default = 30
Of course not all the Excel Functions can be converted. The list of the Excel functions2
converted is:
* , / , + , - , ^ , ABS , ACOS , ACOSH , ASIN , ASINH , ATAN , ATANH , AVERAGE
, COMBIN , COS , COSH , EXP , FACT , INT , LN , LOG , MDETERM , MINVERSE ,
MMULT , MOD , PI , PRODUCT , ROUND , SIN , SINH , SQRT , STDEV , STDEVP , SUM
, TAN , TANH , TRUNC , VAR , VARP

The conversion engine of this macro was originally developed by John Jones in an smart m4 macro language which
makes the VB native j-code more simple (!) to write. It appears in this package thanks to his courtesy

The functions are indicated with their original English names, that usually are different from the local names.

36

Xnumbers Tutorial
Example. Assume to have an Excel formula in the cell B6 calculating the norm of a 3dim vector
like the following worksheet.

We would transform this formula into a multiprecision one using the x-functions of the
Xnumbers add-in
Select the cell B6 and start the macro X-Converter from the menu X-Edit

Range to convert: the range containing the one or more formulas to convert
Digit Parameter: specifies the global precision digits for all the functions under conversion.
You can leave "(Default)" or blank meaning that all the function will use the internal default
precision (usually 30 digits). But you can set a reference cell (example $A$1) where you have
set your own precision. Or you can insert a name of a cell (example "mydigits") that you have
defined. You can also insert directly a number (example 25)
Output Option: sets the output where the multiprecision function will be copied. Select "on
site" if we want to convert the formula directly in the selected cell. Select "new sheet" if we want
to perform the conversion in a new worksheet. The original formula will be preserved.
Let's select "on site" and click "run". The worksheet look will look like as the following

As we can see the original standard formula

=SQRT(B2+B3+B4)
in cell B6 has been substituted with the multiprecision formula
37

Xnumbers Tutorial
=xsqr(xadd(xadd(B2,B3),B4)).
working with the default precision digits (30)
Note that there are 3 nested functions in the converted formula.
The maximum number of nested functions in Excel is 8.
If an error raises during the cell conversion, the formula is left unchanged and the text is turned
in red. The macro always shows a short list of the errors encountered during the conversion.
The macro works also with a range selection
Example. In the following worksheet we have same functions in the range A3:A18
Note that the cell A18 contains the
function COUNTA that are not the
similar x-function
Note that the range A14:A15
contains an array function { ... },
thus the cells A14 and A15 are
inseparable.
The cells A7 contain a simple
constant

We want to convert them, where possible, in a new worksheet without affect the original
worksheet.
For that, select all the range A3:A18 and start the X-Converter
Select "new sheet" and click "run". The new worksheet look like as the following

Note that the cell A7, A14, A15, A18 are unchanged
Note that the original cells A9 = B7^12 and A13 =A11^2.6 having the same operator symbol
"^", are substituted with two different x-functions: xpow for integer power and xexpa for float
power.

38

Xnumbers Tutorial

Statistical Functions

Factorial
xfact(n, [Digit_Max])
Returns the factorial of an integer number xfact(n)= n!
This example shows all 99 digits of 69!
xfact(69, 100) = 71122452428141311372468338881272839092270544893520369393648040923257279754140647424000000000000000

If the parameter Digit_Max is less than 99, the function returns the approximate result in
exponential format:
xfact(69) = 1.71122452428141311372468338881E+98

For large number (n>> 1000) you can use the faster function xGamma(x). The relation between
the factorial and the gama function is:
(n) = (n-1)!

Factorial with double-step

xfact2(n, [Digit_Max])
Returns the factorial with double step.
if n is odd

xfact2(n) = 13579...n

if n is even xfact2(n) = 2468 ...n

Note: In many books, this function is indicate improperly as "double factorial", or - even worse - with the
confusing symbol "!!".

Combinations
xComb(n, k, [Digit_Max])
Returns the binomial coefficients, a combination of n, class k. xcomb = C n,k
The example below shows all the 29 digits of the combination of 100 objects grouped in class
of 49 elements:
xComb(100,49) = 98913082887808032681188722800

Combinations of N = 100
objects in class of 10, 20,
90
Note the typical parabolic
outline of the binomial
coefficients

For large argument (n and k >>1000) use the faster function xcomb_big(n,k) .
39

Xnumbers Tutorial

Permutations
xPerm(n, [k], [Digit_Max])
Returns the permutation of n, class k. xperm(n,k)= Pn,k.
If k is omitted, the function assume k = n and in this case will be P(n) = n!
Examples:
xPerm(100, 20, 60) = 1303995018204712451095685346159820800000
xPerm(100) = 9.33262154439441526816992388562E+157

Arithmetic Mean
xmean(x, [Digit_Max])
Returns the arithmetic mean of n numbers, extended or not. The argument is a range of cells.

M=

Geometric Mean
xgmean(x, [Digit_Max])
Returns the geometric mean of n numbers, extended or not.

GM =

x1 x 2 x 2 ... x n

Quadratic Mean
xqmean(x, [Digit_Max])
Returns the quadratic mean of n numbers, extended or not.

QM =

Standard Deviation
xstdev(range, [Digit_Max])
xstdevp(range, [Digit_Max])
Return the deviation of n numbers, extended or not. Range is a range of cells. The optional
parameter Digit_Max, from 1 to 200, sets the number of significant digits (default 30)

xstdev returns the standard deviation

(x x)
n 1

40

Xnumbers Tutorial

xstdevp returns the population standard deviation

p =

(x x)

Variance
xvar(range, [Digit_Max])
xvarp(range, [Digit_Max])
Return the variance of n numbers, extended or not. Range is a range of cells. The optional
parameter Digit_Max, from 1 to 200, sets the number of significant digits (default 30)

xvar returns the standard variance

v =

(x x)

n 1

xvarp returns the population variance

vp

(x x)
=

Probability distributions
Xnumbers contains several type of probability distribution functions
DSBeta(x, a, b, [dtype])
DSBinomial(k, n, p, [dtype])
DSCauchy(x, m, s, n, [dtype])
DSChi(x, r, [dtype])
DSErlang(x, k, l, [dtype]))
DSGamma(x, k, l, [dtype]))
DSLevy(x, l, [dtype]))
DSLogNormal(x, m, s, [dtype]))
DSLogistic(x, m, s, [dtype]))
DSMaxwell(x, a, [dtype]))
DSMises(x, k, [dtype]))
DSNormal(x, m, s, [dtype]))
DSPoisson(k, z, [dtype]))
DSRayleigh(x, s, [dtype]))
DSRice(x, v, s, [dtype]))
DSStudent(t, v, [dtype]))
DSWeibull(x, k, l, [dtype]))

Beta distribution

0 < x < 1 , a > 0, b > 0,

Binomial distribution

k integer, n integer , 0 < p < 1

Cauchy (generalized) distribution

n integer , s > 0

Chi distribution

r integer, x > 0

Erlang distribution

k integer, x > 0

Gamma distribution

x > 0, k > 0, l > 0

Levy distribution

x > 0, l > 0

Log-normal distribution

x > 0, m 0, s > 0

Logistic distribution

x > 0, m 0, s > 0

Maxwell-Boltzman distribution

x > 0, a > 0

Von Mises distribution

k > 0, - < x <

Normal distribution

s>0

Poisson distribution

k integer, z > 0

Rayleigh distribution

x > 0, s > 0

Rice distribution (j=1 cumulative)

x > 0, v 0 , s > 0

Student distribution (j=1 cumulative)

v integer degree of freedom

Weibull distribution (j=1 cumulative)

x > 0, k integer, l > 0

The optional parameter dtype = 0 (default) returns the density distribution f(x); dtype = 1
returns the cumulative distribution F(x).
x

F ( x) = P(x < x ) = f (t )dt

a

The lower limit "a" depends by the definition domain of the density function f(x).

41

Xnumbers Tutorial

Probability density definition

Rayleigh

Gamma

Cauchy

Beta

Lvy

Logistic

42

Xnumbers Tutorial

MaxwellBoltzmann

Rice

Erlang

Normal

Von Mises

Weibull

43

Xnumbers Tutorial

Log-normal

Binomial

f ( k , n, p ) =

Student

f (t , v) =

n!
p k (1 p ) n k
k!(n k )!

Poisson

( 12 (v + 1))
2 v ( 12 v)(1 + 12 t 2 ) ( v +1) / 2

f ( x, v ) =

Chi squared

x v / 21e x / 2
f ( x, v ) = v / 2 1
2 ( 2 v)

44

xve x
v!

Xnumbers Tutorial

45

Xnumbers Tutorial

Univariate Statistic
xstatis(range, [digit_max])
This function returns the univariate statistic summary of a range of data.
The statistics computed are:
Total of numerical elements of the range

Arithmetic Mean

n
Standard Deviation

(x x)

(x x)

n 1

Popul. Standard Deviation

Autocorrelation lag1

n 1

( xi x )( xi +1 x )

(x x)
n

This function returns a vector of 5 elements. Use the ctrl+shift+enter key sequence for insert it.

46

Xnumbers Tutorial

Linear Regression Coefficients

xRegLinCoef( Y, X, [DgtMax], [Intcpt])
Computes the multivariate linear regression with the least squares method in multi-precision
arithmetic.
Y is a vector (n x 1) of the dependent variable.
X is a list of independent variable. It may be a vector for monovariable regression (n x 1), or a
matrix for multivariate regression (n x m).
DgtMax sets the precision (default 30).
Intcpt = true/false. If true (default ) the algorithm calculates the intercept; otherwise the intercept
is set to 0
For monovariate regression, this function returns two coefficients [a0, a1]; the first one is the
intercept of Y axis, the second one is the slope.
Simple Linear Regression
Example. Evaluate the linear regression for the following xy data table
y

x
0.1

The model for this data set is

1991

0.2

1991.001046

y = a0 + a1 x

0.35
0.4
0.45
0.6

1991.001831
1991.002092
1991.002354
1991.003138

Where [a0 , a1] are the unknown coefficents that

can be evaluate by the xRegLinCoef function

0.7
0.8
0.9
1
1.5
1.8
2
3

1991.003661
1991.004184
1991.004707
1991.00523
1991.007845
1991.009414
1991.01046
1991.01569

We can also compute the factor r2 in order to

measure the goodness of the regression
This can be done by the xRegLinStat function

Let's see other examples

47

Xnumbers Tutorial

Example of univariate regression

Example of multivariate regression

Example of linear regression with intercept = 0

Note that the x-regression functions always returns a vector of m+1 values [a0, a1, ...am]
independently if the intercept is 0 or not.
Multivariate Regression
This function can also compute a multivariate regression. This is when y depends by several
variables x1, x2, xn. Look at this example
x1

x2

x3

-4

4000.8

0.1

-2

4000.7

0.2

0.5

-1

4001.55

0.3
0.4
0.5

0.5
1
1

0
1.5
2

4001.65
4002.4
4002.59

The model for this data set is

y = a0 + a1 x1+ a2 x2+ a3 x3
Where [a0 , a1 , a2 , a3] are the unknown
coefficients

48

Xnumbers Tutorial

Linear Regression - Standard Deviation of Estimates

xRegLinErr( Y, X, Coeff, [DgtMax], [Intcpt])
Returns the standard deviation of the linear regression of estimate
Y is a vector (n x 1) of dependent variable values.
X is a list of independent variable values. It may be a vector for monovariable regression (n x 1)
or a matrix for multivariate regression (n x m).
Coeff is the coefficients vector of the linear regression function [a0, a1, a2...am].
DgtMax sets the digits precision of computing.
Intcpt = true/false. If true (default ) the algorithm calculates the intercept deviation; otherwise
intercep a0 is set to 0
See above examples
Polynomial Regression
The same function finding the linear regression can easily be adapted to the polynomial
regression. In the example below we will find the best fitting polynomial of 3rd degree for the
given data
x

10

1120

11

1473

12

1894

13

2389

14

2964

15

3625

16

4378

17

5229

18

6184

19

7249

20

8430

The model for this data set is

y = a0 + a1 x+ a2 x2+ a3 x3 where [a0 , a1 , a2 , a3] are the unknown coefficients
First of all, we add at the given table two extra columns containing the power x2, x3
They can easily be computed in an Excel worksheet as shown above.
The polynomial coefficients can be computed by xRegLinCoef.
The exact result is y = 10 + x + x2 + x3

49

Xnumbers Tutorial

Linear Regression Formulas

Generally, the multivariate linear regression function is:

y = a0 + a1 x1 + a2 x2 + ...am xm
where:

[a0 , a1 , a2 ... am ]

The coefficients of regression can be found by the following algorithm

Make the following variables substitution:

X i = xi x

for i = 1..m

Y = y y
where the right values are the averages of samples y and x respectively:

y=

1
yk
n k

xi =

1
xi ,k
n k

After that, the coefficients a= [a1, a2, ....an] are the solution of the following linear system

[C] a = b
where [C] is the cross-covariance matrix, and b is the XY covariance

j X 12, j

=
C=
=

j X 1, j X 2, j j X 1, j X 3, j .. j X 1, j X m, j

j X 22, j
j X 2, j X 3, j .. j X 2, j X m, j
.. j X 3, j X m, j
=
j X 32, j

=
=
j X m2 , j

j Y j X 1, j
YX
j j 2, j
b = j Y j X 3, j

....

j Y j X m, j

and the constant coefficient is given by:

m

a0 = Y ai X i
i =1

For m=1 we obtain the popular formulas of the monovariate linear regression

a1 =

j Yj X j
j X 2j

a0 = Y a1 X

This is the linear solution known as the Ordinary Least Squares (OLS). The analysis of this kind
of approach shows that, for large dimensions of n (many measurement values) the matrix C
becomes nearly singular

50

Xnumbers Tutorial

Linear Regression Covariance Matrix

xRegLinCov ( Y, X, Coeff, [DgtMax], [Intcpt] )
Returns the (m+1 x m+1) covariance matrix of a linear regression of m independent variables

y = a 0 + a1 x1 + a 2 x 2 ... + a m x m
For a given set of n points Pi = ( x1i x2i

xmi , yi )

Parameter Y is an (n x 1) vector of dependent variable. Parameter X is a matrix of independent

variables. It may be an (n x 1) vector for monovariable regression or an (n x m) matrix for
multivariate regression.
Parameter Coeff is the (m+1) vector of the linear regression coefficents
Cross Covariance Matrix
Given the matrix X of the independent variables points
Intercept calculated

Intercept = 0

1 x11
1 x
12
X =
... ...

1 x1n

... xm1
... xm 2
... ...

... xmn

The covariance matrix C is

C = s2 X X T

... xm1
... xm 2
... ...

... xmn

x11
x
X = 12
...

xn 2
The covariance matrix C is

Where:

C = s2 X X T

Where:

(y
=
i

2
y i )

n m 1

(y
=

y i ) 2

nm

Note that the square roots of the diagonal elements of the covariance matrix

si = cii
are the standard deviations of the linear regression coefficients

51

Xnumbers Tutorial

Linear Regression Statistics

xRegLinStat( Y, X, Coeff, [DgtMax], [Intcpt])
Returns some statistics about the linear regression
R2

Square of the linear correlation factor

S y,x

Standard deviation of the linear regression

Parameter Y is a vector (n x 1) of dependent variable.

Parameter X is a list of independent variable. It may be an (n x 1) vector for monovariable
regression or a (n x m) matrix for multivariate regression.
Coeff is the coefficients vector of the linear regression function [a0, a1, a2...am].
Formulas
The regression factor (better: the square of regression factor) R2 lie between 0 and 1 and
roughly indicates how closely the regression function fits the given values Y.
Generally, it can be computed by the following formula:

( y yi* ) 2
= 1 i i
i ( yi y ) 2

=1

y y*
2
y

Where y is the value estimated by the regression function and y is the mean of y values.

y* = a0 + a1 x1 + a2 x2 + ...am xm
y=

1
yk
n k

For monovariate regression (m=1), the above formula returns the popular formula:
R2 =

x2
y

( x )2
n

( y )2

n

Standard error of the linear regression is:

Intercept calculated

Intercept constrained to 0

(y

(y

s y,x =

yi*

s y,x =

n gl 1

Where gl = number of independent variables

52

yi*

n gl

Xnumbers Tutorial

Linear Regression Evaluation

= xRegLinEval(Coeff, X)
Evaluates the multivariate linear regression in multi precision arithmetic.
Parameter Coeff is the coefficients vector [a0. a1, a2, ....] of the linear regression
Parameter X is the vector of independent variables. It is one value for a simple regression
The functions return the linear combination.

y = a0 + a1 x1 + a2 x2 + ...an xn
Example: Plot the linear regression for the following data set
x
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1

y
0.58
0.65
0.88
1.25
1.32
1.14
1.31
1.51
1.54
1.48
1.98

In this worksheet, each value of linear regression y is computed by the function xRegLinEval.
The regression coefficients are computed by xRegLinCoef. The results are converted in double
by the function xcdbl.
Selecting the range A1:C12 and plotting the data we get the following regression graphs
2.4

y*

1.6
1.2
0.8
0.4
0
-1

-0.75

-0.5

-0.25

53

0.25

0.5

0.75

Xnumbers Tutorial

Polynomial Regression - Coefficient

xRegPolyCoef( Y, X, Coeff, [Degree], [DgtMax], [Intcpt])
Computes the polynomial regression with the least squares method in multi precision
arithmetic.

y = a0 + a1 x + a2 x 2 ... + am x m
with m > 1
Y is a vector (n x 1) of the dependent variable.
X is a vector (n x 1) of the independent variable
Degree is the degree m >1 of the polynomial
DgtMax sets the precision (default 30).
Intcpt = true/false. If true (default) the algorithm calculates the intercept; otherwise the intercept
is set to 0
The function returns the coefficient vector [a0, a1, a2...am]
Example. Find the 3rd degree polinomial fitting the given data set (xi, yi)

Polynomial Regression - Standard Deviation of Estimates

xRegPolyErr ( Y, X, Coeff, [Degree], [DgtMax], [Intcpt])
Returns the standard deviation of the polynomial regression estimate
Y is a vector (n x 1) of dependent variable values.
X is a vector (n x 1) of the independent variable
Coeff is the coefficients vector [a0, a1, a2...am]. of the polynomial regression
DgtMax sets the precision (default 30)..
Intcpt = true/false. If true (default) the algorithm calculates the intercept; otherwise the intercept
is set to 0
see above example

Polynomial Regression Statistics

xRegPolyStat ( Y, X, Coeff, [Degree], [DgtMax], [Intcpt])
Computes the R-squared factor of the polynomial regression and the standard deviation of the
regression.
54

Xnumbers Tutorial

R2

R-Squared coefficient of the polynomial regression

S y,x

Standard Deviation of the polynomial regression

Y is a vector (n x 1) of dependent variable values.

X is a vector (n x 1) of the independent variable
Coeff is the coefficients vector [a0, a1, a2...am]. of the polynomial regression
DgtMax sets the precision (default 30).
Intcpt = true/false. If true (default ) the algorithm calculates the intercept deviation; otherwise
this is set to 0
If you wanto to see the standard deviation select two cells and give the CTRL+SHIFT+ENTER
sequence.

Macro - Regression
Xnumbers contains two macros performing the linear and polynomial regression in
multiprecision arithmetic. From the Xnumbers menu, select

Macro \ Regression \ Linear

Macro \ Regression \ Polynomial

Data XY is an array containing the column of the variable X at the left with the adjacent column
of Y at the right. For a m-multivariate regression also the columns of X must be exactly m.
Select this range before starting the macro. In this case the input box will be automatically filled
Data XY for univariate and
polynomial regression

Data XY for
multivariate regression

Output cell indicates the starting upper left cell of the output area.
Intercept. If checked the macro calculates the intercept otherwise it is set to 0
Convert to double. Converts the multiprecision in double before the output
55

Xnumbers Tutorial
R-squared + residual std. dev. Calculate the correspondent statistics
Estimate std. dev. Calculate the standard deviation of estimate
Digits Max. from 1 to 200, sets the precision (default 30)
Degree. The panel of the Polynomial Regression macro is similar to the Linear except for the
input degree box that allows to chose the polynomial degree.

An example of the regression output is.

To obtain this result follow this steps:.

Select the range A2:C7 and start the macro.
Select all the option boxes and deselect the "double conversion" in
order to see all the 30 digits.
Note. The text has been added by hand only for clarity. The macro do not write them. We do it.
The coefficient a0 (intercept) will be always output. If the intercept is switched off, the result in
the cell F2 will be 0

56

Xnumbers Tutorial

Sub-Tabulation
One important application of linear regression is the sub-tabulation, which is the method to
extract a table of values with smaller step from an original table with bigger steps. In other
words, we can obtain a fine tabulation from a table with a few values of a function. Lets see
this example.
Example: Extract from the following dataset, a table having 11 values with step 0.1, from 0 to 1
y

x
0
0.2
0.5
0.6
0.7
1

First of all, we find the linear regression coefficients

5.1
4.7
4.5
4.3
4.2
3.6

[ a 0 , a1 ]
Than we re-calculate the values
y i = a0 + a1 h , i = 110, h = 0.1

5.50
5.00
4.50
4.00
3.50
3.00
0

0.2

0.4

0.6

0.8

The graph shows the extra points added by the sub tabulation. Note that this method is
different from the interpolation because the regression line does not pass through any of the
original points. The new values of the table B are different from the ones table A even in the
same x-values.
This feature came in handy when we want to regularize the row data.

Data Conditioning
The conditioning of the data consists of subtracting the mean from the values of the sample. It
can improve the accuracy of the linear regression, but the regression coefficients obtained conditioned coefficients - are different from the regression coefficients of the row data. They
can be back-transformed by the following method:
Given X and Y two data vectors, the linear regression polynomial of n degree well be:

57

Xnumbers Tutorial
n

p( x) = ai x i
i =0

We made the data conditioning, making the average of X and Y

x=

1
xi
n

y=

1
yi
n

Substituting the old variables with the new variable u and v

u i = xi x

vi = y i y

Than, the new linear regression polynomial well be:

n

p(u ) = bi u i
i =0

The original ai coefficients can be obtained from the new bi coefficients by the following
formulas.
n

a0 = y + (1)i bi x i
i =0

n
i
ak = (1)i + k bi x i k
i =k
k

This method is often useful for increasing the gloabal accuracy of the linear regression

58

Xnumbers Tutorial

Linear Regression with Robust Method

RegLinRM(x, y, [Method])
This function1 performs the linear regression with three different robust methods:
- SM: simple median
- RM: repeated median
- LMS: least median squared
Robust methods are suitable for data containing wrong points. When data samples have noise
(experimental data), the basic problem is that classic LMS (least minimum squared) is highly
affected by noisy points. The main goal of robust methods is to minimize as much as possible
the influence of the wrong points when fitting the function
The parameter x and y are two vectors of the points to fit.
The optional parameter "Method" sets the method you want to use (default = SM)
The functions returns an array of two coefficients [a1, a0] where

y a1 x + a0
Use CTRL+SHIFT+ENTER to paste it.
Example: Suppose you have sampled 5 experimental values (xi, yi), with a (suspected) large
error in the last value 6.5.
x

1
2
3
4
5

1.1
2
3.1
3.8
6.5

7
6
5
4

Ordenadas
SM
RM
LM S
M
.CC.
SM
tandard

Line a r Re gre ssion

3
2

1
In the graph are shown the
regression lines obtained with
0
all robust methods in
0
1
2
3
4
5
6
comparison with the standard
OLS regression.
As we can see all the lines SM, RM, LMS (Robust Methods) minimize the influence of the value
(5, 6.5)

The routines for robust linear regression were developed by Alfredo lvarez Valdivia. They appear in this collection
thanks to its courtesy

59

Xnumbers Tutorial

Linear Regression Min-Max

RegLinMM(x, y)
This function performs the linear regression with the Min-Max criterion (also called Chebychev
approximation) of a discrete dataset (x, y)
The parameter "x" is a (n x 1) vector of the independent variable,
The parameter "y" is a (n x 1) vector of the dependent variable
The function returns the coefficients [a0, a1] of the max-min linear regression

~
y = a0 + a1 x
As known, those coefficients minimize the max absolute error for the given dataset

E = max | ~
y ( xi ) yi |
Example. Find the better fitting line that minimize the absolute error

The liner regression is

y 0.428 + 1.142 x

with an error max

Emax 0.7

The scatter plot shows the lineare regression approximation

As we can see, all the points lie in the plane strips of Emax around the min-max line (pink
line). (Emax 0.7)

60

Xnumbers Tutorial

NIST Certification Test

The multiprecision regression functions of Xnumbers have been tested against the NIST linear
regression data sets. The table below gives the LRE results from xRegLinCoef and
xRegPolyCoef functions
NIST test for linear regression (row data sets)

StRD Datasets

Level of
difficulty

Model of
class

Coeff.

Function

Norris

low

Linear

xRegLinCoef

LRE
15.0

Pontius

low

Quadratic

xRegPolyCoef

15.0

NoInt1

Average

Linear

xRegLinCoef

15.0

Filip

high

Polynomial

11

xRegPolyCoef

15.0

Longley

high

Multilinear

xRegLinCoef

15.0

Wampler1

high

Polynomial

xRegPolyCoef

15.0

Wampler2

high

Polynomial

xRegPolyCoef

15.0

Wampler3

high

Polynomial

xRegPolyCoef

15.0

Wampler4

high

Polynomial

xRegPolyCoef

15.0

Wampler5

high

Polynomial

xRegPolyCoef

15.0

The table below gives the LRE results on the NIST test univariate data sets obtained from
xStats function
NIST test for univariate data sets

Difficulty Size

Stand.
Dev.

Mean

Autocor.
Coef.

Dataset

Category

PiDigits

Univariate

5000

15

15

15

Lottery

Univariate

218

15

15

15

Lew

Univariate

200

15

15

15

Maveo

Univariate

50

15

15

15

Michelso

Univariate

100

15

15

15

NumAcc1

Univariate

15

15

15

NumAcc2

Univariate

1001

15

15

15

NumAcc3

Univariate

1001

15

15

15

NumAcc4

Univariate

1001

15

15

15

61

Xnumbers Tutorial

Transcendental Functions

Logarithm natural (Napiers)

xLn(x, [Digit_Max])
Returns the natural logarithm (or Napiers) , in base "e"
The argument may be either normal or extended number.
Example:
xLn(30) = 3.4011973816621553754132366916

Logarithm for any base

xLog(x, [base], [Digit_Max])
Returns the logarithm for any base (default 10)

y = log base ( x)
The argument may be either normal or extended number.
Example
xlog(30) = 1.47712125471966243729502790325

Exponential
xexp(x, [Digit_Max])
Returns the exponential of x in base "e"
Example
e10
1000

xexp(x) = e

= xexp(10) = 22026.4657948067165169579006452
= xexp(1000) = 1.97007111401704699388887935224E+434

Note the exponent 434 of the second result. Such kind of numbers can be managed only with
extended precision functions because they are outside the standard limits of double precision.

Exponential for any base

xexpa(x, [a], [Digit_Max])
x

Returns the exponential of x in any in base

xexpa(x, a) = a
The arguments a and x may be either normal or extended numbers, with a > 0.
If the base "a" is omitted the function assumes a = 10.
21.234

= xexpa(1.234, 2) = 2.3521825005819296401155858555

100.54

= xexpa(-0.54) = 0.288403150312660594239196924659

62

Xnumbers Tutorial

Constant e
xe([Digit_Max])
Returns Euler's constant "e", the base of the natural logarithm.
The optional parameter Digit_Max, from 1 to 415, sets the number of significant digits (default
30).
xe()

= 2.71828182845904523536028747135

xe(60) = 2.71828182845904523536028747135266249775724709369995957496696

Constant Ln(2)
xLn2([Digit_Max])
Returns the constant Ln(2).
The optional parameter Digit_Max, from 1 to 415, sets the number of significant digits (default
30).

Constant Ln(10)
xLn10([Digit_Max])
Returns the constant Ln(10).
The optional parameter Digit_Max, from 1 to 415, sets the number of significant digits (default
30).

Hyperbolic Sine
xsinh(x, [Digit_Max])
Returns the hyperbolic sine of x in multiprecision arithmetic

sinh =

e x e x
2

Hyperbolic ArSine
xasinh(x, [Digit_Max])
Returns the hyperbolic arsine of x in multiprecision arithmetic

asinh( x) = ln x + x 2 + 1

63

Xnumbers Tutorial

Hyperbolic Cosine
xcosh(x, [Digit_Max])
Returns the hyperbolic cosine of x in multiprecision arithmetic

cosh( x) =

e x + e x
2

Hyperbolic ArCosine
xacosh(x, [Digit_Max])
Returns the hyperbolic Arcosine of x in multiprecision arithmetic
The argument x, normal or extended, must be x >1

acosh = ln x + x 2 1 , x > 1

Hyperbolic Tangent
xtanh(x, [Digit_Max])
Returns the hyperbolic tangent of x in multiprecision arithmetic

tanh( x) =

e x e x
e x + e x

Hyperbolic ArTangent
xatanh(x, [Digit_Max])
Returns the hyperbolic artangent of x in multiprecision arithmetic
The argument x, normal or extended, must be |x| < 1

1 1+ x
atanh ( x) = ln

2 1 x

, x <1

Euler constant
=xeu( [Digits_Max] )
Returns the Euler-Mascheroni constant
(The same constan returned by xGm function)
Example
xeu()

= 0.57721566490153286060651209008

xeu(60) = 0.57721566490153286060651209008240243104215933593992359880576

64

Xnumbers Tutorial

Trigonometric Functions

Sin
xsin(a, [Digit_Max])
Returns the sine of the angle a
xsin(a) = sin(a)
The argument a, in radians, may be either a normal or an extended number.
xsin(1.5)

= 0.997494986604054430941723371141

Cos
xcos(a, [Digit_Max])
Returns the cosine of the angle a
xcos(a) = cos(a)
The argument a, in radians, may be either a normal or an extended number.
xcos(1.5) = 7.07372016677029100881898514342E-2

Computation of cos(/2)
Example: compute cos (89,99999995) with the standard built-in function COS function
COS(89.99999995) = COS(1.570796326) = 7.94896654250123E-10

The correct answer, accurate to 15 digits, is 7.94896619231321E-10

As we can see, only 7 digits are corrected. The remaining 8 digits are meaningless. On the
contrary, with the multiprecision function xcos(x) we have the correct result with all its
significant digits.
xcos(1.570796326)

7.94896619231321E-10

The table below shows the computation effect when a approaches /2

angle

(deg)

xcos()

COS() built-in

Err %

1.57
1.570
1.5707
1.57079
1.570796
1.5707963
1.57079632
1.570796326
1.5707963267
1.57079632679

89.95437383553930
89.95437383553930
89.99448088119850
89.99963750135470
89.99998127603180
89.99999846476560
89.99999961068120
89.99999995445590
89.99999999456290
89.99999999971950

7.96326710733325E-4
7.96326710733325E-4
9.63267947476522E-5
6.32679489657702E-6
3.26794896619225E-7
2.67948966192313E-8
6.79489661923132E-9
7.94896619231321E-10
9.48966192313216E-11
4.89661923132169E-12

7.96326710733263E-04
7.96326710733263E-04
9.63267947476672E-05
6.32679489666849E-06
3.26794896538163E-07
2.67948965850537E-08
6.79489670660314E-09
7.94896654250123E-10
9.48965963318629E-11
4.89658888522954E-12

7.75E-14
7.75E-14
-1.55E-13
-1.45E-11
2.48E-10
1.28E-09
-1.29E-08
-4.41E-08
2.41E-07
6.20E-06

As we can see, the accuracy of the standard function COS decreases when the angle
approaches the right angle. On the contrary, the xcos function keeps its accuracy.

65

Xnumbers Tutorial

Tan
xtan(a, [Digit_Max])
Returns the tangent of a
xtan(a) = tan(a)
The argument a, in radians, may be either a normal or an extended number.

Arcsine
xasin(a, [Digit_Max])
Returns the arcsine of a
xasin(a) = arcsin(a)
The arcsine is defined between -/2 and /2
The argument a, where | a | 1 , may be either a normal or an extended number.

Arccosine
xacos(a, [Digit_Max])
Returns the arccosine of a
xacos(a) = arccos(a)
The arccosine is defined between 0 and
The argument a, where | a | 1 , may be either a normal or an extended number.

Arctan
xatan(a, [Digit_Max])
Returns the arctan of a
xatan(a) = arctan(a)
The arctan(a) is defined between

/ 2 < arctan(a)< / 2

Constant
These functions return the following multiples of
xpi([Digit_Max])

xpi =

xpi2([Digit_Max])

xpi2 = /2

xpi4([Digit_Max])

xpi4 = /4

x2pi([Digit_Max])

x2pi = 2

The optional parameter Digit_Max, from 1 to 415, sets the number of significant digits (default
30).
Example. Compute the Hermite-Ramanujan constant with 36 significant digits

163

xexp(xmult(xpi(36),xsqr(163,36),36),36) = 262537412640768743.999999999999250005

66

Xnumbers Tutorial

Complement of right angle

xanglec(a, [Digit_Max])
Returns the complement of angle a to the right angle
where 0 /2 .

xanglec() = /2

Example:
xanglec(1.4) = 0.17079632679489661923132169163
For angles not too near the right angle this function is like the ordinary subtraction. The use of
this function is computing the difference without loss of significant digits when the angle is very
close to the right angle. For example, computing in Excel the following difference:
(PI()/2 1.570796) = 1.57079632679490 1.570796 = 0.00000032679490
we have a loss of 7 significant digits, even though the computation has been made with 15
significant digits. On the contrary, if we use:
xanglec(1.570796 , 15) =

3.26794896619231E-7

we get the full precision with 15 significant digits. The "lost" digits are automatically replaced

67

Xnumbers Tutorial

Polynomial Rootfinder
The roots of polynomials are of interest to more than just mathematicians. They play a central
role in applied sciences including mechanical and electrical engineering where they are used in
solving a variety of design problems.
Xnumbes provides several macros based on the following polynomial rootfinder algorithms.
RootFinder JT

Jenkins and Traub algorithm

(translated in VB from the original rpoly FORTRAN 77 subroutine)

RootFinder GN

Generalized Newton-Raphson method

RootFinder ADK

Aberth, Durand, Kerner algorithm

Rootfinder RF

Ruffini's method for real integer roots.

Rootfinder LB

Lin-Bairstow algorithm

Rootfinder SK

Siljak algorithm

Rootfinder LA

Laguerre algorithm

These macros are able to find, in a few seconds, all the roots - real or complex - of a dense
polynomial up to 15th - 20th degree, in double or multi-precision. It is remarkable that
sometimes the results has shown in an exact way, even if the computation is intrinsically
approximate.
The characteristics of each rootfinder are reassumed in the following table
Macro

Roots

Coefficients

Arithmetic

RootfinderJT

Complex

Real

Standard

RootfinderGN

Complex

Real

Multi-precision

RootfinderDK

Complex

Complex

Multi-precision

RootfinderRF

Real, Integer

Real, Integer

Multi-precision

RootfinderLB

Complex

Real

Standard

RootfinderSK

Complex

Complex

Multi-precision

RootfinderLA

Complex

Real

Standard

68

Xnumbers Tutorial

Input parameters
The polynomial rootfinder interface is simple and straight.

Method:
JT
GN
ADK
RF
LB
SK
LA

Jenkins-Traub
Gen. Newton-Raphson
Aberth-Durand-Kerner
Ruffini
Lin-Bairstow
Siljak
Laguerre

Standard
Multiprecision
Multiprecision
Standard
Standard
Multiprecision
Standard

Real coefficients
Real coefficients
Complex coefficients
Real Integer coefficients
Real coefficients
Complex coefficients
Real coefficients

Coefficients input:
is an array containing the polynomial coefficients with increasing
degree, from top to bottom. May be also a single cell containing the polynomial formula string,
such as:
-120+274x-225x^2+85x^3-15x^4+x^5
RootfinderDK and Siljak can also accept complex coefficients. In that case the input is an
(n x 2) array. Examples of possible input are (thick black box):

Remarks.
The formula string is more adapt for sparse polynomials.
Real coefficients can be put in horizontal or vertical vector. Complex coefficients, only in
vertical vectors
Results Output: It is the upper left corner of the output area. If blank, the routine assumes the
cell nearest the given coefficients range.
Error: Sets the relative roots accuracy. The algorithm terminates when the relative difference
between two iterations is less then this value.
Iter: The algorithm stops when the iterations counter reaches this value.
Multi-Precision: Enable/disable the multi-precision arithmetic
MP-out: If checked, the results are written in multi-precision, otherwise they are converted in
standard double precision.

69

Xnumbers Tutorial

Output
The rootfinder macros output their results in the following simplified layout
The roots and their estimated relative errors are written in a table starting from the left upper
cell indicated in the input window. In the right-bottom cell is written the total elaboration time in
seconds
Note: we have formatted the table
only for clarity. The macros do not
perform this task. We do it.

Integer Rootfinder output

The macro RootfinderRF outputs all integer roots of the polynomial (if any) at the left and the
coefficients of the remainder polynomial (deflated polynomial) at the right
Polynomial

RootfinderRF
Integer
roots

Reduced
Polynomial

This result means that the given polynomial

x 8 20 x 7 + 147 x 6 582 x 5 + 1510 x 4 328 x 3 6280 x 2 + 11904 x 8704

can be factorized as
( x + 2)( x 2)( x + 8) 2 ( x 4 4 x 3 + 23 x 2 38 x + 34)

70

Xnumbers Tutorial

How to use rootfinder macros

Using polynomial rootfinder macros is simple. Select the polynomial coefficients and start the
rootfinder that you prefer. All input fields are filled and the only work that you have to do in the
most cases - is to press "Run".
Example. Select the range C3:C11 and start the RootfinderJT . The coefficients input-box is
filled with C3:C11 and the result-output box is filled with the cell E3.

Press run and - after a while - the routine ends and the roots will be displayed at the right,
like in the following fugure

Sparse polynomials. We can pass to the rootfinder macros also symbolic polynomial string,
(that it is the faster way for sparse high degree polynomials). Lets see this example
Find all roots of the following 16th degree polynomial
x^16-6817x^8+1679616
Write this string in a cell, select it and start a rootfinder macro

71

Xnumbers Tutorial

In this case we have used the Durand-Kerner algorithm obtaining a very high accuracy
(practically the highest accuracy in standard double precision)

Root Error Estimation

The third column produced by the rootfinder macros is an estimation of the relative root error,
defined as:
eri =| xi ~
x | / | xi | for | xi |> 0
where

~
x

is the true unknown root and

xi

is the approximate root given by the rootfinder

We have to say that this number should be regarded as an estimation of goodness of the
root found; small values (for example 1E-9 , 1E-12 ) indicate an high precision of the
correspondent root. On the contrary, larger values (for examples 1E-3 , 1E-5) indicates a
difficult roots that require an extra investigation.
For example assume to find the root of the following 6th degree polynomial

Clustering effect: In this case, the accuracy is enough good, but quite lower than the previous
example. The reason is that the roots:
1, 1.01, 1.02, 1.03, 1.04, 1.05
are very close each other (0.1% of difference)

72

Xnumbers Tutorial

Complex polynomials. The macro RootfinderDK and RootfinderSK can solve also complex
polynomials. Example: find the roots of the following polynomial with complex coefficients

P( z ) = (12 + 4i ) + 4 z + (15 5i ) z 2 5 z 3 + (3 + i ) z 4 + z 5
Select both real and imaginary coefficients columns and start the macro RootfinderKD

The roots are z = 1 , z = 2 , z = 3 j . Observe that the results are shown in exact mode even
if the computation is intrinsically approximated.
A polynomial of n degree, having as roots
the first 1, 2,..n integers, belongs to the
Wilkinson class that, as known, is
ill-conditioned. This dense polynomial is
usually assumed as standard reference for
polynomial rootfinder algorithms. We have
tabulated the LRE (log relative error)
obtained with all the rootfinder macros.
As we can see, for a Wilkinson polynomial of
20th degree, we have good about four
significant digits (0.1% accuracy)

Wilkinson polynomial test

16.0
14.0

LRE

12.0
10.0
8.0
6.0
4.0
2.0
0.0
0

10

15

20

But all polynomials are so hard to solve? Fortunately not. Many polynomials with higher
degree, can be solved with good accuracy
For example, if we try to get all real roots of the 16th degree Legendres polynomial
L16(x) = 6435 -875160 x2 +19399380 x4 -162954792 x6+ 669278610 x8
-1487285800 x10 +1825305300 x12 -1163381400 x14 +300540195 x16

We have a general accuracy of more than 13 digits

Legendre polyn. Coeff.
Real
Imm
6435 -0.989400934991646
0 -0.944575023073157
-875160 -0.865631202387904
0 -0.755404408355024
19399380 -0.617876244402639
0 -0.458016777657228
-162954792 -0.281603550779259
0 -0.095012509837637
669278610
0.095012509837637
0
0.281603550779259
-1487285800
0.458016777657228
0
0.617876244402640
1825305300
0.755404408354981
0
0.865631202387767
-1163381400
0.944575023073325
0
0.989400934991655
300540195

73

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

Rel. Err.
2.9585E-17
1.6352E-13
3.673E-14
1.1559E-14
1.0779E-14
7.4625E-16
1.9313E-16
8.5038E-18
8.5038E-18
1.4485E-16
1.5356E-15
5.0196E-15
2.5615E-14
4.6793E-14
8.4139E-14
5.5583E-15

25

Xnumbers Tutorial
(remember that the higher degree coefficients are at bottom)
In the last column there are the errors estimation given by the rootfinder DK. They are slight
different from the true roots errors, but we have to remember that this column must be regard
as an index of the root approximation: low values mean a good accuracy, larger values often
(but not always) indicate a poor approximation

Integer roots
In applied science it's rarely to came across in polynomials having exact integer roots.
Nevertheless, they are frequent in math, didactical examples and algorithm testing .
Xnumbers has a dedicated special macro for finding the integer real roots of a polynomial. It
uses the Ruffini's method with the QD algorithm for roots isolation.
This method is generally less efficient then JT or DK but it can gain in accuracy.
The roots found by this method have no round-off errors so the deflated polynomial is exact.
Therefore, in that case, the root-finding-deflating process is without errors.
For polynomial having a mix of integer real roots, complex roots and real roots the method
returns the integer roots and the coefficients of the deflated polynomial that can be solved with
the aid of the general purpose macros: DK, GN or JT. Because the deflated polynomial has a
lower degree, the roots accuracy will be generally higher than by solving directly the given
polynomials.

original
P(x)

Deflated
Q(x)

Extract integer roots

(if exist)

find real and complex

roots (if exist)

Let's see how it works practically. Assume to have the following polynomial
degree
a0
a1
a2
a3
a4
a5
a6
a7
a8

coeff
8678880
-13381116
8844928
-3279447
746825
-107049
9437
-468
10

The exact roots are:

integer

real

complex

5, 6, 7, 8, 9

2.8

4.5 i 0.5

If we try to solve this 8th degree polynomial with a general

rootfinder, probably the best accuracy that we can obtain
is about 1e-10, that it is a good result but we can do better
if we extract the integer roots before and then, solving for
the remaining roots

Select the range B2:B10 and start the macro for extracting the integer roots and deflated
polynomial

74

Xnumbers Tutorial
The original polynomial is now cracked into the following factors

( x 5)( x 6)( x 7)( x 8)(10 x 3 118 x 2 + 457 x 574)

Now let's find the roots of the following 3rd degree polynomial by, for example, the general JT
rootfinder. We obtain:
Re

Im
2.8
4.5
4.5

Rel. Err.
0
0.5
-0.5

1.14E-17
1.44E-15
1.44E-15

The general accuracy is better than 1e-14 , thousand times than the direct method.
Clearly is a good thing to keep attention to the integer roots (when they are).
Global roots accuracy
versus the solving
methods:

1.0E+00
1.0E-02
1.0E-04
1.0E-06

Rootfinder RF + DK
Rootfinder JT
Rootfinder DK
Rootfinder NG

1.0E-08
1.0E-10
1.0E-12
1.0E-14
1.0E-16
Rootfinder Rootfinder Rootfinder Rootfinder
RF + DK
JT
DK
NG

The multiprecision should be used when the coefficients exceed 15 digits (remember that the
coefficients must be exact in order to extract the exact integer roots)
Let's see the following 18th degree polynomial having the roots
Polynomial roots

Coefficients
-612914519230813800000
91181999821816015800
-5186948337826516202
137665995531841931
-1622627967498318
6214402509219
-11208193158
10605849
-5122

integer

real

complex

25, 27,29, 31,

1000, 1001, 1002, 1003, 1004

none

none

Note that same coefficients have 16 - 18 significant digits and

they must be inserted as x-numbers, (that is as string) in
order to preserve the original precision.
We have also to set the multiprecision check-box in the
macro RootfinderRF

Note that this a so called "clustered polynomial" because some of its integer roots (1000, 1001,
1002, 1003, 1004) are very closed each other (difference less then 1%). This situation is quite
difficult for many algorithms and the accuracy is generally quite poor. On the contrary, the
Ruffini's method works very fine in that case.

75

Xnumbers Tutorial

Multiple roots
The macro "Factors" performs the decomposition of a polynomial with multiple roots into
smaller factors having all single roots. It uses the GCD method with the Euclid' algorithm
Example, the polynomial:

648 + 3132 x 6534 x 2 + 7737 x 3 5744 x 4 + 2779 x 5 878 x 6 + 175 x 7 20 x 8 + x 9

Has the roots x = 1 (m = 2), x = 2 (m = 3) and x = 3 (m = 4)
It can be decomposed into the product of the following factors

(3 + x)(6 5 x + x 2 )(6 + 11x 6 x 2 + x 3 ) 2

Using this macros is simple. Select the polynomial coefficients and start the macro. All the
input and output cells are automatically filled. Press Run

Each factors contains only single roots and thus can be solved with high precision by any
rootfinder macro1.

For further details see "Non Linear Equations", Foxes Team, 2006

76

Xnumbers Tutorial

Polynomial Functions

Polynomial evaluation
=POLYN(z, Coefficients, [DgtMax])
Computes the polynomial at the value z.

P ( z ) = a0 + a1 z + a2 z 2 + ...an z n
The parameter Coefficients is the (n+1) column vector containing the polynomial coefficients
from the lowest to the highest order.
This function accept also complex coefficients. In that case the parameter Coefficients is an
(n+1 x 2) array.
The optional parameter DgtMax set the precision. If omitted, the function works in the faster
double precision.
This function works also for complex arguments. In that case, z must be a complex number
(two adjacent cells) and the function returns two values. To see both real and imaginary part,
select two cells and give the CTRL+SHIFT+ENTER key sequence. If you press only ENTER,
the function returns only its real part.
Example: compute the following real polynomial

P( z ) = 2 z 4 + z 3 5 z 2 + 2 z + 4
for z = 4 2i

Otherwise, if you want to compute a real polynomial for a real argument, e.g. z = 10 - simply
pass a single value

77

Xnumbers Tutorial
Example: compute the following complex polynomial

P( z ) = 2 z 4 + (1 i ) z 3 5 z 2 + (2 i ) z + (4 5i )
for z = 4 2i

Polynomial derivatives
=DPOLYN(z, Coefficients, Order, [DgtMax])
Computes the polynomial derivative at the value z.

P( z ) = a0 + a1 z + a2 z 2 + ...an z n
D j ( z) =

d j P( z )
dz j

The parameter "Coefficients" is the (n+1) vector containing the polynomial coefficients from the
lowest to the highest order.
This function accept also complex coefficients. In that case the parameter Coefficients is an
(n+1 x 2) array.
The parameter "Order" sets the order of the derivative.
The optional parameter "DgtMax" set the precision. If omitted, the function works in the faster
double precision.
This function works also for complex arguments. In that case, z must be a complex number
(two adjacent cells) and the function returns two values. To see both real and imaginary part,
select two cells and give the CTRL+SHIFT+ENTER key sequence. If you press only ENTER,
the function returns only its real part.
Example. Compute the derivatives of the following polynomial

P( z ) = 3 + 2 z + z 2 + z 3
For z= 3, we have:

Example: calculate the 2nd derivative of the following complex polynomial at the point

z = 4 2i

P ( z ) = 2 z 4 + (1 i ) z 3 5 z 2 + (2 i ) z + (4 5i )

78

Xnumbers Tutorial

With DPOLYN and POLYN it is very easy to implement, for example, the Newtons algorithm
for finding the polynomial root with high precision
Example: find the real root of the following polynomial with Newtons algorithm

x 7 5 x 6 + 64 x 3 8000
The popular iterative Newtons formula is

xi +1 = xi +

p( xi )
p' ( xi )

Starting from the point x = 10. Note that we cannot use the handly x = 0 , because the
derivative is zero

The exact digits caught by the algorithm, are shown in blue. Note the impressive acceleration.
Try this example with 60 and more digits if you like.

Polynomial coefficients
=PolyTerms(Polynomial)
Returns the vector of the polynomial coefficients
The argument is a polynomial string like "1-3x+5x^2 +x^5" in any order.
Example

79

Xnumbers Tutorial

Note the braces { } in the formula. This indicates that the function return a vector. We must
select the range before enter the function with "shift+ctrl+enter".

Polynomial writing
=PolyWrite(Coefficients, [variable])
It returns the polynomial string from its coefficients.
The first argument may be a (1 x n) vector or an (2 x n) array. In the last case, the first row
indicates the coefficient position and the second row contains the correspondent coefficient
value.
The second optional argument specifies the variable string (default is "x").

Note that the second argument "t" must be insert as string, that is between quotes "..."

Polynomial addition
=PolyAdd(Poly1, Poly2)
Performs the addition of two polynomials. The arguments are monovariable polynomial strings.
Example:
PolyAdd("1-3x" , "-2-x+x^2")

"-1-4x+x^2" .

Polynomial multiplication
=PolyMult(Poly1, Poly2)
Performs the multiplication of two polynomials The arguments are monovariable polynomial
strings.
Example:
PolyMult("1-3x" , "-2+5x+x^2") =

"-2+11x-14x^2-3x^3" .

(1 3x )( 2 + 5 x + x 2 ) = 2 + 11x 14 x 2 3x 3

80

Xnumbers Tutorial

Polynomial subtraction
=PolySub(Poly1, Poly2)
Returns the difference of two polynomials. The arguments are monovariable polynomial strings.
Example:
PolySub("1-3x" , "-2+5x+x^2") = "3-8x-x^2" .

Polynomial division quotient

=PolyDiv(Poly1, Poly2)
Returns the quotient of two polynomials. The arguments are monovariable polynomial strings.
Example:
PolyDiv("x^4-1" , "x^2-x-1") = "2+x+x^2" .
In fact:

)(

x 4 1 = x 2 x 1 2 + x + x 2 + 1 + 3x

Polynomial division remainder

=PolyRem(Poly1, Poly2)
Returns the remainder of two polynomials
The arguments are monovariable polynomial strings.

81

Xnumbers Tutorial

Hermites and Cebychevs polynomials

By the basic operations we can build any other polynomial.
Example: Calculate the first 9th degree Cebychevs and Hermites polynomials
Cebysevs polynomials can be obtained by
the iterative formula

Hermites polynomials can be obtained by

the iterative formula

T0 = 1 , T1 = x
Tn +1 = 2 x Tn Tn 1

H 0 = 1 , H1 = x

H n +1 = 2 x H n 2n H n 1

The two iterative formulas can be arrange as:

=polysub(PolyMult("2x",Tn),Tn-1)
=polysub(PolyMult("2x",Hn),PolyMult(2*n,Hn-1))
These functions are inserted from the cell B4 to B9 and C5 to C9

Legendres Polynomials
Legendres polynomials can be obtained by the following well known iterative formula
Pn ( x) =

2n 1
n 1
x Pn 1 ( x)
Pn 2 ( x)
n
n

P0 = 1 , P1 = x

The first five polynomials are:

P0 = 1 , P1 = x , P2 =

1
1
1
3 x 2 1 , P3 = 5 x 3 3 x , P4 = 35 x 4 30 x 2 + 3
2
2
8

The above formula is very popular, but from the point of view of numeric calculus has one
disadvantage: its coefficients are decimal and this causes round-off errors leading inaccuracy
for higher polynomial degree. It is convenient to rearrange the iterative formula to avoid
fractional coefficients.
Lets assume that a Legendres polynomial can be written as

Pn ( x ) =

1
Ln ( x)
kn

(1a)

Where kn is an integer number and Ln(x) is a polynomial having integer coefficients

The Legendres polynomial Pn(x) is completely defined by the couple of (kn , Ln(x))
Starting with

82

Xnumbers Tutorial

k0 = 1

L0 = 1

k1 = 1

L1 = x

We can show that the following iterative process, with n 2 , gives the couples (kn , Ln(x))

U n ( x) = k n 2 ( 2n 1) x
an = k n 1 (n 1)
Vn ( x) = U n ( x) Ln 1 ( x) an Ln 2 ( x)
bn = n k n 1 k n 2

cn = GCD(bn , coef (Vn ) )

Where the coeff operator returns the coefficients vector of the polynomial Vn(x), and the GCD is
the greatest common divisor.
Simplifying, we get, finally the couple (kn , Ln(x))

kn =

bn
cn

Ln ( x) =

1
Vn ( x)
cn

This iterative algorithm, working only with integer values, is adapted to build Legendres
polynomials with high degree.
Lets see how to arrange a worksheet for finding Legendres polynomial
In the first column we insert the degree n, beginning from 0 to 2, for the moment
In the last two columns k and L(x) we have added the starting values.

The row 6 contains all the functions that the process needs.
In particularly we note:
The function polyterms(D6) gives the coefficients vectors [-1, 0, 3] of V(x) = -1+3x^2
The function xMCD returns the greatest common divisor of [-1, 0, 3, 2] 1
Select the row 6 and drag it down. We generate the Legendres polynomial in the form (1a)

83

Xnumbers Tutorial
Here is a table of Legendres polynomials obtained with the above method
n

L(x)

-1+3x^2

-3x+5x^3

3-30x^2+35x^4

15x-70x^3+63x^5

16

-5+105x^2-315x^4+231x^6

16

-35x+315x^3-693x^5+429x^7

128

35-1260x^2+6930x^4-12012x^6+6435x^8

128

315x-4620x^3+18018x^5-25740x^7+12155x^9

10

256

-63+3465x^2-30030x^4+90090x^6-109395x^8+46189x^10

11

256

-693x+15015x^3-90090x^5+218790x^7-230945x^9+88179x^11

12

1024

231-18018x^2+225225x^4-1021020x^6+2078505x^8-1939938x^10+676039x^12

13

1024

3003x-90090x^3+765765x^5-2771340x^7+4849845x^9-4056234x^11+1300075x^13

We can also extract a table of Legendres coefficients by the Polyterms() function

84

Xnumbers Tutorial

Polynomial shift
=PolyShift(Poly, x0)
Performs the polynomial translation to x0,
The argument "Poly" can be the polynomial strings or the vector of polynomial coefficients.
This function returns the coefficient vector of the translated polynomial.
If you select one cell, the output will be a polynomial string
Example. Given the polynomial:
188784918 47389623 x + 4952504 x 2 275809 x 3 + 8633 x 4 144 x 5 + x 6

substituting x with z+24, we have

18 + 9 z 16 z 2 z 3 9 z 4 + z 6

This function is useful for transforming polynomial for reducing the coefficients amplitude and
improving the precision of rootfinder methods. In this example we work with coefficients of two
maximum digits, instead of 9 digits. We note also that the second polynomial, having the
second coefficient = 0, is centered. His roots are the same of the given polynomial, translated
of 24, but can be factorize much better. In fact, we have

( z 2 z + 1)( z 2 + z + 2)( z 2 9)

Polynomial center
=PolyCenter(Coefficients)
Returns the center of the polynomial roots circle
The argument specifies the vector of the polynomial coefficients in the following order:

[a0 , a1 , a2 ... an ]
It can also be a polynomial string
if x1, x2, ...xn are roots of polynomial the center Bx is defined as:

Bx =

x0 + x1 + x2 ... + xn an 1
=
n
n

85

Xnumbers Tutorial

Polynomial roots radius

=PolyRadius(Coefficients)
Returns the approximated radius of the polynomial roots circle.
The argument is the vector of the polynomial coefficients in the following order:

[a0 , a1 , a2 ... an ]
It can also be a polynomial string
If zi are the roots of a polynomial, the radius is defined as:

R = max (| zi |)
i =1...n

.The circle of root is very useful for locating all the roots of a polynomial. For example, given the
following 9 degree polynomial.
degree
a0

coefficents
-3098250

a1

4116825

a2

-2427570

a3

916272

a4

-244674

a5

46934

a6

-6430

a7

608

a8

-36

a9

The center = 4 and the radius 6.8

We can draw the circle containing, with high probability, all polynomial roots
We know that the roots of this polynomial are:
x real
9

x imm
0

-6

-4

-3

We have to point out that this method is probabilistic. It means that, picking-up a random
polynomial, the most part of the roots are found inside the circle but it is also possible to find
same roots outside the circle with 1% of probability.
Example: compute the root circle of the polynomial:

x7-5x6+64x3-8000

86

Xnumbers Tutorial
radius 4.331
center 0.714
The roots are:
x real

x imm

-2.7429701

1.6132552

-2.7429701 -1.6132552
-0.4369651

3.2182957

-0.4369651 -3.2182957
3.17993518

2.2060806

3.17993518 -2.2060806
5

Polynomial building from roots

=PolyBuild(Roots, [Variable])
Builds a polynomial from its roots. Argument "Roots" is an (n x 2) array, contains the
polynomial roots. It can be an (n x 1) vector for real roots.
This function returns the coefficient vector of the polynomial.
If you select one cell, the output will be a polynomial string
Complex roots for real polynomial:

Multiple roots:

Complex roots for complex polynomial

If the complex roots are not symmetrical, the
polynomial has both real and imaginary part.

Zero roots .If you want a polynomial with

multiple zero roots, simply repeat many couple
[0, 0] as it needs.

This function returns the vector of polynomial coefficients if you select more than two vertical
cells. It is useful for higher degree polynomial

87

Xnumbers Tutorial

In this example we get the 10

coefficients of the 9th degree
polynomial having the 9 roots in the
range A2:B10.
Select the range E2:E11 and insert
the function PolyBuild with the
CTRL+SHIFT+ENTER sequence.

If complex roots are not conjugate,

the polynomial has complex
coefficients. This function returns also
the imaginary part of the coefficients.
Simply select the range D2:E6.

88

Xnumbers Tutorial

Polynomial building with multi-precision

PolyBuildCfx()
This macro generates the polynomial coefficients from the polynomial roots.
This macro works like the function PolyBuild except that it works in multi-precision. It is useful
for high degree polynomial, when the coefficients become longer than 15 digits.
For using this macro select the range
containing the roots.
Then, start the macro. Choose the digits
precision (default=30) and the range you want
to paste the coefficients (default is the range
at the right side of the roots range selected).

In the following table we have calculated the coefficient of the polynomial having as roots the
first 19 integer numbers. That is:
x1 = 1 , x2 = 2, x3 = 3 ,.... x19 = 19
Roots

PolybuildCfx (30 digits)

PolyBuild

Diff.

-121645100408832000

-121645100408832000

431565146817638400

431565146817638000

400

-668609730341153280

-668609730341153000

-280

610116075740491776

610116075740492000

-224

-371384787345228000

-371384787345228000

161429736530118960

161429736530119000

-40

-52260903362512720

-52260903362512700

-20

12953636989943896

12953636989943900

-4

-2503858755467550

-2503858755467550

10

381922055502195

381922055502195

11

-46280647751910

-46280647751910

12

4465226757381

4465226757381

13

-342252511900

-342252511900

14

20692933630

20692933630

15

-973941900

-973941900

16

34916946

34916946

17

-920550

-920550

18

16815

16815

19

-190

-190

As we can see there are a little difference (digits in red) between the exact coefficients
computed by this macro PolyBuildCfx (multiprecision arithmetic with 30 digits) and those
returned by the function PolyBuild (standard double precision).

89

Xnumbers Tutorial

Polynomial Solving
=PolySolve (Polynomial)
This function returns the roots of a given real polynomial using the Jenkins-Traub algorithm.

a0 + a1 x + a2 x 2 + ...an x n
The arguments can be a monovariable polynomial strings like "X^2+3x+2" or a vector of
coefficients
This function returns an (n x 2) array.
It uses the same algorithm of the RootfinderJT macro. It works fine with low-moderate degree
polynomials, typically up to 10th degree. For higher degree it is more convenient to use the
macro.
Example. Find all roots of the given 10 degree polynomial

Integer polynomial
=PolyInt(Polynomial)
This function returns a polynomial with integer coefficients having the same roots of the given
polynomial. This transformation is also know as "denormalization" and can be useful when the
coefficients of the normalized polynomial are decimal.
Example: Given thefollowing polynomial:
-0.44+2.82x-3.3x^2+x^3
To eliminate decimal coefficients we denormalize the polynomial
-22+141x-165x^2+50x^3 = PolyInt("-0.44+2.82x-3.3x^2+x^3")
Take care with the denormalization because the coefficients became larger and the
computation may lose accuracy. See the example below
The following polynomials have the same root x = 11/10:
Pb(x)

= -2.4024+10.1524x-17.1x^2+14.35x^3-6x^4+x^5

Pa(x)

= -6006+25381x-42750x^2+35875x^3-15000x^4+2500x^5
90

Xnumbers Tutorial

If we compute both polynomials for x = 11/10, with standard double precision we get:
Pa(1.1) = -2.664E-15
Pb(1.1) =

4.547E-12

As we can see, the first value, obtained by the decimal polynomial, is 1000 times more precise
then the one obtained by the integer polynomial

Polynomial System of 2nd degree

=SYSPOLY2(Poly1, Poly2)
Solves a system of two 2nd degree polynomials.

a11 x 2 + a12 xy + a13 y 2 + a14 x + a15 y + a10 = 0

a21 x 2 + a22 xy + a23 y 2 + a24 x + a25 y + a20 = 0

It returns a (4 x 4) array containing the four solutions.
The parameters "Poly1" and "Poly2" can be coefficients vectors or polynomials strings
The coefficients must be passed in the same order of the above equation.
Polynomial strings, on the contrary, can be written in any order. Examples of 2nd degree x-y
polynomials strings are:
13+x+y^2-y+x^2+2x*y
x^2 + y^2 - 10
4x^2+8x*y+y^2+2x-2
Note: the product symbol * can be omitted except for the x*y mixed term
A 2nd degree system can have up to four solutions. It can also have no solution (impossible) or
even infinite solutions (undetermined). The function returns #N/D if a solution is missing
Example: solve the following system

x 2 + 2 xy + y 2 + x y = 0
2
x + y 2 10 = 0
Using SYSPOLY2 the solutions real or
complex can be obtained in a very quick
way
Real solutions represent the intersection
point of the curve poly1 and poly2.
They are: P1 = (3 , 1) , P2 = (1 , 3)
The system has also two complex solutions that have not a geometrical representation
P3 = (2.5 +j 1.118034 , 2.5 +j 1.118034) , P4 = (2.5 j 1.118034 , 2.5 j 1.118034)
The degree of the given system is 4
Example: solve the following system

xy 1 = 0

2
2 xy + y + x + y 1 = 0

91

Xnumbers Tutorial
The apparent degree of the system is 2 x 2 = 4

As we can see, the function SYSPOLY2 returns only three solutions: one real and two
complex.
P1 = (1, 1) ,

P2 = (j , j ) ,

P3 = = ( j , j )

Thus, the actual system degree is 3.

Bivariate Polynomial
=POLYN2(Polynomial, x, y, [DgtMax])
Returns the value - real or complex - of a bivariate polynomial P(x, y).
The parameter "Polynomial" is an expression strings. Valid examples are:
13+x+y^2-y+x^2+2x*y , x^2+y^2-10 , 8x*y+y^2+2x-2 , 10+4x^6+x^2*y^2
Note: the product symbol * can be omitted except for the x*y mixed terms
The third optional parameter DgtMax sets the multiprecision. If missing, the computation is
performed in faster double precision.
The variables x , y can be real or complex. The function can return real or complex numbers.
Select two cells if you want to see the imaginary part and give the CTRL+SHIFT+ENTER
sequence
Example: Compute the polynomial

P = x 2 + 2 xy + y 2 + x y
at the point
x = (2.5 + j 1.11803398874989)
y = (2.5 + j 1.11803398874989)
Verify that it is a good approximation of the
polynomial root

92

Xnumbers Tutorial

Partial fraction decomposition

Partial fraction decomposition is the process of rewriting a rational expression as the sum of a
quotient polynomial plus partial fractions..

N ( x)
R( x)
= Q ( x) +
= Q( x) + Fi
D( x)
D( x)
where each Fi is a fractions of the form

Am
A1
A2
+
+ ...
2
x + p ( x + p)
( x + p) m
or

B x + Cm
B1 x + C1
B x + C2
+ 22
+ ... 2 m
2
2
x + bx + c ( x + bx + c)
( x + bx + c) m

being m is the multiplicity of the correspondent root

The denominators D(x) is determined from the poles of the fractions itself. In fact, p is a real
root of D(x), while the quadratic factor can be obtained from the complex root using the
following relation

i b = 2 , c = 2 + 2

(1)

Many calculators and computer algebra systems, are able to factor polynomials and split
rational functions into partial fractions. A solution can also be arranged in Excel with the aid of
Xnumbers functions. Let's see
Real single poles. Find the fraction decomposition of the following rational fraction

N ( x) 3 x 3 + 276 x 2 1433 x + 1794

=
D( x) x 4 15 x 3 + 65 x 2 105 x + 54
First of all, we try to find the roots of the denominator using, for example, the function
polysolve. We find that the roots are pi = [1, 2, 3, 9]. They are all real with unitary multiplicity,
therefore the fraction expansion will be

A3
N ( x)
A1
A2
A4
=
+
+
+
D( x) x + p1 x + p2 x + p3 x + p4
where pi are the roots and Ai are unknown
Several methods exist for finding the fraction coefficients Ai . One of the most straight and
elegant is due to Heaviside that, for a real single root, simply states:

Ai =

N ( pi )
D' ( pi )

where D'(x) is the derivative of D(x)

A possible arrangement in Excel is the following

93

Xnumbers Tutorial

Therefore, the searched decomposition is

3x 3 + 276 x 2 1433 x + 1794

40
8
5
40
=
+

+
4
3
2
x 15 x + 65 x 105 x + 54
x +1 x + 2 x + 3 x + 9
You can prove yourself that this expression is an identity, thus always true for every x, except
the poles.
Complex single poles. Find the fraction decomposition of the following rational fraction

x 3 21 x 2 + 52 x + 123
N ( x)
= 4
D( x) x 2 x 3 29 x 2 42 x + 650
First of all, we try to find the roots of the denominator using, for example, the function
polysolve. We find that the roots are p = { 5 2i , 4 3i } . They are complex with unitary
multiplicity, therefore the fraction expansion will be

N ( x)
B x + C1
B x + C2
= 2 1
+ 2 2
D ( x) x + b1 x + c1 x + b2 x + c2
where bi and ci , calculated by the (1), are b1 = 10 , c1 = 26 , b2 = 8 , c 2 = 25
The coefficients Bi and Ci are unknown. For solving them we used here the so called
undetermined coefficients method
Renamed, for simplicity:

D1 ( x) = x 2 + b1 x + c1 , D2 ( x) = x 2 + b2 x + c2
The fraction expansion may be rewritten as

N ( x)
B x
C1
B x
C2
= 1 x+
+ 2 +
D ( x) D1 ( x )
D1 ( x) D2 ( x ) D2 ( x)
Giving 4 different values to x, the above relation provides 4 linear equations with the unknown
B1, C1, B2, C2, that can be easily solved. We can choose any value that we want; for example xi
= { 0, 1, 2, 3 } and we get the following linear system
0
1/17
1/5
3/5

1/26
1/17
1/10
1/5

0
1/34
2/45
3/58

1/25
1/34
1/45
1/58

B1
C1
B2
C2

123/650
9/34
3/10
63/290

Solving this linear system by any method that we like, for examply by SYSLIN , we get the
solution
[B1 , C1 , B2 , C2] = [-2 , 7 , 1 , -2]
Substituting these values, we have finally the fraction decomposition
94

Xnumbers Tutorial

x 3 21 x 2 + 52 x + 123
x2
2x + 7
= 2
+ 2
4
3
2
x 2 x 29 x 42 x + 650 x 10 x + 26 x + 8 x + 25
You can prove yourself that this expression is always valid for any compatible value of x
A possible arrangement for solving this
problem In Excel is a bit more
complicated then the previous one.
Let's see. First of all we compute the
roots with the function Polysolve; then
we compute the trinomials D1(x) and
D2(x) by the formulas (1)

Then we compute the polynomials N, D, D1, D2 for each values of x by the function polyn. We
get the 4x5 table visible at the right

From this table we get the complete system matrix in the following way.

The 4 x 4 linear system can be solved by any method that you want. For example by matrix
inversion as shown in the example.

95

Xnumbers Tutorial

Orthogonal Polynomials
Orthogonal polynomials are a class of polynomials following the rule:
b

w( x) p

( x) pn ( x) dx = mn cn

Where m and n are the degrees of the polynomials, w(x) is the weighting function, and c(n) is
the weight. m n is the Kronecker's delta function being 1 if n = m and 0 otherwise.
The following table synthesizes the interval [a, b], the w(x) functions and the relative weigh c(n)
for each polynomials family
polynomial

interval w(x)

Chebyshev polynomial of the

first kind

[-1, 1]

(1 x 2 ) 1/ 2

for n = 0

/ 2 for n 0

Chebyshev polynomial of the

second kind

[-1, 1]

(1 x 2 )1/ 2

/2

(1 x 2 ) 1/ 2

21- 2 (n + 2 )
for 0
n!( n + ) 2 ( )

[-1, 1]

Gegenbauer polynomial

2 / n 2 for = 0
2

2 n n!

Hermite polynomial

(-,+)

e x

Jacobi polynomial

(-1, 1)

(1 x) (1 + x)

hn

Laguerre polynomial

[0, +)

e x

generalized Laguerre
polynomial

[0, +)

xk e x

(n + k )!
n!

Legendre polynomial

[-1, 1]

2
2n + 1

Where

hn =

2 + +1
(n + + 1) (n + + 1)
2n + + + 1
n!(n + + + 1)

96

Xnumbers Tutorial

Orthogonal Polynomials evaluation

This set of functions1 calculate the orthogonal polynomials and their derivatives at the given
point . They return two values: the first one is the polynomial value, the second is its 1st
derivative. If you want to see both values select two adjacent cells and give the
CTRL+SHIFT+ENTER sequence. If you give ENTER, you will get only the polynomial value
Function Poly_ChebychevT(x, n)

Chebychev polynomial of the first kind

Function Poly_ChebychevU(x, n)

Chebychev polynomial of the second kind

Function Poly_Gegenbauer(a, x, n)

Gegenbauer polynomial

Function Poly_Hermite(x, n)

Hermite polynomial

Function Poly_Jacobi(a, b, x, n)

Jacobi polynomial

Function Poly_Laguerre(x, n, m)

Laguerre generalized polynomial

Function Poly_Legendre(x, n)

Legendre polynomial

Example: Tabulate the Legendre polynomial of 6th degree, for 0 x 1, with step h = 0.1

As we can see we have insert Poly_Legendre as a standard function, because in this

exercise we do not need the derivative information
Example. Find the greatest zero of the 5th degree Legendre polynomial
We can use the Newton-Raphson method,
starting from x = 1, as shown in the
worksheet arrangement.
Both polynomial and derivative are obtained
from the Poly_Legendre simply selecting
the range B5:C5 and pasting the function as
array with CTRL+SHIFT+ENTER sequence.
The other cells are filled simply by dragging
down the range B5:C5

Many thanks to Luis Isaac Ramos Garcia for his great contribution in developing this software

97

Xnumbers Tutorial

Function Poly_ChebychevT(x, [n])

Function Poly_ChebychevU(x, [n])
Evaluate the Chebychev orthogonal polynomial of 1st and 2nd kind
Parameters:
x (real) is the abscissa,
n (integers) is the degree. Default n = 1
Function Poly_Gegenbauer(L, x, [n])
Evaluate the Gegenbauer orthogonal polynomial of 1st and 2nd kind
Parameters:
x (real) is the abscissa,
n (integers) is the degree. Default n = 1
L (real) is the Gegenbauer factor and must be L < 1/2
Function Poly_Hermite(x, [n])
Evaluate the Hermite orthogonal polynomial of 1st and 2nd kind
Parameters:
x (real) is the abscissa,
n (integers) is the degree. Default n = 1
Function Poly_Jacobi(a, b, x, [n])
Evaluate the Jacobi orthogonal polynomial of 1st and 2nd kind
Parameters:
x (real) is the abscissa,
n (integers) is the degree. Default n = 1
a (real) is the power of (1-x) factor of the weighting function
b (real) is the power of (1+x) factor of the weighting function
Function Poly_Laguerre(x, [n], [m])
Evaluate the Laguerre orthogonal polynomial of 1st and 2nd kind
Parameters:
x (real) is the abscissa,
n (integers) is the degree. Default n = 1
m (integer) is the number of generalized polynomial. Default m = 0
Function Poly_Legendre(x, [n])
Evaluate the Legendre orthogonal polynomial of 1st and 2nd kind
Parameters:
x (real) is the abscissa,
n (integers) is the degree. Default n = 1

98

Xnumbers Tutorial

Weight of Orhogonal Polynomials

This set of functions calculate the weight c(n) of each orthogonal polynomial p(x, n)
b

cn = w( x)[ pn ( x)] dx
2

Function Poly_Weight_ChebychevT(n)

Chebychev polynomial of the first kind

Function Poly_Weight_ChebychevU(n)

Chebychev polynomial of the second kind

Function Poly_Weight_Gegenbauer(n, l)

Gegenbauer polynomial

Function Poly_Weight_Hermite(n)

Hermite polynomial

Function Poly_Weight_Jacobi(n, a, b)

Jacobi polynomial

Function Poly_Weight_Laguerre(n, m)

Laguerre generalized polynomial

Function Poly_Weight_Legendre(n)

Legendre polynomial

If we divide each orthogonal polynomial family for the relative weight we have an orthonormal
polynomial family

Zeros of Orthogonal Polynomials

This macro finds all roots of the most popular orthogonal polynomials
Its use is very easy.
Simply start the Zero macro from the menu
"tools > Ortho-polynomials..."
Choose the family and the degree that you
want and fill the optional parameters
Then press OK

This is an example of output for a

Laguerre polynomial of 6th degree
( m = 0)
Note: the format is added for clarity.
The macro does not do this

99

Xnumbers Tutorial

Coefficients of Orthogonal Polynomials

This macro calculate the coefficients of the most common orthogonal polynomials
Its use is very easy.
Simply start the Coeff macro from the menu
"tools/Ortho-polynomials..."
Choose the family and the degree that you
want and fill the optional parameters.
Then, press OK

This macro return also the polynomial weight

This is an example of output for a Laguerre
polynomial of 4th degree (m = 0)
The ortho-polynomial can be written as

L6 ( x) =

100

1 4
x 16 x 3 + 72 x 2 96 x + 24
24

Xnumbers Tutorial

Complex Arithmetic and Functions

Xnumbers provides a large collection of complex functions
Complex Hyperbolic Sin
Complex Hyperbolic Cos
Complex Hyperbolic Tan
Complex Inverse Hyperbolic Cos
Complex Inverse Hyperbolic Sin
Complex Inverse Hyperbolic Tan
Complex digamma
Complex Exponential Integral
Complex Error Function
Complex Complem. Error Function
Complex Gamma Function
Complex Logarith. Gamma Function
Complex Zeta Function
Complex Quadratic Equation
Complex Expression Evaluation

Complex Addition
Complex Subtraction
Complex Multiplication
Complex Division
Polar Conversion
Rectangular Conversion
Complex absolute
Complex power
Complex Root
Complex Log
Complex Exp
Complex inv
Complex negative
Complex conjugate
Complex Sin
Complex Cos
Complex Tangent
Complex Inverse Cos
Complex Inverse Sin
Complex Inverse Tan

How to insert a complex number

For definition a complex number is an ordered couple of numbers: (a,b)
In Excel a couple of numbers is represented by two vertical or horizontal adjacent cells and the
complex number (a, b) is a range of two cells. The figure below shows both vertical and
horizontal representations:
(234 , 105) in range "B7:C7" and in range "B2:B3"
( -100 , 23) in the range "E7:F7" and in range "D2:D3"

For entering complex functions you must select two cells, insert the complex function and give
the CTRL+SHIFT+ENTER keys sequence. If you press the ENTER key, the function returns
only the real part of the complex number.
Symbolic rectangular format
Xnumbers support the format "x+jy " only in expression strings passed to the function cplxeval.
Except this case, you must always provides a complex number as a couple of real numbers
(one or two cells).
The reason for this choice is that the rectangular format is more adapt for symbolic calculation
while the array format is more convenient for numerical computation characterized by long noninteger numbers.
But, of course, you can convert a complex number (a,b) into its symbolic format "a+jb" by the
Excel function COMPLEX, as shown in the following example
101

Xnumbers Tutorial

XNUMBERS has two sets of complex functions: for standard double precision (prefixed by
"cplx") and for multiprecision (prefixed by "xcplx").

Complex Addition
xcplxadd(a, b, [Digit_Max])
cplxadd(a, b)
Performs the complex addition:

(a 1 , a 2 ) + (b1 , b 2 ) = (a 1 + a 2 , b1 + b 2 )
Complex Subtraction
xcplxsub(a, b, [Digit_Max])
cplxsub(a, b)
Performs the complex subtraction.

(a 1 , a 2 ) (b1 , b 2 ) = (a 1 a 2 , b1 b 2 )
Complex Multiplication
xcplxmult(a, b, [Digit_Max])
cplxmult(a, b)
Performs the complex multiplication:

(a 1 , a 2 ) * (b1 , b 2 ) = (a 1 b1 - a 2 b 2 , a 1 b 2 + a 2 b1 )
Complex Division
xcplxdiv(a, b, [Digit_Max])
cplxdiv(a, b)
Performs the complex division

(a 1 , a 2 ) a 1 b1 + a 2 b 2 a 2 b1 a 1 b 2
=
,
(b1 , b 2 ) b12 + b 22
b12 + b 22
102

Xnumbers Tutorial

Polar Conversion
xcplxpolar(z, [angle], [Digit_Max])
cplxpolar(z, [angle])
Converts a complex number from its rectangular form to the equivalent polar form.
The optional parameter angle sets the angle unit (RAD, DEG) (default RAD).
x

(deg)

1
0.866025
0.707107
0.5
0
-0.5
-0.70711
-0.86603
-1
-0.86603
-0.70711
-0.5
0
0.5
0.707107
0.866025

0
0.5
0.707107
0.866025
1
0.866025
0.707107
0.5
0
-0.5
-0.70711
-0.86603
-1
-0.86603
-0.70711
-0.5

1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1

0
30
45
60
90
120
135
150
180
-150
-135
-120
-90
-60
-45
-30

(x, y) ( , )
Where

= x2 + y 2
y
x

= atan , x > 0
= sgn( y )

, x=0
2
, y = 0, x < 0

=
y
atan x + sgn( y ) , y 0, x < 0

Rectangular Conversion
xcplxrect(z, [angle], [Digit_Max])
cplxrect(z, [angle])
Converts a complex number from its polar form to the equivalent rectangular form. The optional
parameter angle sets the angle unit (RAD, DEG) (default RAD).
( , ) (x, y)
Where

x = cos( )
y = sin ( )

Complex absolute
xcplxabs(z, [Digit_Max])
cplxabs(z)
Returns the absolute value of a complex number

| z |=

z12 + z 22

103

Xnumbers Tutorial

Complex power
xcplxpow(z, [n], [Digit_Max])
cplxpow(z, [n])
Returns the nth integer power of a complex number zn (default n = 2)

z n = ( x + iy ) = n e n
n

Where

= x2 + y 2

x
y

= atan

Complex Roots
xcplxroot(z, [n], [Digit_Max])
cplxroot(z, [n])
Returns all the nth roots of a complex extended number z1/n (default n = 2)
The function returns a matrix of (n x 2) values. Remember to press the sequence
CTRL+SHIFT+ENTER for insert properly this function.
The root of a complex number is computed by the De Moivre-Laplace formula.
n

z=

x + iy =

+ 2k
+ 2k
+ i sin
n
n

cos

, k = 0,1...n 1

where

= x2 + y 2

x
y

= atan

Note: If you select only one row, the function return only the first complex root (given for k = 0).
Example: compute all the 3 complex cubic roots of the number z = 8

104

Xnumbers Tutorial

Complex Log
xcplxLn(z, [Digit_Max])
cplxLn(z)
Returns the natural logarithm of a complex number

log(z ) = log(x + iy) = log( ) +

Where:

= x2 + y2

y
x

= atan

Complex Exp
xcplxExp(z, [Digit_Max])
cplxExp(z)
Returns the exponential of a complex number

e z = e x +iy = e x cos( y ) + ie x sin( y )

Complex inverse
xcplxinv(z, [Digit_Max])
cplxinv(z)
Returns the inverse of a complex number

1
1
x
y
=
= 2
i 2
2
z x + iy x + y
x + y2

Complex negative
xcplxneg(z)
cplxneg(z)
Returns the complex negative

z = (x + iy ) = x iy

Complex conjugate
xcplxconj(z)
cplxconj(z)

105

Xnumbers Tutorial
Returns the conjugate of a complex number

z = x + iy = x iy

Complex Sin
=cplxsin(z)
Returns the sine of a complex number

Complex Cos
cplxcos(z)
Returns the cosine of a complex number

Complex Tangent
cplxtan(z)
Returns the tangent of a complex number

Complex ArcCos
cplxacos(z)
Returns the arccosine of a complex number

Complex ArcSin
cplxasin(z)
Returns the arcsine of a complex number

Complex ArcTan
cplxatan(z)
Returns the arctangent of a complex number

Complex Hyperbolic Sine

cplxsinh(z)
Returns the hyperbolic sine of a complex number
Parameter z can be a real or complex number (two adjacent cells)

106

Xnumbers Tutorial

Complex Hyperbolic Cosine

cplxcosh(z)
Returns the hyperbolic cosine of a complex number
Parameter z can be a real or complex number (two adjacent cells)

Complex Hyperbolic Tan

cplxtanh(z)
Returns the hyperbolic tangent of a complex number

Complex Inverse Hyperbolic Cos

cplxacosh(z)
Returns the inverse of the hyperbolic cosine of a complex number

Complex Inverse Hyperbolic Sin

cplxasinh(z)
Returns the inverse of the hyperbolic sine of a complex number

Complex Inverse Hyperbolic Tan

cplxatanh(z)
Returns the inverse of the hyperbolic tangent of a complex number

Complex digamma
cplxdigamma(z)
Returns the logarithmic derivative of the gamma function for complex argument.

( x) =

' ( x)
d
ln ( ( x ) ) =
dx
( x)

Complex Exponential Integral

cplxei(z)
Returns the exponential integral of a complex number

et
dt
x t

Ei ( x) =

107

Xnumbers Tutorial

Complex Error Function

cplxerf(z)
Returns the "error function" or "Integral of Gauss's function" of a complex number

erf ( z ) =

t 2

dt

Complex Complementary Error Function

cplxerfc(z)
Returns the complementary error function for a complex number

erfc( z ) = 1 erf ( z )

Complex Gamma Function

cplxgamma(z)
Returns the gamma function for a complex number

( x) = t x 1 e t dt
0

Complex Logarithm Gamma Function

cplxgammaln(z)
Returns the natural logarithm of the Gamma function for a complex number

Complex Zeta Function

cplxzeta(z)
Returns the Riemann zeta function (s) for a complex number. It is an important special
function of mathematics and physics which is intimately related with very deep results
surrounding the prime number, series, integrals, etc.
For |s|>1 the function is defined as:

108

Xnumbers Tutorial

Complex Quadratic Equation

cplxEquation2(a, b, c, [DgtMax])
Returns the multiprecision solution of the quadratic equation with complex coefficients

a z2 + b z + c = 0
where a, b, c are complex
The solutions are found by the resolution formula

z=

b
b 2 4ac

2a
2a

This function returns an (2 x 2) array

The optional parameter DgtMax, from 1 to 200, sets the number of the significant digits. If
missing, the computation is in standard double precision.
Example: Find the solution of the following complex equation with 20 digits precision

z 2 + (9 2 i ) z + 4 + i = 0

Of course the function can also works with real coefficents equations

z2 + 2 z + 3 = 0

109

Xnumbers Tutorial

Number Theory
Maximum Common Divisor
xMCD(a1, [a2])
MCD(a1, [a2], [a3]...)
Returns the Maximum Common Divisor (also called Greatest Common Divisor, GCD) of two or
more extended numbers
The arguments "a1" and "a2" may be single numbers or arrays (range). At least, two values
must be input. If "a1" is a range, "a2" may be omitted

Minimum Common Multiple

xMCM(a1, [a2])
MCM(a1, [a2], [a3]...)
Returns the Minimum Common Multiple (also Least Common Multiple, LCM) of two or more
extended numbers
The arguments "a1" and "a2" may be single numbers or arrays (range). At least, two values
must be input. If "a1" is a range, "a2" may be omitted
Example

Note that LCM may easily overcome the standard precision limit even if its arguments are all in
standard precision.

110

Xnumbers Tutorial

Rational Fraction approximation

xFract(x, [Digit_Max])
Fract(x, [ErrMax])
Returns the fractional approximation of a non-integer number x, the functions returns a vector
of two integer numbers, numerator N and denominator D :
xN/D
The optional parameter ErrMax sets the accuracy of the fraction conversion (default=1E-14).
The function tries to calculate the fraction with the maximum accuracy possible.
The algorithm uses the continued fraction expansion1

N 0 = 0 , N1 = 1

Where ai are found by the following algorithm:

D0 = 1 , D1 = 0

ai +1 = int ( xi / yi )

N i +1 = ai N i + N i 1

xi +1 = yi

Di +1 = ai Di + Di 1

yi +1 = xi yi ai +1

In the example below we want to find

the fraction form of the number 0.126.
The function returns the solution:
N = 63 , D = 500
0.126 63 / 500
Often the rational form is not so easy to find, and depends strongly on the precision we want to
reach.
See, for example, the fractions that approximate 2 with increasing precision
Digit

N/D

Error

1.500000000000000

0.08579

1.400000000000000

0.01421

41

29

1.413793103448280

0.00042

99

70

1.414285714285710

7.2E-05

239

169

1.414201183431950

1.2E-05

1393

985

1.414213197969540

3.6E-07

3363

2378

1.414213624894870

6.3E-08

8119

5741

1.414213551646050

1.1E-08

10

47321

33461

1.414213562057320

3.2E-10

11

114243

80782

1.414213562427270

5.4E-11

12

275807

195025

1.414213562363800

9.3E-12

13

1607521

1136689

1.414213562372820

2.8E-13

14

3880899

2744210

1.414213562373140

4.2E-14

15

9369319

6625109

1.414213562373090

1.3E-14

You can regulate the desiderate approximation with the parameter ErrMax

from The art of Computer Programming, D.E.Knuth, Vol.2, Addison-Wesley, 1969

111

Xnumbers Tutorial

Continued Fraction
xFractCont(x, [Digit_Max])
FractCont(x)
These functions perform the continued fraction expansion of a decimal number x
Usually the continued fraction is written in the
compact form x = [a0, a1, a2, a3...]
Each step gives the rational fraction fi = si / ti ,
where the fractions fi converge to the number x

x = a0 +

1
a1 +

1
a2 +

1
a3 + ....

Both the above functions return an array containing the integer values [ai, si, ti].
The algorithm stops when the fraction approximates x with the maximum precision allowed,
that is 10^-15 and 10^-(DgtMax) respectively for the standard and the multiprecision function.
Example. Compute the continued fraction
expansion of the number e2 = 7.38905...
with standard precision (error < 1E-15)
As we can see, the continued fraction of the
transcendental real number e2 can be written
as:
e2 = [7, 2, 1, 1, 3, 18, 5, 1, 1, 6, 30, 8...]
while the correspondent approximating
fractions are
e2
e2
e2
e2
....

15/2 (1.5%),
22/3 (0.75%),
37/5 (0.15%),
133/18 (0.002%),

Sometime the continued fraction expansion requires the multiprecision. Expand, for example,
the Ramanujian number
digits

58

24591257751.9999998222132414696... with 30 significant

As we can see, the last fraction: 101111070168039846853 / 4111667292 approximates

58

with a relative error of about 1E-30

112

Xnumbers Tutorial

Continued Fraction of Square Root

FractContSqr(n)
This function returns the continued fraction
expansion of the square root of an integer
number

n = a0 +

n = [a0 , a1 , a2 , a3 ...2a0 ]

1
a1 +

1
a2 +

1
a3 + ....

As known, the expansion is infinite and periodic.

The algorithm stops when it come across into a term double of the first term a0
This function returns the (m+1) vector [a0, a1, a2,...am ] where am = 2*a0
Use the ctrl+shift+enter sequence.

77

Example. Calculate the continued fraction of

Therefore the infinite, periodic continued fraction is

77 = [8, 1, 3, 2, 3, 1, 16, 1, 3, 2, 3, 1, 16, 1, 3, 2, 3, 1, 16....]

Note that the function returns only the first terms + the period.

Check Prime
Prime(n)
Prime(n) =

"P" if n is prime, or the lowest factor if n is not prime

Returns "?" if the function is not able to check it.

Example
prime(134560093) ="P"
prime(134560079)= 89

Next Prime
NextPrime(n)
This function1 returns the prime number greater than n or "?" if the function is not able to
calculate it
nextprime(9343560093) = 9343560103

113

Xnumbers Tutorial

Modular Addition
xaddmod(a, b, m)
Performs the modular addition1

(a + b) (mod m)
where a, b integer and m positive integer

Modular Subtraction
xsubmod(a, b, m)
Performs the modular subtraction

(a b) (mod m)
where a, b integer and m positive integer

Modular Multiplication
xmultmod(a, b, m)
Performs the modular multiplication

(a b) (mod m)
where a, b integer and m positive integer

Modular Division
xdivmod(a, b, m)
Performs the modular division

(a / b) (mod m)
where a, b integer and m positive integer
Remember that the modular division is always possible if, and only, the module m is prime.
Otherwise the division could be impossible. In that case the function returns "?"

Modular Power
xpowmod(a, p, m)
Performs the modular integer power of ap
That is defined as the remainder of the integer division of ap by m

a p
r = a m
m
p

Modular function xaddmod, xsubmod, xmultmod, xdivmod appear thanks to the courtesy of John Jones

114

Xnumbers Tutorial
Example: compute

324 (mod 9005)

xpowmod(3,24,9005) = 3306
It's easy to prove that

324 (mod 9005) = 282429536481 (mod 9005) = 3306

When the number a or p become larger it is impossible to compute the integer power directly.
But the function xpowmod can return the correct result.
Examples: compute

123939040 (mod 3001)

It would be impossible to compute all the digits of this power. Using multiprecision we have
xpow(12,3939040) =

1.24575154970238125896669174496E+4250938

This result shows that 123939040 has more then 4 million of digits!
Nevertheless the remainder of this impossible division is
xpowmod(12, 3939040,3001) = 947
Examples: Miscellanea of modular function

Observe that m = 1000453 is prime.

115

Xnumbers Tutorial

Perfect Square
xIsSquare(n)
Checks if a number n is a perfect square
xisSquare(1000018092081830116) = TRUE
Because: 1000018092081830116 = 10000090462
xisSquare(2000018092081830116) = FALSE

Check odd/even
xIsOdd (n)
Checks if a number n is odd (TRUE) or even (FALSE)

Check Integer
xIsInteger(x)
Checks if a number x is integer

116

Xnumbers Tutorial

Macro - Factorize
Factorize()
This macro factorize a number of active cell returning the list of prime number and their
exponents

n = p1e1 p2e 2 p3e3 ... pkek

Select the cell contains the number you want to factorize and the run the macro Factorize from
the menu or Prime... > Factorize

Example. Assume to have in the cell A2 the following extended number

13320493696756339619246691430

Click copy if you want to copy the list in the worksheet, starting from the cell just below cell
A2.
This macro uses the trial division method with the prime table generated by the Eratostene's
sieve algorithm
This method is adapt for numbers having factors no more that 6 digts max. For higher factor
the elaboration time becames extremely long and it cames usefull a second factorization
method, the so called Pollard rho algorithm, for craking a number into two lower factors (not
necessary prime). Each factors, if not prime, can be factorized separately with the trial division
method.
Example. The number
18446744073709551617 = 274177 * 67280421310721
can be factorized with both methods: it requires about 33 sec with trial division; but less then 3
sec with Pollard method
The following number instead can be factorize only with Pollard method (about 40 sec).
10023859281455311421 = 7660450463 * 1308520867
Note that in this case both factors have 10 digits. The factos are prime so the factorization
stops itself.
For testing the primality see the probabilistic Prime test

117

Xnumbers Tutorial
Large numbers having factors more than 20 digits can be only manipulate with Xnumbers but
no more factorized. This task requires sophisticated algorithms joined with extreme fast
routines, usually written in C++ or Assembler.
When the numbers are very large, no efficient integer factorization algorithm is published; a
recent effort which factored a 200 digit number (RSA-200) took eighteen months and used over
half a century of computer time. The supposed difficulty of this problem is at the heart of certain
algorithms in cryptography such as RSA.
Not all numbers of a given length are equally hard to factor. The hardest instances of these
problems are those where the factors are two randomly-chosen prime numbers of about the
same size, but not too close
Factoring software is available either in commercial math packages or in standalone freeware
programs.
One of the most interesting program released in the public domain, supporting Quadratic Sieve
(QS) and Number Field Sieve (GNFS), that we have used for many years is Msieve.

Batch Factorization with Msieve

Msieve is a very power freeware program, created by Jason Papadopoulos, for factoring large
integers.
You can download the latest version of Msieve.exe from www.boo.net/~jasonp
No particular installation is required. Simply copy it in any folder that you like, for example,
/msieve.

What Msieve Does

Factoring is the study (half math, half engineering, half art form, and half... genial tricks) of
taking big numbers and expressing them as the product of smaller numbers. As the number to
be factored becomes larger, the difficulty involved in completing its factorization explodes and
the elaboration time increase sharply. The multiprecision library contained in Xnumber, written
in VBA, is no more sufficient for performing the factorization of number larger of 18-20 digits.
Msieve can with high probability find the complete factorization of any input number up to about
125 digits in size. The actual number of digits supported is much higher (up to 164 digits), but
problems larger than 125 digits are likely to fail.
Trial division is used on all inputs; if the result is less than 25 digits in size, tiny custom routines
do the factoring. For larger numbers, the code switches to more powerful methods. Prior to
version 1.04, those methods were limited to the quadratic sieve (QS). From that point on,
however, an implementation of Pollard-Brent algorithm and the number field sieve NFS are
also available.
A description of QS and NFS can be found in the Msieve Library itself with also a good amount
of Quadratic Sieve references.
The oldest users will be pleasant to know that Msieve.exe is a consolle program. It can be
started at the prompt command (the old DOS enviroment) using the following general syntax
>> msieve [options] [number]
Msieve supports many useful options. For a complete list, run the command
>> msieve -h
For factoring one number give the following command
>> msieve -q

8004000546054003543176004301

8004000546054003543176004301
p10: 1000400017
prp10: 2000000011
prp10: 4000400023
In this case the program only outputs all the factors found. The code "p10" indicates a prime
factor of 10 digits; "prp10" indicates a probable prime factor of 10 digits.

118

Xnumbers Tutorial
With the option -v (verbose), the program will also output the factoring trace with many useful
information (for factoring experts, of course).
>> msieve -v

8004000546054003543176004301

One useful option is also the time limit setting -d

>> msieve -v -d 10 number
This elaboration will stop after 10 minute, ended or not, and the intermediate results will be
saved in a working file, ready for a successive restart.
Onother way to stop a running elaboration is sending the key sequence CTRL+C to the active
command window.
The younger users will be happy to know that this elaboration can be performed directly from
Excel by the macro Numbers/Factors...
Note. The first time that you run Msieve from this macro, you have to provide the folder where
Msieve.exe is located. Follow the simple configuration instructions.
After that, the macro can start Msieve in batch input mode, returning output in Excel at the end
of the elaboration.
Using is very simple. Select the cell containing the number to factorize

Start the macro Numbers/Factors..., choose the Msieve and click on "Factorize".

If we like we can modify the job timeout (default 15 min).

The "Factorization Job Manager" lists and visualizes all the jobs submitted to Msieve.

Lists Refresh
Imports the results in
Excel
Kills a running job
Deletes a job
Verbose mode on/off
This Help

The jobs list is automatically refreshed every second. If the refreshing activity stops for some
reason, you can manually refresh it.
119

Xnumbers Tutorial
The factors obtained can be imported into Excel by the apposite buttom. The macro outputs
each factor in the format [exponent, factor]

An active factorization job runs in a minimize

window. If you like, you can open it at
normal size and observe what Msieve is
doing
The precedent submited jobs remain memorized until they are delete. They can be recalled
and viewed by the buttom Batch... of the Number Factorization panel
The factorization time is usually very fast for numbers up to 50 digits and remains reasonably
fast for numbers in the range 50 - 75 digits, as shows the following statistical graph
1000

Time (sec)
1.2 GHz, 256

100
10
1

Digits

0.1
0

10

20

30

40

50

60

70

80

Factorization function
Factor(n)
This function performs the decomposition in prime factors of a given integer number
Returns an array of two columns: the first column contains the prime factors and the second
column contains the exponents
This function is useful for factorizing on-line integer numbers N < 1E14 directly in the
worksheet.
It uses the so called "brute force attack" and the Fermat-Lehman method.
In this example, the given number is
decomposed in 5 factors
2277785128000 = 26 53 232 732 101
The #ND symbol indicates the end of factors
list. To make sure to get all factors you have
to extend the selection until you see this
symbol
Using this function is very simple. Select a range of 2 columns and several rows (for example 7
rows). The insert the function Factor using the ctrl+shift+enter sequence.
120

Xnumbers Tutorial

Macro - Prime Numbers Generator

PrimeGenerator
This macro is useful to generate your own table of prime numbers. The table begins from the
cell A1 of the active worksheet.
The macro computes 10.000 (default)
prime numbers for each time.
The macro can be stop and restart as
you like
It always restarts from the last prime
number written in the worksheet.

Prime Test
This macro perform the probabilistic prime test with the Fermat or Miller-Rabin method. These
tests are especially adapt for long number.
Using is very simple. Select the number that you want to test and start the macro from the
menu Macros / Numbers / Prime test
Select the method that you want and press "Test". After few seconds you get the results
Note that this test is exact for
detecting composite numbers,
but it can detect a prime number
with a finite probability (usually
very high).
The numbers satisfying the
Fermat or Miller-Rabin are called
"pseudo-prime"

The probability is correlated to the trials number T with the following approximate formula
Fermat test

Miller-Rabin test

p = 1 2 T

For T = 100 the probability is about

p = 1 4 T

For T = 50 the probability is about

1 7.8 10 31
1 7.8 10 31

You can also select a list of cells containing several numbers to test. This is useful for finding
prime in a set of consecutive integers.
121

Xnumbers Tutorial

Example: find the next prime number after 200000000000532100000000

A prime number must be odd, so let's begin to prepare a sequence of 20 or more odd numbers
starting from 200000000000532100000001
The frequency of prime numbers, in this range, is about 5%, so we hope to find a prime number
in our list. If this does not happen we try with a successive set of numbers, and so on, until a
prime comes out.

In this case we have found a probable prime 200000000000532100000003

You may prove by yourself that it is a true prime

122

Xnumbers Tutorial

Diophantine Equation
DiophEqu(a, b, c)
This function solves the Diophantine linear equation

a x+b y = c

x, y Z

where a, b, c , x, y are all integer numbers

The integer solutions can
be expressed as

xk = x0 + k Dx

y k = y0 + k D y

for k = 0, 1, 2...

This function return an (2 x 2) array of four integer values.

The first row contains a particular solution, while the second row
contains the integer increments for generating all the solutions.
If you only want a particular solution [x0, y0] simply select an array of 2
adjacent cells. If the equation has no solution the function return "?"

x0
D
x

y0
D y

Example. Find all the integer solutions of the equation 2x+3y = 6

As we can see, the function returns one
solution (-6, 6) and the increments (3, -2).
So all the integer solutions of the above
equation can be obtained from the following
formulas for any integer value of k

xk = 6 + 3k

y k = 6 2k
Often is not so easy to find the solution of a diophantine equation. Let's see the following.
Long numbers. This function works also with extended numbers.
Example. Find a solution of the equation ax+by = c having the following coefficients
a
b
18760000596690052 13650847757772
Note that the first coefficients has 17
digits and the second one has 14
digits. Without multiprecision it would
be difficult to solve this problem.
You can enjoy yourself to prove that
the result returned from the function
DiophEqu is correct

123

c
64

Xnumbers Tutorial

Brouncker-Pell Equation
= PellEqu(d, [n])
This function solves the so called Brouncker-Pell quadratic equation on the integer domain

x 2 d y 2 = 1 , x, y , d N
The function returns the n-th integer solution (default n = 1) using the continued fraction method.
As know, this equation has non trivial solutions if d is not a perfect square.
The function works for low-moderate integer d values but, even with this constraints,
sometimes this smallest solution is quite small, and sometimes it is huge. Let's see.
Compute the first 3 integer solutions of the equation x 31 y = 1
2

As we can see the solutions grow sharply also for very small d.
Sometime the first solution is huge also with moderate d. In that case is better to arrange the
solution in vertical cells, as shown in the following examples

Note that the "smallest" solution of the equation x 991 y = 1 has 30 digits !
2

When the solution exceed 250 digits or the time of 200 iterations the functions returns "?".

124

Xnumbers Tutorial

Euler's Totient function

= Totient(n)
Returns the number of integer not exceeding and relatively prime to n
This function is intended only for low-moderate numbers.

Remember that two numbers n , m are relative prime if, an only if, GCD(n, m) = 1
With this function is easy to set a table of the totient (n) values for 1 n 100

(n)

10

10

12

16

18

20

12

10

22

20

12

18

12

28

30

30

16

20

16

24

12

36

18

24

16

40

40

12

42

20

24

22

46

16

42

20

50

32

24

52

18

40

24

36

28

58

16

60

60

30

36

32

48

20

66

32

44

24

70

70

24

72

36

40

36

60

24

78

32

80

54

40

82

24

64

42

56

40

88

24

90

72

44

60

46

72

32

96

42

60

40

125

10

Xnumbers Tutorial

Integer relation
A set of real numbers [x1, x2, ... xn ] , is said to possess an integer relation if there exist a set of
integers [a1, a2, ...an ] , not all zero, such that
a1 x1 + a2 x2 ... + an xn = 0
The integer relation algorithm can be used to solve subset sum problems, as well as to
determine if a given numerical constant is equal to a root of a univariate polynomial with integer
coefficients; in this case we say the constant to be "rational" suche as "12/51" or "algebraic",
such as 2 . On the contrary, if such polynomial does not exist, we say the constant
"transcendental", such as .
Therefore this algorithm can be used for algebraic-transcendental screening of real numbers.
Because we always have an approximated value of the constant, the found integer relation will
not exactly zero but very close to it only if we use a sufficient number of digits.
For example, we find the following 5th degree relation
[a1, a2, a3, a4, a5 ][1, x, x2, x3, x4]T = a1+ a2 x + a3 x2 + a4 x3 + a5 x4 0
where [a1, a2, a3, a4, a5 ] = [-1, 0, -2, 0, 1] , and x = 1.553773974...
Of course, a numerical discovery of a relation by this algorithm does not in general constitute a
proof of this relation; one of the reasons being that the computer operates in finite precision. In
many cases, however, the relations we first discovered numerically subsequently received
rigorous mathematical proofs. Moreover, many complicated relations probably would never,
have been dreamed of without the assistance of the computer.
Algorithms for finding integer relations include HJLS algorithm, LLL algorithm, PSLQ algorithm.
All these algorithms requires high-precision arithmetic, but the PSLQ, developed by Ferguson
and Forcade (1979) and improved by Ferguson and Bailey (1992), is proved the most efficient
and stable integer relation algorithm. The Xnumbers VB macro has been developed from the
PSLQ descripted by Bailey, D. and Plouffe at http://www.cecm.sfu.ca/organics/papers/bailey
Norm and Bounding.
The norm of the relation is defined the Euclidean norm of the vector |x| = [x1, x2, ... xn ]
Among the relations satisfinyg the constraint | xi ai | < , we are clearly interested in those
relations having small norm, and, if possible, with the smallest nth degree.
The constraint of small norm has been introduced for filtering between relations.
In fact, if we can easily observe that, if the vector x becames exagerately large, then an integer
relation will be surely find for any real number.
| 314159265358979323 100000000000000000 | < 10-16 ,
Of course, the norm of this relation is very high (about 3.2 1017 ) and thus we have to reject it
PSLQ computes the norm of any examined relation and returns the bound of the coefficients.
Even if a relation is not found, the resulting bound means that cannot possibly be the root of a
polynomial of degree n, with coefficients of size less than the established bound. Even negative
results of this sort are often of interest.
Examples
given a real value x we build the vector [1, x, x2, x3, ... xn] with n in increasing order n = 3, 4,
5...., we are searching for possible integer relations having small error < 1E-14 and small
norm |x|
4

Epprox. value: 3.189207115002720, Exact value:

Relation: [14, -32, 24, -8, 1],

Polynomial:

14 32 x + 24 x 2 8 x 3 + x 4

Epprox. value: 2.67413461226797, Exact value:

Relation: [4, 24, -12, 4, 6, 0, -1] ,

2 +2

Polynomial:

2+ 2

4 + 24 x 12 x 2 4 x 3 + 6 x 4 x 6

126

Xnumbers Tutorial
Epprox. value: 2.991971857463750, Exact value:

2+ 3

Relation: [-23, -36, 27, -4, -9, 0, 1], Polynomial: 23 36 x + 27 x 4 x 9 x + x

2

Epprox. value: 0.809016994374947, Exact value:

Relation: [1, 2, -4] , Polynomial:

( 5 + 1) / 2

1 + 2x 4x2

Epprox. value: 0.923879532511287, Exact value:

( 2 + 2) / 4

Relation: [-1, 0, 8, 0, -8], Polynomial: 1 + 8 x 8 x

Epprox. value: 0.900968867902419, Exact value:

Relation: [1, -4, -4, 8], Polynomial:

cos( / 7)

1 4 x 4 x 2 + 8x3

Epprox. value: 1.839286755214160,

Relation: [1, 1, 1, -1], Polynomial:

Exact value:

1 + x + x 2 x3

19
33 3 19
33 1

+
+
+
27
9
27
9
3

Epprox. value: 0.012345678901235, Exact value: 1/81

Relation: [1, -162, 6561], Polynomial:

1 162 x + 6561x 2 = (1 81x) 2

A very simple trick for testing a relation consists in increasing the degree to n+1; if the algorithm
still returns the same relation then, with high probabability, the relation is correct. But pay
attention, because, sometime a numeric value can have two different relations with different
degree. This means that the given value is the root of two different polynomials. Let's see the
following example
Epprox. value: 7.7099759466767, Exact value:
Relation: [221, -108, 18, -1], Polynomial:

5+6

221 108 x + 18 x 2 x 3

But the same value also returns the following relation of degree 5
Relation: [221, 113, -90, 17, -1], Polynomial:

221 + 113x 90 x 2 + 17 x 3 x 4

We can verify that both relations are correct, in this case.

Higher degree relations
All the previous examples can be found using PSLQ in standard 64 bit precision. However,
when we search for higher the Polynomial degree ( n > 6), the multiprecision is required
Epprox. value: 3.00161461434129452355339436348, Exact value:
2

4+ 2

Relation: [ -8, 48, -20, 56, -6, 8, 5, 0, -1], Polynomial: -1+104x -18x +8x6-x8
Time: 0.4 sec
We have seen that for 8th degree polynomial we have needed about 30 digits, in this case.
Generally speaking, the required precision increases with the degree and the norm of the
coefficients of the polynomial, as reported in the following table

127

Xnumbers Tutorial
Degree

Digits

10
15
16
20
25
28
30

40
80
85
120
170
220
250

Time (sec)
0.5
4.5
5.4
23.0
81.7
126.1
217.7

Absolute limits. The table can be accessed in two ways: given a polynomial degree we get
the minimum precision of the number (number of significant digits); on the contrary, given a
precision of a number (digits) we get the maximum obtainable polynomial degree.
For example, if we are searching for a polynomial of 15 degree, we must have a number with
80 digits, at least. If we have have calculate a number with 100 digits we can search for
polynomials having degree n < 18.
In the above table and graph we have also reported the time1 in second of the PSLQ (two-level)
algorithm. As we can see, the time increases sharply with the degree. The time stays less the
few seconds for n <16, but requires about 3-4 minutes for n = 30 and 250 digits (the maximum
possible for Xnumbers)
Some other high results are shownin the following list. In these trials, the problem is to recover
the polynomial of degree n satisfied by , where = a1/r b1/s. In
other words, the input vector x = (1, , 2, 2.... n-1), where n = (r s + 1). Here the working
precision was set to the level required by PSLQ and the input number was computed with the
needed precision, as determined by the above table.
Degree: 12. Exact value: = 31/ 6 + 21/ 2 2.61515051... (50 digits)
Relation: [ -25, 0, 552, 0, -60, 0, 166, 0, -60, 0, 12, 0, -1] ,
Time: 2.8 sec
Degree: 12. Exact value: = 31/ 4 + 21/ 3 2.57599506... (50 digits)
Relation: [ -11, -216, -360, -32, 27, -288, 24, 0, -9, -8, 0, 0, 1],
Time: 3 sec
Degree: 15. Exact value: = 31/ 5 + 21/ 3 2.50565198951... (80 digits)
Relation:, [ -59, -360, -540, 80, -1620, 27, -80, -540, 0, 40, -9, 0, -10, 0, 0, 1]
Time: 5 sec
Degree: 16. Exact value: = 31/ 4 21/ 4 0.126866897949771... (85 digits)
Relation: [ 1, 0, 0, 0, -3860, 0, 0, 0, -666, 0, 0, 0, -20, 0, 0, 0, 1],
Time: 5.6 sec
Degree: 20. Exact value: = 31/ 5 21/ 4 0.0565238246127... (120 digits)
Relation: [ 49, -1080, 3960, -3360, 80, -108, -6120, -7440, -80, 0, 54, -1560, 40, 0, 0, -12, -10,
0, 0, 0, 1],
Time: 16.8 sec
Degree: 20. Exact value: = 31/ 5 + 21/ 4 2.434938054618... (120 digits)
Relation: [ 49, -1080, 3960, -3360, 80, -108, -6120, -7440, -80, 0, 54, -1560, 40, 0, 0, -12, -10,
0, 0, 0, 1],
Time: 23 sec
1

Computation on a PC with AMD Athlon, 2GHz, 512 MB Ram

128

Xnumbers Tutorial

It is interesting to note that the two numbers 31/ 5 21/ 4 gives the same relations. In fact they
are both roots of the same 20th degree polynomial.
Degree: 25. Exact value: = 31/ 5 + 21/ 5 2.39442929461255233... (170 digits)
Relation: [-3125, 0, 0, 0, 0, 21875, 0, 0, 0, 0, -57500, 0, 0, 0, 0, -3500, 0, 0, 0, 0, -25, 0, 0, 0, 0,
1],
Time: 81.7 sec
Degree: 25. Exact value: = 31/ 5 21/ 5

0.0970325846184823... (170 digits)

Relation: [ 1, 0, 0, 0, 0, -116255, 0, 0, 0, 0, -11240, 0, 0, 0, 0, -3760, 0, 0, 0, 0, 5, 0, 0, 0, 0, -1],

Time: 68.7 sec
Degree: 28. Exact value: = 31/ 7 21/ 4

-0.01927630224403418... (220 digits)

Relation: [ -47, -2688, -13104, -7560, 448, -79296, 136584, -108, -672, -184128, -25956, 0,
560, -74592, 54, 0, -280, -4872, 0, 0, 84, -12, 0, 0, -14, 0, 0, 0, 1],
Time: 126 sec
Degree: 30. Exact value: = 31/ 6 21/ 5

0.0522386001789677198... (250 digits)

Relation: [ 179, -4860, 30780, -65520, 30240, -192, -405, -98820, -538380, -149520, -240, 0,
270, -84240, 90360, -160, 0, 0, -90, -7380, -60, 0, 0, 0, 15, -12, 0, 0, 0, 0, -1],
Time: 218 sec

129

Xnumbers Tutorial

Macro Integer Relation Finder

The macro Integer Relation Finder, from the menu "Numbers", searches for two type of integer
relations:
Polynomial relation
Input: a real number x, and the n-th degree

a0 + a1 x + a2 x 2 + a3 x 3 + ...an x n

General relation
Input: a vector x of n-th dimension

a1 x1 + a2 x2 + a3 x3 + ...an xn

Using this macro is very simple: select the cell containing the number x or the range of the
column-vector x and start the macro. The macro assumes a simple layout in order to speed-up
the input operation. The input layout is different for the two type of relation
Polynomial relation input layout
The cell B2 contains the multiprecision number x
of 30 digits and just below, the cell B3 contains
the degree n. Select the cells B2 and start the
macro
Multiple searching for one x and several degree

The cell B2 contains the multiprecision number x

of 30 digits and just the row below, the range
B3:D3 contains the degree n = 5, 6, 7. Select the
cells B2 and start the macro. For the same
number x, the macro will search for integer
polynomial with degree 5, 6 and 7 respectively

Multiple searching for several x and several degree

The range B2:B4 contains the multiprecision

numbers x1, x2 and x3 and just the row
below, the range B5:E5, contains the degree
n = 5, 6, 7 an 8. Select the range B2:B4 and
start the macro. For the each number x, the
macro will search for 4 integer polynomial with
degree 5, 6, 7 and 8, respectively

General relation input layout

The range B2:B6 contains the vector x .
Select the range B2:B6 and start the macro.
The dimension of he relation will be 5, just the same of the
vector dimension

Multiple searching for several vectors is possible. Simple select as many adjacent vectors as
we need. Note that each vector can have different dimension. In this case, select the
rectangular range containing all the vectors.

130

Xnumbers Tutorial
Output results
The macro writes the result in the worksheet are beside the input data using a simple standard
layout almost similar for polynomial and general relation.
Let' s see with an example
Example 1. We want to search for a smallest integer polynomial having the approximate root:
x 2.99197185746375045829465694878... (30 digits)
We restrict our searching to the polynomials from 3 to 8 degree
Insert the input data x and the degree
3, 4, 5, 6 and 7 as in the sheet at the
left. Select the cell B1 and start the
macro

The input fields are correctly filled.

Press "start". In few seconds the
macro finds the polynomials adding
same useful information. Fortunately
it have found 2 "good" polynomials,
in this case

The coefficents of the polynomials are listed in the last section of the report, but just above, the
macro write the information returned by the PSLQ routine
Digits: the precision of the calculus
Degree: degree of the polynomial; degree +1 is the dimension of the relation
RC: return code: 0 =OK, 1 =precision exhausted, 2 = iteration overflow, 3= overflow error
This code states the "goodness" of the relation found
Residue: relative residue of the relation, computed as / a , where

a x

i i

, a = max(| ai xi

|)

Bound: this number M states that there can exist no relation vector whose Euclidean norm is
less than M. This is a very important result provided by PSLQ algorithm. Together with RC
helps us to analyze the goodness of the relation found.
Iter: number of iterations needed for reaching the result
Time: elaboration time in seconds
Example 2. We have 3 real numbers x1, x2, x3 computed with 15 precision digits.
x1 = 35.812551166675

x2 = 8.554440155849

x3 = 4.186428394404

We want to investigate if they could be express by the following simbolic formula

x=

1
a a
a2 + a3e + 4 + 5
e
a1

a1 x + a2 + a3e +

a4

a5
=0
e

where [a1, a2, a3, a4, a5] all integer.

For finding this integer relation of dimension n = 5 (that it is compatible with the precision of the
numbers that we have) we compute the following vectors
value with x1, x2, x3 respectively

131

[x, ,

, e, e 1 changing the x

Xnumbers Tutorial
Arrange the 3 vectors as in the
sheet at the left. Select the range
B2:D6 and start the macro.

The input fields are correctly filled

and the option button "General" is
selected. Press "start".

In very few seconds the macro succesfully finds integer relations for the first two numbers,
while, on the contrary, the third numbers has not a small simple relation. Note how large are
the coefficents of the third column. Note also that its bound M = 3.4E+9 its very high; it assures
that no integer relation exists with norm ||a|| < 3.4E+9

132

Xnumbers Tutorial

Linear Algebra Functions

Matrix Addition
xMatAdd(mat1, mat2, [DgtMax])
Performs the addition of two matrices. mat1 and mat2 are (n x m) arrays

c11 .. c1 n a11 .. a1 n b11 .. b1 n

c
=
+

n 1 .. cn m an 1 .. an m bn 1 .. bn m

Matrix Subtraction
xMatSub(mat1, mat2, [DgtMax])
Performs the subtraction of two matrices. mat1 and mat2 are (n x m) arrays

c11 .. c1 n a11 .. a1 n b11 .. b1 n

c
=

n 1 .. cn m an 1 .. an m bn 1 .. bn m

Matrix Multiplication
xMatMult(mat1, mat2, [DgtMax])
Performs the multiplication of two matrices. mat1 (n x p) and mat2 (p x m) are arrays

a11 .. a1 m
c11 .. c1 n a11 a12 .. a1 p

=
a2 1 .. a2 m

n 1 .. cn m an 1 an 2 .. an p a
p 1 .. a p m

Matrix Inverse
xMatInv(A, [DgtMax])
Returns the inverse of a (n x n) square matrix
It returns "?" for singular matrix.
This function uses the Gauss-Jordan diagonalization algorithm with partial pivoting method.

Matrix Determinant
xMatDet(A, [DgtMax])
Returns the determinant of a square matrix.
It returns "?" for singular matrix.

133

Xnumbers Tutorial

Matrix Modulus
xMatAbs(A, [DgtMax])
Returns the absolute value of a matrix or vector.
It is also known as "modulus" or "norm"
Parameters A may be an (n x m) array or a vector

A =

(ai , j ) 2
i =1 j =1

Scalar Product
xProdScal( v1, v2, [DgtMax])
Returns the scalar product of two vectors
n

c = V1 V2 = V1 i V2 i
i =1

Note: The scalar product is zero if, and only if, the vectors are perpendicular

V1 V2 = 0

V1 V2

Similarity Transform
= xMatBAB(A, B, [DgtMax])
Returns the matrix product:

C = B 1 A B

This operation is also called the "Similarity Transform" of the matrix A by the matrix B. This
operation plays a crucial role in the computation of eigenvalues, because it leaves the
eigenvalues of the matrix A unchanged. For real, symmetrical matrices, B is orthogonal. The
Similarity Transform is also called the "orthogonal transform".
A and B must be square matrices.

Matrix Power
= xMatPow(A, n, [DgtMax])
Returns the integer power of a square matrix.
n7
time
64
48
B = A = A A A... A
n

Example

134

Xnumbers Tutorial

Matrix LU decomposition
= xMatLU(A, [Pivot], [DgtMax])
Returns the LU decomposition of a square matrix A
It uses Crout's algorithm

1 0 0 u11 u12
A = L U = l21 1 0 0 u 22

l21 l21 1 0
0

u13
u23

u33

Where L is a lower triangular matrix, and U is an upper triangular matrix

The parameter Pivot (default=TRUE) activates the partial pivoting.
Note: if partial pivot is activated, the LU decomposition refers to a permutation of A
If the square matrix has dimensions (n x n), this function returns an (n x 3n) array where the
first n columns are the matrix L, the next n columns are the matrix U, and the last n columns
are the matrix P.
Globally, the output of the Mat_LU function will be:
- Columns (1, n)
= Matrix L
- Columns (n+1, 2n) = Matrix U
- Columns (2n+1, 3n) = Matrix P
When pivoting is activated the right decomposition formula is: A = P L U , where P is a
permutation matrix
Note: LU decomposition does not work if the first element of the diagonal of A is zero
Example: find the factorization of the following 3x3 matrix A

Note: if you want to get only the L and U matrices select a range (3 x 6) before entering this
function

Matrix LLT decomposition

= xMatLL(A, [DgtMax])
This function returns the LLT decomposition of a square matrix A
It uses Cholesky's algorithm

l11 0
T
A = L L = l21 l22
l31 l32

0 l11 0
0 l21 l22
l33 l31 l32

0
0
l33

Where L is a lower triangular matrix

The function returns an (n x n) array
Note: Cholesky decomposition works only for positive definite matrix
135

Xnumbers Tutorial

Example.

The diagonal elements of the L matrix are all positive. So the matrix A is positive definite and
the decomposition is correct. This function simply stops when detects a negative diagonal
element, returning the incomplete decomposition.
See this example
<Attention !>
A diagonal element of the L matrix is
negative. So the matrix is not positive
definite and the decomposition cannot be
completed

Vector Product
= xProdVect(v1, v2, [DgtMax])
Returns the vector product of two vectors

v11 v12 v21v32 v22 v31

V1 V2 = v21 v22 = v12 v31 v11v32

v31 v32 v11v22 v21v12

Note that if V1 and V2 are parallels, the vector product is the null vector.

Solve Linear Equation System

xSYSLIN( A, B, [DgtMax])
Solves a linear system in multiprecision.
The input parameter A is an (n x n) array, B may be a vector (n x 1) or an (n x m) array.
It returns a vector (n x 1) or an (n x m) array depending by the argument B
A set of m linear systems in n unknowns looks like this:

[ A ] x1 = b1 , [ A ] x1 = b 2 , ... [ A ] x1 = b m
It can be rewritten as:

a11
[ A ] [ x ] = [ B ] ...

an 1

.. a1 n x11

.... ...

.. a n n x n1

.. x1m b11
.... = ...

.. x n m b n 1

.. b1 m
....

.. b n m

This function uses the Gauss-Jordan diagonalization algorithm with partial pivoting method.
136

Xnumbers Tutorial
Example. Find the solution of the following 7x7 linear system
A
462
924
1716
3003
5005
8008
12376

792
1716
3432
6435
11440
19448
31824

1287
3003
6435
12870
24310
43758
75582

2002
5005
11440
24310
48620
92378
167960

3003
8008
19448
43758
92378
184756
352716

b
4368
12376
24290
12376
31824
62856
31824
75582
149877
75582
167960
333918
167960
352716
702429
352716
705432 1406496
705432 1352078 2697968

The solution is the vector [1, 1, 1, 1, 1, 1, 1]. Solving with standard arithmetic, we have an
average accuracy of about 1E-8, while in multiprecision we have an accuracy better than 1E-28

Square Delta Extrapolation

ExtDelta2(x)
xExtDelta2(x, [DgtMax])
This function returns the Aitken's extrapolation, also known as "Square Delta Extrapolation".
The parameter x is a vector of n value (n 3), in vertical consecutive cells. (n = 3 for the multiprecision function xExtDelta2).
This formula can be applied to any generic sequence of values (vector with n>2 ) for
accelerating the convergence.
2

( x1 , x2 , x3 ... , xn ) (v1 , v2 , v3 ... , vn 2 )

Note that this algorithm produces a vector with n-2 values. If n = 3, the result is a single value.
Taking the difference i = xi +1 xi , the Aitken's extrapolation formula is:

(xi xi 1 )
2i 1
= xi
(xi 2 xi 1 + xi 2 )
i 1 i 2
2

vi = xi

This formula can be applied to the second sequence to obtain a new sequence with n-4 values,
and so on. The process stops when the last sequence has less than 3 values.
Example. we want to find the numeric solution of the equation x = cos(x)
We choose the central point method. Starting from x0 = 0 we build the iterations
xn+1 = cos(xn)

137

Xnumbers Tutorial
As we can see in the following table, the convergence is evident but very slow (after 12
iterations the precision is about 3E-5) .
The functions of this worksheet are:
The cell B2 contains the starting value x0
The cell B3 , C2, D3 contain the formulas
[B3]=C2
[C2]=COS(B2)
[D3]=ASS(B2-C2)
Selecting the range B3:D3 and dragging
down we can easily perform the iterative
process as we like.
We observe that the convergence is
evident but quite slow.
As we can see the last 12th value has an
error of about 3.5E-3. But if we perform
the delta extrapolation of the three last
values we get a new value having an
accuracy better than 1E-5.
Now let's repeat the iterative process using systematically the square delta extrapolation
In this process, we have systematically
repeated the 2 extrapolation every 3
iterations.
We have inserted in the cell B5
=ExtDelta2(B2:B4)
In the cell B8
=ExtDelta2(B3:B7)
In the cell B14
=ExtDelta2(B9:B11)

The acceleration is superb!. After only 12 steps, the precision is better than 1E-15.
The graph below shows better than many words the acceleration effect
0

10

12

1
0.01
0.0001
1E-06
1E-08
1E-10
1E-12
1E-14

Error (extrap)
Error

1E-16

The Aitken's extrapolation formula works very well with the Gauss-Seidel iterative method, and
for accelerating the convergence of many iterative processes.

138

Xnumbers Tutorial

Macro for Multiprecision Matrix Operations

This application collects a set of useful macros performing multiprecision matrix operations
Gauss-Jordan algorithm

Determinant

det(A)

Addition

A+ B

Subtraction

A B

Multiplication

A B

Scalar multiplication

kA

Inverse

A1

Similarity transform

B 1 A B

Linear System

AX = B

Gauss-Jordan algorithm

Linear System overdetermined.

Ax = b

rows > columns

LU decomposition

A = LU

Crouts algorithm

Cholesky decomposition

A = LLT

Choleskys algorithm

Norm

Gauss-Jordan algorithm

Scalar product

AT B

SVD

U V T

Golub-Reinsch algorithm

The use of this macro is quite simple. If the operation requires only one matrix (determinant,
inversion, etc.) select the matrix, start the macro and choose the appropriate operation.
Other operations (addition, multiplication, etc.) require two matrices. In that case you have also
to select the second matrix in the second input-box.
The internal calculus is performed in multiprecision. The result is converted in standard
precision (15 significant digits max) for more readability, but, if you like, you may also leave it in
full multiprecision format.
Example: Solve the following linear system Ax = b.

Select the matrix A and start the macro

Choose the operation Linear System and then move into the right field to select the vector b.

139

Xnumbers Tutorial

Indicate, if necessary, the upper-left

cell of the range where you want to
write the result.

Then, press OK. The result will be output starting from the output cell I2.
The special button near the input field is useful for selecting large
matrices .Select the first cell, or an internal cell of the matrix and then
press this button. The entire matrix will be selected

Smart
Selector

Elaboration time
Multiprecision computation does slow down the computation considerably. It takes much more
time then the standard double precision. The time depends on the matrix dimension and on
the precision digits. The following graphs show the average time for the inversion and for the
multiplication of two dense matrices.
160

140

Time (sec)

30 digits

time (sec)

120
100

invers.

80

multip.

60

20 x 20

40

20

Nx N
0

25

50

75

digits

100

50

100

150

200

standard precision

Multiprecision

As we can see, the inversion of a (100 x 100) dense matrix, with 30 precision digits, takes
about 150 seconds. Clearly, for this kind of tasks, macros are more suitable than functions.

140

Xnumbers Tutorial

Integrals & Series

Discrete Fourier Transform

=DFT(samples, [central])
=FFT(samples, [central])
Returns the complex matrix of the DFT transform of N samples.
This function returns an (N x 2) array. The first column contains the real part; the second
column the complex part .
The optional parameter "central" (default False) outputs a central transform.
If N is an integer power of 2, that is N = 2P, use the fastest FFT.
FFT uses the Cooley and Tukey decimation-in-time algorithm.
Formulas
Given N samples ( f(0), f(1), f(2),....f(N-1) ) of a periodic function f(t) with a normalized
sampling rate (T=1), the DFT is:

F (k ) = Fr (k ) + i Fi (k ) =

N 1

f (n) [cos(2 nk / N ) i sin(2 nk / N )]

n=0

The components (Fr , Fi) are called the harmonic spectrum of f(t)
By the Fourier series, we can approximate a periodic function f(t):

f (t ) a0 +

K 1

ak cos(k t ) + bk sin(k t )

k =1

where the coefficients (ak , bk ) are the components 2Fr and 2Fi of DFT
Example: Find the 16-FFT of the following periodic function (T = 1 sec)

f (t ) = 3 + cos( t ) + 0.5 cos(3 t )

where

2
T

First of all we have to sample the given function. Setting N = 16 , we have a sampling period of
t =

1
T
=

N 16

ti = i t

, i = 0, 1,...N 1

f i = 3 + cos( ti ) + 0.5 cos(3 ti )

Appling the FFT function at the samples set ( f0, f1, f2,....f15 ), we get the complex discrete
Fouriers transform

141

Xnumbers Tutorial

Note that the FFT returns a (16 x2)

matrix. The first column contains the real
part of FFT while the second column the
imaginary one.
The magnitude and phase can be easily
obtained with the following formulas

(FFTre )2 + (FFTim )2

Ai =

FFTim

FFTre

i = arctan

Observe that the first row of the FFT

contains the mean of f(t).
Note also that the rows from 10 to 16 are
the mirror-copy of the previous rows.

Rearranging the mirror part we can obtain the central Fourier's spectrum.
For example, below there are two representations of the same DFT, obtained setting the
"central" parameter "false" and "true", respectively.
Bilateral DFT

Central DFT

Amp

Amp

0.2

0.2

0.15

0.15

0.1

0.1

0.05

0.05

10 13 16 19 22 25 28 31

142

10 13 16 19 22 25 28 31

Xnumbers Tutorial

Discrete Fourier Inverse Transform

=DFT_INV(samples, [central])
=FFT_INV(samples, [central])
Returns the inverse of the DFT transform of N complex samples.
This function returns an (N x 2) array containing the samples of the function f(t).
The optional parameter "central" (default False) outputs a central transform.
If N is an integer power of 2, thus N=2P, use the fastest FFT_INV function
FFT_INV uses the Cooley and Tukey decimation-in-time algorithm.
Formulas

f ( n) =

N 1

[Fr (k ) + i Fi (k )] [cos(2 nk / N ) + i sin(2 nk / N )]

k =0

Where the components (Fr , Fi) are the harmonic spectrum of f(t)
Example: Find the inverse transform of the FFT computed in the previous example

As we can see, the first column of FFT_INV returns the samples of f(t) that have originated the
FFT itself

143

Xnumbers Tutorial

Discrete Fourier Spectrum

=DFSP(samples, [dB], [Angle])
This function returns the harmonic spectrum of a samples set
The parameter "samples" is a vector of N equidistant samples
The optional parameter "dB" (default FALSE) sets the amplitude conversion in decibel
The optional parameter "Angle" (default RAD) sets the angle unit (RAD, GRAD, DEG)
The function returns an (N x 2) array, containing the amplitude and phase.
The spectrum (An, n) is computed for real positive frequencys, being

f (t ) f (0) + An sin (n t + n )
Example: Find the harmonic spectrum of the following 32 samples
Amplitude
1.2
1
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Phase
100
50
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
-50
-100

Inverse Discrete Fourier Spectrum

=DFSP_INV(spectrum, [dB], [Angle])
This function builds the temporal sequence from its real spectrum (amplitude, phase)

( An , n )

f (ti )

The parameter "spectrum" is an (M x 2) array. Each row contains a harmonic. The first column
contains the amplitude and the second column the phase
The optional parameter "dB" (default FALSE) sets the output in decibel
The optional parameter "Angle" (default RAD) sets the angle unit (RAD, GRAD, DEG)
The function returns the vector (N x 1) where N = 2M

144

Xnumbers Tutorial

2D Discrete Fourier Transform

=FFT2D (samples)
This function performs the 2D-FFT of a bidimensional data set (xi, yi).
The parameter "samples" is an (N x M) array where N and M are integer powers of 2 (4, 8, 16,
32, 64). The function returns an (2N x M) array. The first N rows contain the real part, the last
N rows contain the imaginary part.
Note: This function requires a large amount of space and effort. Usually it can works with
matrices up to (64 x 64).
Example: Analyze the harmonic component of the following (8 x 8) table

The 2D-FFT can be

computed in a very straight
way. Simply select a 16 x 8
array and insert the FFT2D
where the input parameter is
the given matrix (range
B10:I17).
We can easily extract the
harmonic components:

H (0,0) = 1
H (1,0) = 0.05
H (0,1) = 0.1

H (1,1) = 0.354 + 0.354 j

H ( 2,2) = 0.25 j

If we compute the inverse transform DFT2D_INV(L10:S25) we will obtain again the given
starting matrix.

2D Inverse Discrete Fourier Transform

=FFT2D_INV (samples)
This function FFT2D_INV performs the inverse of the FFT2D. It accepts in input an (2N x M)
array having the real part in the first N rows and the imaginary part in the last N rows. It returns
an (N x N) array

145

Xnumbers Tutorial

Macro DFT (Discrete Fourier Transform)

This macro performs:

the DFT of a data set of N samples

the DFT-Inverse of a data set of N complex samples

the 2D-DFT of a matrix of N x M samples

the 2D-DFT-Inverse of a two matrices of N x M samples

DFT
It works for any number N.
If N is a powers of 2 (8, 16, 32, 64, etc.) the macro uses the faster FFT algorithm and the
elaboration is more efficient.
Its use is quite simple. Select the vector of samples and then start the macro

The column x is not strictly necessary. If present, the macro uses it to calculate the sampling
parameters (see the top-right box).
For long input vector, you can select only the first cell C9 and start the macro. The entire input
data set (32 x 1) will be automatically selected.
The macro writes the results in the following way
f
=
Hre =
Him =
Amp =
Phase =

frequency sample
Real part of DFT transform
Imaginary part of DFT transform
Amplitude (if polar is checked)
Phase (if polar is checked)

The amplituted can be converted in dB. That is: Amp dB = 20 Log(Amp)),

or also in dBm, the power ratio relative to 1 mW (a sine wave of 1 Vpeak , with 50 ohms load, is
equivalent 10 dBm).
The option check box "central" specifies if the DFT output is bi-lateral or central.
See the following examples

146

Xnumbers Tutorial
Example of a Fourier
transform of pulse.
Period = 1 sec
Duty cicle = 5.5%
Amplitude = 1
Sampling time = 1/256

f(t)
1.2
1
0.8
0.6
0.4
0.2
0
-0.5

Amplitude of the central

DFT
(-100 < f < 100)

-0.3

-0.1

0.1

0.3

0.5

Amp(f)
0.06
0.05
0.04
0.03
0.02
0.01
0
-100

Amplitude of the bi-lateral

DFT
(0 < f < 255)

-50

50

100

Amp
0.06
0.05
0.04
0.03
0.02
0.01
0
0

Amplitude in dB of the
central DFT
(-100 < f < 100)

50

100

150

200

250

Amp (dB)
-25
-30
-35
-40
-45
-50
-55
-60
-100

-50

147

50

100

Xnumbers Tutorial

Operation DFT-inverse
In this case you have to select two columns: the real and imaginary part of the DFT (Hre, Him).
Then start the macro as usually.
If the DFT is in polar form (Amplitude, Phase), you have to check the polar option box and
choose consequently the appropriate units: dB and angle
Operation 2D-DFT
Dimension N and M must be integer power of 2
In this case you have to select a matrix of N x M values (do not select the axes-scales) and
start the macro as usually.
If you want the DFT in polar form (Amplitude, Phase), you have to check the polar option and
choose consequently the appropriate units: dB and angle

The macro generates two matrices containing

the real and imaginary parts of the 2D-DFT

Operation 2D-DFT inverse

In this case you have to select a matrix of 2N x M values (do not select the axes values)
containing both real and imaginary part and start the macro as usually.
If the DFT is in polar form (Amplitude, Phase), you have to check the polar option and choose
consequently the appropriate units: dB and angle

148

Xnumbers Tutorial

Macro Sampler
This is a simple but very useful macro for function sampling
It can generate samples of multivariate functions such as:
f (x) , f (x1, x2) .... f (x1, ... xm)
The samples can be arranged in a list or in a table (only for 2 variables).
Examples of lists and tables generated by this macro are shown in the following sheet

The tables at the top are the skeletons to generate the samples-list or the samples-table just
below. The skeleton contains the following parameters for the sampler.
Start
Samples
Period
Step
Cyclic

starting point of the variable X0

number of samples to generate: N
length of the sampling: P
length between two consecutive point H = X1 X0
True or False (default), specifies if the function is periodic with period P.

The difference between a cyclic or no-cyclic function is in the step formula

S=P/N
for cyclic function
S = P / (N-1) for no-cyclic function
For example, the sampling with N = 5, X0 = 0 and P = 2, needs a step H = 0.5

The first point and the last point, in this case, are always taken.
But, for a periodic function, the same sampling needs a step of H = 0.4

Practically, the last point X = 2, in this case, is discharged, because is f(0) = f(2). Usually
periodic functions require to set Cyclic = True for the FT analysis
The skeleton can be drawn by hand or
automatically. In this case you have only to give
the number of variables that you want.
The check-box Function seed tells the macro
to created also the cell in which you can insert
the function to sample

149

Xnumbers Tutorial
A simple skeleton for one variable is:
In the cell C2 you must insert the function
f(x) to sample. The reference for the
independent variable x is the cell B2. For
example, if the function is y = x +2x2
You have to insert the formula
= B2+2*B2^2 in the cell C2
Parameters N (Samples), P (Period), H (Step) are not all independent. Only two parameters
can be freely chosen.
The macro chooses the first two parameters found from top to bottom
The remain parameter is obtained by the step-formula
Synthetically you can have one of the following three cases
Given parameters
Samples, Period (N, P)
Samples, Step
(N, H)
Period, Step
(P, H)

Calculated parameter
Step
(H)
Period (P)
Samples (N)

Look at the following three examples below for better explanation. The given parameters are in
blue while the calculated parameter is in red.

After you have set and filled the skeleton, select it and start the sampler macro again
(remember that range must always have 6 rows, including the header)
The macro show the following window

The check-box Add formula tells to the macro to leave the formula in the sample set,
otherwise it contains only the values. Formula can be added only for a monovariable list or for a
table.

150

Xnumbers Tutorial

Integral function
=IntegrData(xi, yi, [IntType])
Computes numerically the integral functions F(xi) of a given dataset (xi, yi)
xi

F ( xi ) = yi dx
0

The sampling interval must be constant

The IntType parameter (default = 2) sets the integration formulas 1, 2, 3
Case 1 - 2 points integration formula of 1st degree

I12 = h( f1 + f 2 ) / 2

symmetric

E h 3 f ( 2 ) / 12

Case 2 - 4 points integration formulas of 3rd degree

symmetric

I 23 = h( f1 + 13 f 2 + 13 f 3 f 4 ) / 24

Left side

I12 = h(9 f1 + 19 f 2 5 f 3 + f 4 ) / 24

Right side

I 34 = h( f1 5 f 2 + 19 f 3 + 9 f 4 ) / 24

E 11 / 720 h 5 f ( 4 )

Case 3 - 6 points integration formulas of 5th degree

I 34 = h(11 f1 93 f 2 + 802( f 3 + f 4 ) 93 f 5 + 11 f 6 ) / 1440

symmetric

E 191 / 60480 h 7 f ( 6)

I12 = h( 475 f1 + 1427 f 2 798 f 3 + 482 f 4 173 f 5 + 27 f 6 ) / 1440

Left side

I 23 = h(27 f1 + 637 f 2 + 1022 f 3 258 f 4 + 77 f 5 11 f 6 ) / 1440

I 45 = h( 11 f1 + 77 f 2 258 f 3 + 1022 f 4 + 637 f 5 27 f 6 ) / 1440
I 56 = h(27 f1 173 f 2 + 482 f 3 798 f 4 + 1427 f 5 + 475 f 6 ) / 1440

Right side

Depending on the case, the algorithm uses these formulas at the beginning, at the center and
the end of the integration interval.
2 points formula
f1

f2

x
h

4 points formulas
f1

f2

f3

f1

f4

f2

f3

x
h

f1

f4

f2

x
h

151

f3

f4

Xnumbers Tutorial
6 points formulas
f1

f2

f3

f4

f5

f1

f6

f1

f2

f3

f4

f5

f3

f4

f5

f6

f6

x
h

h
f1

f2

f2

f3

f4

f5

f6

f1

f2

f3

f4

f5

f6

x
h

Example. Compute the integral function of the following tabulated function y(xi)

Using the integration formulas 1, 2 and

3, we have, of course, different precision.
The following graph shows the error
behaviour of the three schemas.
The average errors are about 0.05 for
the 1st degree formula, 5E-5 for the 3rd
degree and 4E-6 for the 5th degree.
Clearly, for smooth, regular functions,
the highest degree formulas reach the
highest precision but we have to
consider that they are also more
expensive

1.E+00

err1
err2
err3

1.E-01
1.E-02
1.E-03
1.E-04
1.E-05
1.E-06
0

Usually the 3rd degree formulas are the best compromise between cost and precision
Avoid discontinues. But, generally, do highest integration degree formulas always give the
best precision? Not always. When the function to integrate shows jumps or direction
discontinuities, the better choice is the linear formula. See the following example.
Given the following functions

y ( x ) = (| x 4 | + x ) / 2 | x 2 |
its exact integral function is
x

F ( x) = y ( x)dx = ( x 4) | x 4 | 2( x 2) | x 2 | + x 2 + 8 / 4
0

152

Xnumbers Tutorial
Sampling the function y(x) and F(x) with h = 0.25 for 0 x 6, we have the following data
sets (xi, y(xi)) and (xi, F(xi))
4.5
4
3.5
3
2.5

y dx

1.5
1
0.5
0
0

As we can see, the function y(x), always linear, has no derivative at the points x = 2, and x = 4.
If we integrate the data set (xi, yi) with the 1st degree formula (IntType = 1) the result coincides
with the exact solution (error = 0). But if we try with the other higher formulas the average error
is, surprisingly, about 0.005

Integral function of a symbolic formula

When we have the function f(x) written in a symbolic formula we can obtain the plot of its
integral function in two ways.

Sampling the given function f(xi) for xi = x0 + ih with a suitable step h

Solving the associated ODE equation
Integrating f(x) from x0 to xi for xi = x0 + ih with a suitable step h

The first method is simple and it is already shown in the previous chapter; the second way
need some more explanation.
The function y(x) that we want to plot is defined as
x

y ( x) =

f ( x)dx

x0

Taking the derivatives of both sides and remebering Leibniz's rule, we have

y ' ( x) = f ( x)

y ( x0 ) = 0

As we can see, the computing of any integral function is equivalent of solving an ODE.
Therefore we can use any method that we have used for solving ODE: Runge-Kutta, PredictorCorrector, Taylor, etc. (see chap. ODE for further details)
The third method, less efficient then the others, may be successfully used when the function
f(x) or its derivatives f'(x) has some singularities in the integration range.
The following example explain better the concept. Assume to have to plot the integral function
of the following function
x

y ( x) = ln( x 2 )dx
0

We note that,. lim ln( x ) =

2

x 0

Therefore, x = 0 is a singular point. From theory we know that the integral exists, thus we can
project to tabulate the given integral for several points xi = ih with a suitable step h, for
example h = 0.1
For this scope we may use the function integr_de or Integr that are more suitable for such
singularities

153

Xnumbers Tutorial
Insert the function definition in cell B1
Ln(x^2)
Build a sequence x = 0, 0.1, 0.2...4 in the
range A4:A44
Insert the function integr_de in cell B5 and the
function = LN(A5^2) in cell C5 and drag down
the range B5:C5.
Note that we cannot complete the first cell C4
because of #NUM! error.

Now, plotting the the range A5:C44, we have the following graph
Comparing with the exact
solution

4
f(x)

y ( x) = x ln( x 2 ) 2 x

f(x)

2
1
0
-1
-2
-3
-4
0

0.5

1.5

2.5

3.5

we observe a precision better

then 1E-15.
From the graph we see that the
integral function y(x) has a zero
between 2.5 and 3
From the table we may take a
more precise bracketing of the
zero
2.7 < xz < 2.8

Zeros of integral function

We may obtain a more close solution of the integral equation using the Newton-Raphson
iterative algorithm
x

y ( x) =

f (t )dt = 0

xn +1 = xn

x0

y ( xn )
f ( xn )

In the cell B1 insert the

definition of the function f(x)
Ln(x^2)
Starting from x0 = 2.7 we
calculate the integral from 0 to
2.7 in the cell C5, The
derivative is the function f(x)
itself, calculated in the cell B5.
The new value x1 is calculated
in the cell A6
Iterating the process, the solution happears in the last cells of the column A
As we can see, after few iteration the zero converges to the exact solution x = e with a
precision better then 1E-15

154

Xnumbers Tutorial

Function Integration (Romberg method)

=Integr_ro(Funct, a, b, [Param], [rank], [ErrMax])
This function computes the numeric integral of a function f(x) by the Romberg method.
b

I = f ( x)
a

The parameter "Funct" is a math expression string in the variable x, such as:
"x*cos(x)", "1+x+x^2", "exp(-x^2)", ecc.. .
Remember the quote " " for passing a string to an Excel function.
"Funct" may be also a cell containing a string formula
"Param" contains label and values for parameters substitution (if there are)
"Rank", from 1 to 16 (default), sets the maximum integration rank.
"ErrMax" (default 1E-15) , sets the maximum relative error.
For further details about writing a math string see Math formula string
The algorithm starts with rank = 1 and incrementing the rank until it detects a stop condition.
|R(p, p) - R(p, p-1)| < 10^-15

absolute error detect

or
(|R(p, p) - R(p, p-1)|) / |R(p, p)| < 10^-15

if |R(p, p)| >> 1

relative error detect

or
rank = 16
Example. Compute the integral of "x*cos(x)" for 0 x 0.4
Integr_ro("x*cos(x)", 0, 0.4) = 0.0768283309263453
This function can also display the number of sub-intervals and the estimate error.
To see these values simply select three adjacent cells and give the CTRL+SHIFT+ENTER key
sequence.

This result is reached with rank = 4 , s = 16 sub-intervals, and an estimate error of about
E = 3.75E-16
The function accepts also external parameters. Remember only to include the label in the
parameter selection.

155

Xnumbers Tutorial

Function Integration (Double Exponential method)

= Integr_de(funct, a, b, [Param])
This function1 computes the numeric integral of a function f(x) by the Double Exponential
method. This is specially suitable for improper integrals and infinite, not oscillating integrals.
b

I = f ( x)dx

f ( x)dx

I=

I=

f ( x)dx

The parameter "funct" is a math expression string in the variable x, such as:
"x*cos(x)", "1+x+x^2", "exp(-x^2)", ecc.. .
Remember the quote " " for passing a string to an Excel function. "funct" may be also a cell
containing a string formula.
The limits "a" and "b" can also be infinite. In this case insert the string "inf"
"Param" contains labels and values for parameters substitution (if there are)
For further details about writing a math string see Math formula string
The Double Exponential method is a fairly good numerical integration technique of high
efficiency, suitable for integrating improper integrals, infinite integrals and "stiff" integrals having
discontinue derivative functions.
This ingenious scheme, was introduced first by Takahasi and Mori [1974]
For finite integral, the formula, also called "tanh-sinh transformation", is the following
b

f ( x)dx = f ( x(t )) h(t )dt

where:

x(t ) =

b+a ba
+
tanh (sinh(t ) )
2
2

h(t ) =

ba
cosh(t )
2 cosh 2 (sinh(t ) )

Example
1

0.5

(1 x) 0.3 dx = 0.474421154996...

The above integral is very difficult to compute because the derivative is discontinue at 0 and 1
The Romberg method would require more than 32.000 points to reach an accuracy of 1E-7. On
the contrary, this function requires less then 100 points for reaching the high accuracy of 1E-14

This function can also evaluate infinite and/or semi-infinite integral.

1

This function uses the double exponential quadrature derived from the original FORTRAN subroutine INTDE of the
DE-Quadrature (Numerical Automatic Integrator) Package , by Takuya OOURA, Copyright(C) 1996

156

Xnumbers Tutorial
Example

dx

As known, the integral exist if n > 1 and its value is I = 1/(n-1). The parameter "n" is called
"order of convergence". For n = 1.1 we get I = 10

Note that we need to pass the parameter with its label "n". (Param = "D1:D2")
This function cannot give reliable results if n is too close to 1.
The minimum value is about n = 1.03. For lower values the function returns "?".
The DE integration works very well for finite improper integral
Example
1

( )

( )

2
2
ln x dx = lim+ ln x dx = 2
0

a 0

Note that the function f(x) is not defined for x = 0

Example. Another difficult improper integral
1

ln( x)

x dx = 4 / 9

157

Xnumbers Tutorial

Function Integration (mixed method)

= Integr(Funct, a, b, [Param])
This function computes the numeric integral of a function f(x) over a finite or infinite interval
+

f ( x)dx

f ( x)dx

f ( x)dx

f ( x)dx

This function can compute definite integrals, improper integrals and piece-wise functions
integrals. The parameter "funct" is a math expression string in the variable x, such as:
"x*cos(x)", "1+x+x^2", "exp(-x^2)", ecc.. .
Remember the quote " " for passing a string to an Excel function.
"funct" may be also a cell containing a string formula
The limits "a" and "b" can also be infinite. In this case, insert the string "inf"
"Param" contains labels and values for parameters substitution (if there are)
This function uses two quadrature algorithms
1) The double exponential method1 (see function integr_de )
2) The adaptive Newton-Cotes schema (Bode's formula) (see macro Integral_Inf )
If the first method fails, the function switches on the second method
Oscillating functions, need specific algorithms. See Integration of oscillating functions (Filon
formulas) and Fourier's sine-cosine transform
Example. Compute the integral of xcos(x) for 0 x 0.4

In the given interval the function is continuous, so its definite integral exists. This result is
reached with rank = 4, s = 16 sub-intervals, and an estimate error of about 3.7E-16. This
function returns the integral and can also displays the number of sub-intervals and the estimate
error. To see these values simply select three adjacent cells and give the
CTRL+SHIFT+ENTER keys sequence.
Note that the function Integr is surrounded by { } . This means that it returns an array
The function Integr can accept also parameters in the math expression string.
See the example below.

This function uses the double exponential quadrature derived from the original FORTRAN subroutines INTDE and
INTDEI of the DE-Quadrature (Numerical Automatic Integrator) Package , by Takuya OOURA, Copyright(C) 1996

158

Xnumbers Tutorial
Note that we must include the parameter labels in order to distinguish the parameters "k", "w",
and "q". The integration variable is always "x"
Beware of the poles
Before attempting to evaluate a definite integral, we must always check if the integral exists.
The function does not perform this check and the result may be wrong. In other words, we have
to make a short investigation about the function that we want to integrate. Let's see the
following example.
Assume the following integrals to have to compute
1/ 2

2
2x 1
2

dx ,

2
2x 1
2

dx

We show that the first integral exists while, on the contrary, the second does not exist
For x p =

2 / 2 0.707... the function has a pole; that is:

2
lim 2 =
x x p 2 x 1

2
lim+ 2 = +
x x p 2 x 1

The first integral exists because its interval [0, 0.5] does not contain the pole and the function is
continuous in this interval. We can compute its exact value:

2x

2
2

dx =

| 2x 1|
2

log

2
| 2x +1|

1/ 2

2x

2
2

dx = 2 log 2 1

In this situation the function Integr

returns the correct numeric result
with an excellent accuracy, better
than 1E-14.
For this result the integration
algorithm needs 128 sub-intervals

The interval of the second integral contains the pole, so we have to perform some more
investigation. Let's begin to examine how the integral function approaches the pole xp taking
separately the limit from the right and from the left

lim log
x x p

2 x 1
=
2 x + 1

lim+ log
x x p

2 x 1
=
2 x + 1

As we can see the both limits are infinite, so the second integral does not exists
Note that if we apply directly the fundamental integral theorem we would a wrong result:
1

2
| 2 x 1 |

=
log
dx

0 2 x 2 1
| 2 x + 1 | = 2 ln 2 1
2
0

wrong !

Let's see how the function integr works in this case.

159

Xnumbers Tutorial
The numeric result is, of course,
completely wrong because the
given integral goes to the infinity.
But, even in this situation, this
function gives us an alert: the subintervals have reached the
maximum limit of 65536 (216)
So the result accuracy must be
regarded with a reasonable doubt.

Complex Function Integration (Romberg method)

=cplxintegr(Funct, a, b)
This function returns the numeric integral of a complex function f(z) by the Romberg method.
b

F (b) F (a ) = f ( z )dz
a

The integration function "Funct" must be a string in the variable z and can be defined mixing all
arithmetic operators, common elementary functions and complex numbers like:
"z*cos(z)", "1+(1+i)*z+z^2", "exp(-z^2)", ecc.. .
Remember the quote " " for passing a string to an Excel function.
Parameters "a" and b can be real or complex. Complex values are inserted as arrays of two
cells.
Example: Evaluate the following integral
1+i

1+ i
dz
2
z
1i

Because the integration function is analytic, then the given integral is independent form the
integration path. Therefore it can be calculated by the function cplxintegr
The exact result is the complex
number (1+i)
Note that, thanks to the excellent
accuracy, the result is shown exactly
even if it is intrinsically approximated

160

Xnumbers Tutorial

Data Integration (Newton-Cotes)

=IntegrDataC(x,y, [Degree])
This function returns the integral of a discrete set of points (xi, yi) using the Newton-Cotes
formulas. The points may be equidistant or random. The parameter "Degree", from 1(default) to
10, set the degree of the Newton-Cotes formula written as:
x0 + n h

h n
f j bj
k j =0

f ( x)dx =

x0

where fi = yi , "h" is the integration step, "n" is the degree.

The coefficients (bj , K) can be extracted from the following table:
Degree

10

45

288

140

17280

14175

89600

299376

b0

14

95

41

5257

3956

25713

80335

b1

64

375

216

25039

23552

141669

531500

24

250

27

9261

-3712

9720

-242625

64

250

272

20923

41984

174096

1362000

14

375

27

20923

-18160

52002

-1302750

95

216

9261

41984

52002

2136840

41

25039

-3712

174096

-1302750

5257

23552

9720

1362000

3956

141669

-242625

25713

531500

b2
b3
b4
b5
b6
b7
b8
b9
b10

80335

As we can see, for degree=1, the Newton-Cotes formula coincides with the trapezoidal rule
and, for degree = 2, with the popular Cavalieri-Simpson formula.
Trapezoid rule

Cavalieri-Simpson rule

x2 x0
2
x2
h
f ( x) 3 ( f 0 + 4 f1 + f 2 )
x0

h = x1 x0

h=

x1

h
f ( x) 2 ( f 0 + f1 )

x0

For degree = 4, the table gives the Bode's rule

h=
x4

x 4 x0
4

f ( x)dx 45 (14 f

+ 64 f 0 + 24 f 0 + 64 f 0 + 14 f 0 )

x0

Using the function IntegrDataC is very easy.

Example. Calculate the integral with the Newton-Cotes formulas of degree = 1, 2, 4, 6 for the
dataset obtained by sampling the function sin(x)/x with step 0.2 . In a previous example, using
the Romberg method we have approximated this integral obtaining
Si(1.6) 1.38918048587044 with an accuracy better than 1E-9
Let's see now how the Newton-Cotes formulas work.

161

Xnumbers Tutorial

As we can see, the convergence to the exact result is evident. The most accurate result is
reached with the 6th degree Newton-Cotes formula. The IntegrDataC can work with random
samples.
Example. Given the data table (xi yi) , approximate the integral with the Cavalieri-Simpson
formula.

y
1
1.1
1.2
1.4
1.8
2.4
3.2
4
5

0
0.09531018
0.182321557
0.336472237
0.587786665
0.875468737
1.16315081
1.386294361
1.609437912

1.6
1.2
0.8
0.4
0
0

Note that the data points are not equidistant.

The points have been
extracted from the function
y = ln(x) .
Thus the exact integral is
5*ln(5)-4 4.0471896

Data integration for random points.

Having a set of not equidistant points (xi, yi), we cannot use directly the Newton-Cotes formulas
for fixed step.
In that case, IntegrDataC reorganizes the random data samples in equidistant data samples
and after that, computes the integral using the standard formulas for fixed step
Random Samples
{ (xi , yi) ; i = 0, 1, ... n }

Converted to

Equidistant Samples
{ (xi = x0+i h , yi (xi ) ; i = 0, 1, ... m }

For computing the values f(x0+i h ) at the equidistant grid points, IntegrDataC uses the Aitken
Interpolation algorithm.
Aitken interpolation algorithm.
Given a set of points:
{ (xi , yi)

i = 0, 1, ... n }
162

Xnumbers Tutorial
This method is used to find the interpolation point yp at the value xp .
It is efficient as the Newton formula, and it is also very simple to code.
For j = 1 To n - 1
For i = j + 1 To n
y(i) = y(j) * (x(i) - xp) - y(i) * (x(j) - xp)/ (x(i) - x(j))
Next i
Next j
yp = yi(n)

Function Integration (Newton-Cotes)

=Integr_nc(funct, a, b, Intervals, [Degree])
This function returns the numeric integral of a function f(x) using the Newton-Cotes formulas.
b

F (b) F (a ) = f ( x)dx
a

The parameter "Funct" is a math expression string in the variable x, such as:
"x*cos(x)",

"1+x+x^2",

"exp(-x^2)", ecc.. .

Remember the quote " " for passing a string to an Excel function.
"Funct" may be also a cell containing a string formula
The parameters "a" and "b" are the limits of integration interval
The parameter "Intervals" sets the number of sub-intervals of the integration interval.
The parameter "degree", from 1(default) to 10, set the degree of the Newton-Cotes formula.
degree = 1 coincides with the Trapezoidal rule; degree = 2 coincides with the CavalieriSimpson formula; degree = 4 with the Bode's rule.
Remember that the total knots of the function computation is:
knots = Intervals Degree + 1
Example: Approximate the following integral using 10 sub-intervals and three different
methods: trapezoidal, Cavalieri-Simpson, and Bode's rule.
1

xe

dx

The indefinite integral is known as the closed form:

xe

dx = ( x + 1)e x

So we can compare the exact result, that is

1 2e 1 0.264241117657115356

163

Xnumbers Tutorial

Integr_nc("x*exp(-x)",0,1,10,1) = 0.263408098685072

The trapezoidal rule,

with 10 sub-intervals,
requires 2 knots for
each sub-interval for a
total of 11 function
evaluations (11 knots)

0.4
0.35

(8.3E-04)

f(x)

0.3
0.25
0.2

The accuracy is better

then 1E-3

0.15
0.1
0.05

0
0

0.2

0.4

0.6

0.8

1.2

Integr_nc("x*exp(-x)",0,1,10,2) = 0.264241039074082

The Cavalieri-Simpson
rule, with 10 subintervals, requires 3
knots for each subinterval for a total 21
function evaluation (21
knots)

0.4
0.35

(7.8E-08)

f(x)

0.3
0.25
0.2

The accuracy is better

then 1E-7

0.15
0.1
0.05

0
0

0.2

0.4

0.6

0.8

1.2

Integr_nc("x*exp(-x)",0,1,10,4) = 0.264241117655293

The Bode's rule, with 10

sub-intervals, requires
5 knots for each subinterval for a total of 41
function evaluation (41
knots).

0.4
0.35

(1.8E-12)

f(x)

0.3
0.25
0.2

The accuracy is better

then 2E-12

0.15
0.1
0.05

0
0

0.2

0.4

0.6

0.8

164

1.2

Xnumbers Tutorial

Integration: symbolic and numeric approaches

The usual approach to the calculation of the definite integral involves two steps: the first is the
construction of the symbolic anti-derivative F(x) of f(x)

F ( x) = f ( x) dx
and the second step is the evaluation of the definite integral applying the fundamental
integration theorem.
b

f ( x)dx = F (b) F (a)

a

This approach can only be adopt for the set of the functions of which we know the antiderivative in a closed form. For the most f(x), the integral must be approximated either by
numerical quadrature or by same kind of series expansion.
It is usually accepted that symbolic approaches, when possible, gives more accurate result
then the numeric one. This is not always true. Even if the symbolic anti-derivative is known in a
closed form, it may often be unsuitable for further numerical evaluation. In particular, we have
cases in which such "exact" answers, when numerically evaluated, give less accurate results
than numerical quadrature methods1
Let's see. Assume to have the following integral functions

F ( x) =

( )

3x 2
dx = arctan x 3 + c
6
x +1

We want to calculated the definite integral between a = 2000 and b = 2004

The analytic approach gives:
F(b) -F(a) = arctan(b3) -arctan(a3)
In the following worksheet we have compared the evaluations with the exact anti-derivative and
the numerical quadrature with the Bode's rule. In the cell C2 we have inserted the antiderivative function.
=ARCTAN(B2^3)-ARCTAN(A2^3)
In the cell C2 we have inserted the Bode formula with 20 intervals
=Integr_nc(D1, A2, B2, 20, 4)

In the cell C5 we have also inserted the reference integral value

As we can see, the more accurate result is those obtained with the numerical quadrature;
surprisingly, it is more than 200 millions times more accurate than the one of the exact method!
It is evident from this example that only the symbolic integration could not resolve efficiently the
problem. For numerical integration the quadrature methods are often more efficient and
accurate.

"Improving Exact Integral from Symbolic Algebra System", R.J. Fateman and W. Kaham, University of California,
Berkeley, July 18,2000

165

Xnumbers Tutorial

Integration of oscillating functions (Filon formulas)

=Integr_fsin(Funct, a, b, k, Intervals)
=Integr_fcos(Funct, a, b, k, Intervals)
Oscillating functions can reserve several problems for the common polynomial integration
formulas. The Filon's formulas are suitable to compute efficiently the following integrals.
b

f ( x) sin(kx) dx

f ( x ) cos( kx) dx

for k = 1, 2, 3...N
The parameter "Funct" is a math expression string in the variable x, such as:
"x*cos(x)",

"1+x+x^2",

"exp(-x^2)", ecc.. .

Remember the quote " " for passing a string to an Excel function.
"Funct" may be also a cell containing a string formula
The parameters "a" and "b" are the limits of integration interval
The parameter "k" is a positive integer
The parameter "Intervals" sets the number of sub-intervals of the integration interval.
Remember that the total number of the evaluation nodes is: Nodes = Intervals 2 + 1
To understand the effort in this kind of numerical integration let's see this simple test. Assume
we have to evaluate the following integral for several integer values k, with 0 < k < 25

cos(kx) dx

If we perform the computation with the Cavalieri-Simpson formula (80 nodes) and with the Filon
formula (80 nodes), we get the following result
0
1.E+00

10

15

20

25

30

Cavalieri-Simpson

1.E-01

As we can see, the relative error

increases with the number k much
more for the Cavalieri-Simpson rule
than the Filon formula.
For k = 24 the first formula should have
400 nodes at least, for reaching the
same accuracy of the Filon formula.

Fillon

1.E-02

Relative error versus k

1.E-03
1.E-04
1.E-05
1.E-06
1.E-07
1.E-08

Example: evaluate the integral of the following oscillating function

2

1
sin(8 x)
+1

that can be rearranged as

2

g ( x) sin(8 x)

g ( x) =

where

1
x +1
2

The plot of the integration function and its envelope function g(x) are shown in the following
graph
166

Xnumbers Tutorial

1.2
1

1/(x^2+1)

0.8
0.6

1/(x^2+1)*sin(8*x)

0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1/(x^2+1)

-1
-1.2
0

Below, a simple arrangement to compute the given integral

The approximate error is less then 1E-8, with 300 intervals (default)

167

Xnumbers Tutorial

Integration of oscillating functions (Fourier transform)

= Fourier_sin(funct, k, [a], [param])
= Fourier_cos(funct, k, [a], [param])
These functions1 perform the numerical integration of oscillating functions over infinite intervals
+

f ( x) sin( k x) dx

f ( x) cos(k x) dx
a

If a = 0 (default) , these integrals are called "Fourier's sine-cosine transforms"

The parameter "funct" is a math expression defining the function f(x), not oscillating and
converging to 0 for x approaching to infinity:
"1/x", " 1/(8*x^2)", " exp(-b*x)", ecc.. .
Remember the quote " " for passing a string to an Excel function. "Funct" may be also a cell
containing a string formula
The parameter "k" is a positive number
The "Param" contains labels and values for parameters substitution (if there are)
These functions return "?" if the integral is not converging or if they cannot compute the integral
with sufficient accuracy
For finite integration see also Integration of oscillating functions (Filon formulas)
Example. Prove that is
+

sin x
dx =
x
2
0

The graph of the integration functions is

at the right.
Numerically specking, this integral is
very difficult to calculate for many
algorithms.
For example, the Bode adaptive quadrature needs more than 10.000 points for getting
accuracy of about 1E-4. The Fourier_sin function on the contrary is very efficient for this kind of
integral. The integral can be arranged in the following form
+

sin x
0 x dx =

1
sin x dx
x

That is the Fourier's sine transform of 1/x

These functions use the double exponential quadrature derived from the original FORTRAN subroutine INTDEO of
the DE-Quadrature (Numerical Automatic Integrator) Package , by Takuya OOURA, Copyright(C) 1996

168

Xnumbers Tutorial
We see that the accuracy is better than 1E-15. Note that the function Fourier_sin automatically
multiply the integration function f(x) for the factor sin(k*x), so we have only to pass the f(x)
expression.
Example. Verify that is
+

cos 4 x
dx =
8
x
0

The graph of the integration functions is

Observe that the integration function goes to
infinity for x approaching to 0.

Numerically specking this function is "terrible".

The integral can be arranged in the following form
+

cos 4 x
0 x dx =

1
cos 4 x dx
x

That is the Fourier's cosine transform of 1/x^0.5

The accuracy is better than 1E-15

Infinite Integration of oscillating functions

Generally, the infinite integration of real functions having a certain type of infinite oscillating tails
may give some problem even to the most efficient quadrature algorithms. These problems can
be avoided adopting specific integration tricks
Let's see some of them.
Example. Assume to calculate the following integral
+

cos( x ) cos(2 x )
dx
x

The integration function coverges to zero but it contains two oscillating terms. So we cannot
use directly the integr or integr_de function because they returns "?"
For solving we can use the Fourier's cosine trasform, separating each oscillating term.
The given integral can be re-arranged in the following way
+

cos( x ) cos(2 x )
cos( x ) cos(2 x )
dx =
dx +
x
x
0
1

cos( x )
dx +
x

cos(2 x )
dx
x

Note that the last two integrals cannot have the lower limit 0 because they do not converge for
x approaching to 0.

169

Xnumbers Tutorial
The first integral can be evaluted with the function integr and the two last integrals are
evaluated with the function Fourier_cos with a = 1. Let's see the following spreadsheet
arrangement

Compare the accuracy with the exact result I = Ln(2)

Example. Calculate the following integral
+

sin 4 ( x)
dx
x2

Remembering that is

3 cos(2 x) cos(4 x)
+
sin 4 ( x) =
8
2
8
The given integral can be arranged as
+

sin 4 x
sin 4 x
dx
=
0 x 2 dx +
x2

3
cos(2 x )
1 8 x 2 dx + 1 2 x 2 dx +

cos(4 x )
dx
8x 2
1

The first and second integral can be evaluted with the function integr and the two last integrals
are evaluated with the function Fourier_cos with a = 1.

Compare the accuracy with the exact result I = /2

170

Xnumbers Tutorial
Example. Calculate the following integral
+

sin (x ) dx
2

This function oscillates very badly. Note that the function does not converge to zero, oscillating
continuously from 1 and 1, but we can show that its integral is finite.

Let's perform the substitution

x2 = t

x= t

dx =

1
2 t

So, the given integral becomes

+

sin (x ) dx =
2

sin (t )
dt
2 t

That can be easily computed by the Fourier's cosine transform

Compare the accuracy with the exact result I = ( /8)1/2

171

dt

Xnumbers Tutorial

Double Integral
2D Integration for Normal Domains
Xnumbers contains routines for integrating bivariate functions f(x, y) over a normal domain
(normal to the x-axis and/or to the y-axis) or a circular domain.

g 2(x)

x2+y2 R2

d
h 1(y)

g 1(x)

h 2(y)

x
x

x
a
b
Domain normal to x-axis

Domain normal to y-axis

Circular Domain

For those kinds of 2D-domains the integration formulas can be re-written as the following
b g 2 ( x)

f ( x, y)ds = f ( x, y)dydx

DX

a g1 ( x )

d h2 ( y )

f ( x, y)ds = f ( x, y)dxdy
DY

c h1 ( y )

2 R

f ( x, y)ds =
C

f ( cos( ), sin( )) d d

0 0

Note that a normal domain implies that - at least - one axis must have constant limits.
Rectangular domains are a sub-case of normal domains in which both axes have constant
limits.
The macro Integr2D - suitable for integrating smooth functions f(x, y) and its function version
Integr_2D use the same bidimensional Romberg algorithm, but the function is limited to about
65.000 points.

Macro for Double Integration

Integr2D()
This macro performs the numerical integration of a smooth, regular function f(x, y) over a plane
normal domain D(x, y) .
X max Ymax

f ( x, y)dydx

X min Ymin

The integration functions f(x, y) and eventually also the bounding limits Xmin, Xmax,
Ymin, Ymax -can be written in symbolic expression. For further details about the math string
see Math formula string
The integration function can be:

bivariate functions like x^2+y^2-x*y, log(1+x+y), 1/(1+x^2+2*y^2), etc.

constant numbers like 0 , 2, 1.5, 1E-6 , etc.
constant expressions like 1/2, 2+1, sin(0.1), etc.
172

Xnumbers Tutorial

Boundary limits can be:

constant numbers like 0 , 2 , 10 , 3.141 , etc.

constant expressions like 1/2 , 2+1, pi, sin(1/2*pi) , exp(1) , etc.
mono-variable functions like x/2 , 3y-10 , x^2+x-1 , etc.

A normal domain has, at least, two constant boundary limits.

Function and limits can be passed to the macro directly or by reference. That is: you can write
directly the symbolic expressions or constants into the input-box of the macro panel or you can
pass the cells containing the expressions. This second mode is more easy and straight
Let see how it works.
Approximate the following double integral of the function f(x,y) = ln(1+x+y) in the closed region
delimited by the given constrains
Integration function

ln(1 + x + y )
Integration domain D

0 x 1
x2 y x +1
The domain D is shown in the above plot. As we can see, it is an x-nomal-domain domain.
Verify that the given integral approximates the symbolic expression at the right
1 x +1

f ( x, y)ds = ln(1 + x + y)dydx

D ( x, y )

0 x2

9 ln(3)
3 17
+ 7 ln(2)

4
12
8

The macro assumes as default

the following simple
arrangement (but, of course, it
is not obligatory at all)
Select the A2 cell and start the
Integr2D macro.

As we can see, the entire input boxes

are filled with the right references.
The output result will start from the A4
cell
Optionally we can adjust the Error limit
or the Rank. But usually the only thing to
do is click on the run button
Warning: The Rank increase
exponentially the computation effort,
because points = (2K)^2
"Parameter" input box may be used to
pass values of one or more parameters
eventually present in the integration
function
Activate "polar" for switching to the polar
coordinate system

173

Xnumbers Tutorial

The macro outputs 5 results:

1) Integral
2) Relative error estimation
3) Total points evaluated
4) Elaboration time
5) Error message (if any)

Example 2. Approximate the double integral of the function

[1 + 0.218( y x)]3.148 e 0.852 x

in the closed region delimited by

0 x 1 , 0 y x

The above expression can be written as

"(1+0.218*(y-x))^(-3.148)*exp(-0.852*x)"
or alternatively
"(1+218/1000*(y-x))^(-3148/1000)*exp(-852/1000*x)"
In both cases the result is obtained with high precision

Another elegant way to insert constant in the integration function is using parameters
Example. Evaluate numerically the following double integral

sin(nx) cos(my)dxdy
0 0

for n = 0.75 and m = 0.125

Set two cells, for example, F2 and G2, containing the values of n and m; just above, in the cells
F1 and G1, add their name "n" and "m" and pass the entire range F1:G2 as parameters.

The result will be

174

Xnumbers Tutorial

Compare with the exact result

1
128 2 + 256
3

Macro for Triple Integration

Integr3D()

3D Integration in normal domains

This macro calculates the triple integral of a function f (x, y, z) in a normal domain or in a
spherical domain or in a cylindrical domain.
For this kinds of domains the triple integral can be decomposed by the following formulas

xmax y max z max

f ( x, y, z )dxdydz = f ( x, y, z )dxdydz
D

xmin y min z min

Where the functions:

Xmin and Xmax may depend on variables y and z
Ymin and Ymax may depend on variables x and z
Zmin and Zmax may depend on variables x and y
The normal domain condition implies that, at least, one boundary must be constant
The macro can integrate in spherical coordinates system, using the formulas

x = sin cos
y = sin sin
z = cos
where , from 0 to 2, is the rotation angle on the xy-plane and , from 0 to , is the colatitude
angle
The triple integral changes in

max max max

f ( x, y, z )dxdydz = f ( x, y, z )
D

min

min

sin d d d

min

The macro can also integrate in cylindrical coordinates system, using the formulas

x = cos
y = sin
z=z
175

Xnumbers Tutorial
where , from 0 to 2, is the rotation angle on the xy-plane
The triple integral changes in

max max z max

f ( x, y, z )dxdydz = f ( x, y, z ) dddz
D

min

min

z min

Functions f(x, y, z) and eventually also the boundary functions may be written in symbolic
expression..
For further details about the math string see Math formula string
Integration function can be:
three-variate functions like x^2+y^2-x*z , log(1+x+y+z) , (x+z)/(1+x^2+y^2) , etc.
Constant numbers like 0 , 2 , 1.5 , 1E-6, etc.
Constant expressions like 1/2 , 2+1 , sin(pi/4) , etc.
Boundary limits can be:
Constant numbers like 0 , 2 , 10 , 3.141 , etc.
Constant expressions like 1/2 , 2+1, pi, sin(1/2*pi) , exp(1) , etc.
univariate or bivariate functions like x/2 , 3y-10 , x^2+x-1 , sin(x+z), Ln(x+y) , etc.
A normal domain has, at least, two constant boundary limits.
The Integration function and the boundary limits can be passed to the macro directly or by
reference. That is: we can write directly the symbolic expressions into the input fields, or you
can pass the cells that containing the expressions (simpler), or even a mixed mode.
Lets see how it works
Example 1. Approximate numerically the following triple integral
1 2 3

54 xyz dxdydz

0 0 0

The integration domain is the parallelepiped of lengths 1, 2, 3

The macro assumes as default the following simple layout (but, of course, it is not obligatory)

Now select the cell A2 containing the integration function and start the macro from the
Xnumbers toobar Macros > Integral > Triple.

176

Xnumbers Tutorial
As we can see, all the input
boxes are filled with the right
references.
The output result will start
from the A4 cell
Optionally we can adjust the
Rank, increasing it, if we want
to increase the final accuracy.
But usually the only thing to
do is click on the run button
Warning: The Rank increase
exponentially the computation
effort, because the number of
max points is (2*Rank)^3
Parameter input box may be
used to pass values of one or
more parameters eventually
present in the integration
function
The macro outputs 5 results:
1) Integral
2) Relative error estimation
3) Total points evaluated (function + boundaries)
4) Elaboration time
5) Error message (if any)
The results will appear as the following

Example 1. Approximate numerically the following triple integral

( x

+ y 2 ) dxdydz

where D is the domain defined by the following implict relation

x2 + y2 + z 2 1
This domain is a sphere of unitary radius. Therefore is better to pass to the spherical
coordinates because the integration domain becomes much simpler, being

0 1 , 0 2

The macro automatically performs the spherical transformation; so we have only to set the right
boundaries
Select the cell A2 and start the macro. Select "Polar" and set the Rank = 25. Press "Run"

177

Xnumbers Tutorial

Compare the result with the exact solution 8/15 = 0.53333...

Double integration function

=Integr_2D (Fxy, Xmin, Xmax, Ymin, Ymax, [Polar],[ErrMax])
This function returns the numeric integral of a smooth regular function f(x, y) over a plane
normal domain D(x, y) .
X max Ymax

f ( x, y)dydx

X min Ymin

The integration functions f(x, y) and eventually also the bounding limits Xmin, Xmax,
Ymin, Ymax -can be written in symbolic expression. For further details about the math string
see Math formula string
The integration function can be:
bivariate functions like x^2+y^2-x*y, log(1+x+y), 1/(1+x^2+2*y^2), etc.
constant numbers like 0 , 2, 1.5, 1E-6 , etc.
constant expressions like 1/2, 2+1, sin(0.1), etc.
The boundary limits can be:
constant numbers like 0 , 2 , 10 , 3.141 , etc.
constant expressions like 1/2 , 2+1, pi, sin(1/2*pi) , exp(1) , etc.
monovariable functions like x/2 , 3y-10 , x^2+x-1 , etc.
A normal domain has, at least, two constant boundary limits.
Example. Approximate the following double integral
1 x +1

ln(x + y + 1) dy dx
0 x2

3 17
9 ln(3)
+ 7 ln(2)

4
12
8

The integration domain is shown in the previous example.

In order to avoid long elaboration time, the function limits the total evaluation points to about
65.000 (rank = 8). For heavy computations use the macro Integr2D

178

Xnumbers Tutorial
Note: this function can also return the relative error, the total of evaluation points and the error
message (if any). To see these values simply select a range of two, three or four, adjacent cells
(vertical or horizontal) and give the CTRL+SHIFT+ENTER key sequence.
Example: Approximate the following integral
1

ex

0 y/2

1
dxdy
x + y2 +1
2

The integration domain is represented in the following plot

Integration domain D

0 y 1
y
x ey
2
As we can see the domain is
normal to the y-axes
The calculus of this double integral can be arranged as the following

Sometime, the integral my contain one or more parameters

Example
1 1

( a x
e

+ b y 2 )

dxdy

0 0

Note that the parameters "a" and "b" must be passed with their labels.

179

Xnumbers Tutorial
Example. Compute numerically the following integral

x 2 + y 2 dxdy

Where the domain D is the half-circular

region showed at the right

Set the parameter "Polar" to True.

Compare the precision with the exact result

180

Xnumbers Tutorial

Infinite integral
Integral_Inf()
This macro performs the numeric integration of a smooth, regular, not oscillating function f(x)
over an unlimited (or very long) interval

f ( x)dx ,

f ( x)dx ,

f ( x)dx

This macro can use two different methods:

The Bode formula with adaptive step

The double exponential algorithm

The Bode formula with 8 steps to calculate the integral and the truncation error.

I h1 =

2h
(7 f 0 + 32 f1 + 12 f 2 + 32 f 3 + 7 f 4 )
45

I h2 =

2h
(7 f 4 + 32 f5 + 12 f 6 + 32 f 7 + 7 f8 )
45

I h = I h1 + I h 2
ET

I h I h2
63

After each step the routine detects the truncation error and recalculates the step in order to
keep a constant error (variable step integration method).
The double exponential algorithm, also called "tanh-sinh quadrature". first introduced by
Takahasi and Mori, is based on the hyperbolic variable transformations.
It is more complicated then the polynomial Newton-Cotes schema but, on the other hand, it is
much more efficient.
Using this macro is very easy.
Example: Approximate the given integral

100 x 2 e x dx

The integration function is regular over the entire x-axes; the exponential assures the
convergence. Therefore the infinite integral exists.
Put the symbolic expression 100*x^2*exp(-x) in any cell that you like (A3 for example), and
arrange the worksheet in the following way (but it is not obligatory at all)
The word inf means of course infinity. It is not necessary to specify the sign, because the
macro always assumes b as +inf , a as -inf
Now select the cell A3 and run the macro Integral_Inf . The input fields will be automatically
filled

181

Xnumbers Tutorial
Choose run to start the integration routine. The result will be similar to the worksheet below
(without formatting) where we have compared the results of both methods

As we can see the integral is 100 with an excellent approximation for both methods but the
double exponential is more efficient. It required only 199 function evaluations.
Sometime we have to calculate the integral over the entire x- axes. Let see

x2 + 2x + 3
dx
x4 + x +1

Note that, in this case, we have needed more than 10.000 evaluation points for the variable
step method but only 344 for the DE method. The superiority is ever so evident? Not always.
There are cases in which the adaptive quadrature schema works better. For example when the
integration function has a finite discontinuity (jump) inside the integration interval; this usually
happens for the piecewise functions.
Example, Assume to have to compute the following didactical integral
+

sgn (1.5 x ) x

dx

The integration function is shown in the

following graph
In this case is easy to calculate the integral
simply separating the given interval [1, +] in
two sub-intervals: [1, 1.5] [1.5, +].
Calculating each integral and summing we get
the exact result I = 1/3.
But we want here to investigate how the two methods works in this situation

182

Xnumbers Tutorial

As we can see the variable step method has find the result with high precision using about
7200 steps. The double exponential algorithm fails the convergence
We have to put in evidence that using quadrature algorithms in a blind way, may lead to
wrong result. We should always study the integration function to discover singularities,
discontinuities, convergence rate, etc. If the integration function is sufficiently smooth, then
the numeric integration can give good approximate results.
This routine can also be used over a closed long interval, when other algorithms would take too
long computational time.

Double Data integration

=IntegrData2D(Dataxy, hx, hy)
Given a bidimensional set of points (xi, yj) defined by a rectangular grid

D {( xi , y j ) | xi = x0 + ihx , y j = y0 + jhy , i = 0...n , j = 0...m

this function computes the numerical double integral over the domain D

f ( x , y)dxdy
D

"Dataxy" is a (n x m ) rectangular array containing the function values f(xi, yj)

"hx" and "hy" are the grid intervals respectively of the x-axis and the y axis
Using this function is very simple.
Example assuming to have the following rectangular dataset (x, y) in wich each cell represent a
point of the grid and thus a value of the function f(xi, yj) where i = 0, 1, 2...20, j = 0, 1, 2...20

183

Xnumbers Tutorial

the function = IntegrData2D(B54:V74, C53, A55) returns the value 16.2661406

The scales x and y at the border of the range are not necessary for the function IntegrData2D.
They are designed only for clarity.
This function can also return the estimate error as second, optional value. To see also the error
select two adjacent cells and insert the function with the CTRL+SHIFT+ENTER keys

The given data set was generated by the function

f ( x , y) = e x / 3 + e y
with xi = 0.3 i and yi = 0.2 j
Therefore the above result is an approximation of the following integral
3 4

x/3

+ e y dydx

0 0

that is about 16.266082.

One interesting feature of this function is that
it accepts piecewise-rectangular domains
For example. Assume you have to
approximate the following integral

Ln(x + 2 y + 1)
D

where D is the domain shown at the right

This integral could be split, of course, into three rectangular integrals
2 3

3 3

2 1

0 3

0 1

1 0

f ( x, y)dydx + f ( x, y)dydx + f ( x, y)dydx

which the result is about 18.18803389

184

Xnumbers Tutorial
But it could be approximated in a worksheet sampling the domain for 0 <= x <= 3 and 0 <= y <=
4 with a suitable grid, for example hx = 0.1 , hy = 0.1

The cells out of the domain must be left empty.

The entire range is now B2:AF42 but the function will compute the integral only over the not
empty cells.
Note that in that case the scales x and y at the border of the range are used for calculating the
function f(xi,yi)
The integral and its estimate error will be

Note the good accuracy of this result.

IntegrData2D use the bidimensional trapezoidal formula with one-step of the Richardson
extrapolation. This algorithm works very fine with smooth regular functions and its precision is
similar the Cavalieri-Simpson formula

185

Xnumbers Tutorial

Series Evaluation
=xSerie(Funct, Id, a, b, [Param], [DgtMax])
Returns the numeric series of a function f(n) .
b

S = f ( n)
n=a

The parameter "Funct" is a math expression string such as:

"2^n/n*(-1)^(n+1)", "x^n/n!", "(-1)^(n)*(3+a)*x/(n-1)", ...
Remember the quote " " for passing a string to an Excel function.
For further details about the math string see Math formula string
"Id" indicates the integer index of the sum (usually "n", "k" , "i", etc.)
"a" and "b" are the limits of the sum.
The function may also have other parameters ("x", "y", "a", etc.) that can assume real values.
"Param" contains labels and values for parameters substitution (if there are). If we pass the
variable range without labels, the function will assign the values to the variables in the same
order that they appear in the formula string, from left to right.
The parameter "DgtMax" sets the multiprecision arithmetic. if omitted or zero the function uses
the fastest standard arithmetic
Example 1. Compute
10

(1) n+1
n =1

xn
n

for x = 2, with standard precision (15 digits) and with 25 digits.

The function substitutes x = 2 and then, computes the series f(n) for n =1, 2 ...10
10

(1) n+1
n =1

2n
2 2 23 2 4
210
= 21 + ...
n
2 3 4
10

xSerie("(-1)^(n+1)*x^n/n","n",1,10,2)= -64.8253968253968
Example 2. Compute
10

s=

n =0

xn
n!

for x = 1.5, with standard precision (15 digits) and with 25 digits. As known, this series
approximates the exponential e^(-1.5)

The function xSerie accepts one or more parameters.

186

Xnumbers Tutorial

Example 3. Compute the following series where a and b are parameters

10

s=

n =1

bn + a
n

for a = 0.7 and b = 1.5, in standard precision

Note that we have enclosed the labels "a" and "b" in the range B4:C5 passed to the function as
the argument "Param". The labels indicate to the function the correct assignment between the
variables and their values
Labels are optional. If we pass only the range B5:C5, without labels, the function assign the
values to the variables in the order from left to right.
Note how compact and straight is this calculation using the xSerie function.

Series acceleration with ^2

Many series are very slow to converge requiring therefore methods to accelerate their
convergence. The Aitken's extrapolation formula ( 2 extrapolation) can be used for this scope.
Practically we build a new series S(1), whose partial sums Sn(1) are given by the Aitken's
formula. It is possible to repeat the process starting from the series S(1) to obtain S(2)., and so
on.
Example. We want to approximate the following series:

S =
k =0

(1) k
1+ k

We know the exact result that is

= Log(2) = 0.693147180559945...
In the cell B4 we insert the formula
=Series($B$1;"k";0;A4)
In the cell C4 we insert
=(B4-LN(2))
we fill the rows from 5 to 16 simply selecting the range
B4:C4 and dragging it down.
In the last cell B17 we have inserted the function
=ExtDelta2(B10:B16)
performing the 2 extrapolation using the last 7 values
of the sum
As we can see, the cell B16 shows the sum with 12 terms; its approximation is very poor
having an error of more than 0.01. But if we apply the 2 extrapolation at the last seven partial
sums S(12), S(11), S(10) .... S(6) we have a good approximation with an error less then 1E-8
Note that for reaching this accuracy the given series would need more than 100 million terms!

187

Xnumbers Tutorial

Complex Series Evaluation

=cplxserie(Formula, a, b, [z0])
This function returns the numeric series of a complex function f(z, n) .
b

S = f ( z , n)
n=a

"Formula" is a math expression string defined by arithmetic operators and common elementary
functions such as:
"2^n/n*(-1)^(n+1)", "x^n/n!", "(-1)^(n)*(3+j)*x/(n-1)", ...
Remember the quote """ to pass a string to an Excel function.
The integer variable must be n.
The parameters "a" and "b" set the minimum and the maximum limits of the integer variable "n".
The function may have also a complex variable "z". In that case specify its value in the
parameter z0.
Example: evaluate the given series for z = z0 = 2-i
20

S =

n =1

z
z z
z
= z + + + ...
n
2 3
20

188

Xnumbers Tutorial

Double Series
= xSerie2D(Funct, Id1, a, b, Id2, c, d, [Param], [DgtMax])
Returns the numeric double series of a function f(n, m).
b

S = f (m, n)
m = a n =c

The parameter "Funct" is a math expression string such as:

" x^(n+2*m)/(n!*m!)", "(n+1)/(m+1)!", "comb(n,k)*a^k*b^(n-k)" ...
Remember the quote " " for passing a string to an Excel function.
For further details about the math string see the par. Math formula string
"Id1" , "Id2" indicate the integer indexes of the sum (usually "n", "m", "k" , "i", etc.)
"a" , "b", "c", "d" set the minimum and the maximum limits of the integer variables n and m.
The function may have other parameters ("x", "y", "a", etc.) that can assume real values.
"Param" contains labels and values for parameters substitution (if there are). If we pass the
variable range without labels, the function will assign the values to the variables in the same
order that they appear in the formula string, from left to right.
The parameter "DgtMax" sets the multiprecision. if omitted or zero the function uses the fastest
standard arithmetic
Example. Compute the following double series, in standard (15 digits) and multiprecision (25
digits)
4

s=

10

m = 0 n =1

x (n + 2m )
n!m!

for x = 0.8

Take care to the index limits because, for large interval, this function can slow down your Excel
application

189

Xnumbers Tutorial

Trigonometric series
= Serie_trig(t, period, spectrum, [offset], [angle], [smooth])
It returns the trigonometric series
N

f (t ) = f (0) + an sin (n t + n )

The set

(an , n )

n =1

2
T

, n = 1...N

is called spectrum of the function f(t)

Each couple is called harmonic.

The parameter t can be a single value or a vector values

The parameter period is the period T.
The parameter "spectrum" is an array of (n x 2) elements: the first column contains the
amplitude, the second column the phase.
The optional parameter "offset" is the average level (default 0)
The optional parameter "Angle" sets the angle unit: (RAD (default), DEG, GRAD)
Optional parameter smooth (default False) applies the Lanzos factor to the trigonometric series
N

f (t ) = f (0) + n an sin(n t + n )

n =

n =1

sin(n / N )
n / N

Here is a worksheet arrangement to

tabulate a trigonometric serie having
a spectrum of max 8 harmonics (the
formulas inserted are in blue)
The independent parameters are N
(samples) and T (periodo)
From those, we get the sampling
interval

T = T/(N-1)
The table at the left contains the
parameters for each harmonic: the
integer multiple of the harmonic, its
amplitude and its phase

Period T = 1
4

n Arm.
1
2
3

3.5
3
2.5

Amp
1
0
0.4

Phase
45
0
-45

2
1.5
1

Note that we can always transform

the cosine terms into sine with the
following formula

Period T

0.5
0
0

0.5

1.5

190

cos( ) = sin( + / 2)

Xnumbers Tutorial
The ringing phenomenon is an overshoot of Fourier series occurring at simple discontinuities
(jumps). The series does not converge uniformly to a discontinuous function in a small interval
of the discontinuity point. It can be proven that, at each discontinuity point, the series either
overshoots or undershoots by about 9% of the magnitude of the jump.
The ringing can be removed with the Lanczos sigma factors.
Example. The phenomenon is illustrated for a square wave.
Given the spectrum of the first 20 harmonics An = { 0, 1, 0, 1/3, 0, 1/5, 0, 1/71/19, 0 } and n
= 0, of the square wave with period T = 1, lets plot its Fourier series with and without Lanczos
sigma factors (parameter smooth = true / false respectively)
1.2
1
0.8
0.6
0.4
0.2
0
-0.2 0
-0.4
-0.6
-0.8
-1
-1.2

1.2
1
0.8
0.6
0.4
0.2
0
-0.2 0
-0.4
-0.6
-0.8
-1
-1.2

0.2

0.05

0.4

0.6

0.1

0.8

0.15

0.2

1.2

0.25

As we can see, the series with Lanczos factors (pink line) shows a more stable and smooth
behavior. Note, however, that ringing has interesting physical consequences. Consider, for
example, a linear electric circuit in which, by means of a switch, a fast voltage transition is
created: then the response of the circuit will really exhibit an overshoot

191

Xnumbers Tutorial

Trigonometric double serie

= Serie2D_trig(x, y, Lx, Ly, Spectrum, [offset], [Angle])
It returns the trigonometric double serie
N

f ( x, y ) = f 0 + an ,m cos( n x x + m y y + n , m )
n =1 m =1

where

x =
The set

[a

n ,m

2
Lx

, y =

2
Ly

, n ,m ] n = 0...N , m = 0...M

is called spectrum of the function f(t). Each couple is called "harmonic".

The parameters x and y are vectors
The parameters Lx and Ly are the base lengths of the x-axis and y-axis.
The parameter "Spectrum" is an array of (n x 4) elements: containing the following information:
index n, index m, amplitude and phase.
That is, for example:
n

Amplitude

Phase

45

0.5

-45

0.25

15.5

0.125

30

The optional parameter "offset" is the average level (default 0)

The optional parameter "Angle" sets the angle unit (RAD (default), DEG, GRAD)
The function f(x, y) is returned as an (N x M) array.
Use the CTRL+SHIFT+ENTER key to insert this function.
Example: Here it is a worksheet arrangement to tabulate a trigonometric serie f(x, y) having a
spectrum of max 4 harmonics

192

Xnumbers Tutorial

Discrete Convolution
Convol(f, g, h, [algo])
This function approximates the convolution of two sampled functions f(ti), g(ti)
+

f *g =

f (v) g (t v)dv

The parameters "f" and "g" are column-vectors

The parameter "h" is the sampling step.
The optional parameter "algo" sets the algorithm: 0 = discrete, 1 = linear (default), 2 =
parabolic.
The functions returns a vector with the same dimension of the two vectors f and g.
The convolutions is also called "Faltung"

Here are other examples of convolution plots

In the signals analysis the function f is called "input signal" x(t) and g is called "system impulse
response" h(t). The convolution fg is called "system responce" y(t)
The system behavior is reassumed in the following schema

193

Xnumbers Tutorial

x(t)

h(t)

y(t)

In a linear system, the outputs signal y(t) depends by the input signal x(t)
and by the inpulse responce of the system h(t). That is:
y(t) = x(v) h(t-v) dv
If the system is described by the following differential equation

y ' (t ) + k y (t ) = x(t )
which has an impulse response given by

h(t ) = e k t
We will use convolution to find the zero input response of this system to the square signal of
period T = 1 and amplitude xmax = 1.5

For obtaining this gaph we have used a sampling step of t = 0.02, but this value is not critical.
You can choose the size that you like in order to obtain the needed accuracy.

194

Xnumbers Tutorial

Interpolation

Polynomial interpolation
=PolyInterp(x, xi, yi, [degree], [DgtMax])
=PolyInterpCoef(xi, yi, [degree], [DgtMax])
The first function performs the interpolation of a given set of points (x,y), and returns the value
at the point x.
The Input parameters Xi and Yi are vectors.
The parameter "degree" sets the interpolation polynomial degree (default = 1)
The optional parameter DgtMax sets the maximum digits of the multiprecision arithmetic. If
omitted or 0, the functions works in faster double precision.
The second function returns an array containing the coefficients of the polynomial interpolation.
Use CTRL+SHIFT+ENTER for inserting this function.
This function use the following popular Newton's formula:
m 1
n

p( x) = y1 + D( x1 ,...xm ) ( x x j )
m =1
j =1

where D(x1,...xn) is the "divided difference", given by the following recursive formulas:

D( x1 , x2 ) =

D( x1 , x2 , x3 ) =

y1 y2
x1 x2

, D( x2 , x3 ) =

y 2 y3
,....
x2 x3

D( x1 ,...xm 1 ) D( x2 ,...xm )
D( x1 , x2 ) D( x2 , x3 )
, ..... D( x1 ,...xm ) =
x1 xm
x1 x3

Example : Compute the sub-tabulation between a set of 6 given points (x, y) using a linear
polynomial and a 3rd degree polynomial and a step of h = 0.2. Note that the given points are not
equidistant.

195

Xnumbers Tutorial
Sub-Tabulation with Degree = 1

Sub-Tabulation with Degree = 3

1.2

1.2

0.8

0.8
0.6

0.6
y Interp

0.4

y Interp

0.4

0.2

0.2
0

0
0

10

10

If we wonder which is the 3rd degree polynomial used for interpolating the point x = 3.2 we can
use the function PolyInterpCoef.

The interpolation polynomial will be

p( x) = a0 + a1 x + a2 x 2 + a3 x 3
with the coefficients [a0, a1, a2, a3] returned from the function PolyInterpCoef
We can plot this polynomial by the function Polyn. The interpolation polynomial is shown in the
above graph.
Note that the function has selected the 4 points [x2, x3, x4, x5], discharging the first point x1 and
the last point x6, in order to get the most possible accurate interpolation.
Note. when the number of the points (x, y) is exactly degree+1, then the polynomial is
univocally determined no matter which the interpolation point is: we can pass any number x
that we like, for example x = 0.
In the following example we determine the 2nd degree polynomial interpolating the given 3
points with 30 significant digits

196

Xnumbers Tutorial

Interpolation schemas
If the number N of the points (x, y) is exactly equal to degree+1, then the polynomial is
univocally determined; but if N > degree+1 there are several different polynomials that can be
used for interpolating a given point x. A popular method, adopted by the function PolyInterp is
called "sliding centered polynomial"
Linear interpolation

xi,

xi+1

The linear interpolation is always performed between the nodes xi xi+1 where x [xi xi+1]
For degree > 1 the interpolation strategy is more complicated.
Odd degree polynomial (example degree = 3)

x1, x2, x3, x4

xi-1 xi, xi+1, xi+2

xN-3, xN-2, xN-1, xN

The 3rd degree interpolation is performed between the nodes {xi-1 xi, xi+1, xi+2 } where
x [xi , xi+1]. At the first of the segment, the interpolation is performed between the nodes
{x1, x2, x3, x4 } where x [x1 , x3]. At the end, the interpolation is performed between the
nodes { xN-3, xN-2, xN-1, xN } where x [xN-2 , xN ].
Even degree polynomial (example degree = 2)
average interpolation

x1, x2, x3

xi-1 xi, xi+1,

Left polynomial

xi, xi+1, xi+2

xN-2, xN-1, xN

Right polynomial

The 2rd degree interpolation is performed between the nodes {xi-1 xi, xi+1 } and { xi, xi+1, xi+2 }
where x [xi , xi+1]. The first set gives the left interpolation polynomial while the second set
gives the right interpolation polynomial. In the range x [xi , xi+1] the final interpolation is
obtained taking the average of the two polynomials. At the first of the segment, the interpolation
is performed between the nodes {x1, x2, x3 } where x [x1 , x2]. At the end, the interpolation is
performed between the nodes { xN-2, xN-1, xN } where x [xN-1, xN ].
The average interpolation schemas adopted for even degree polynomials can often increase
the gloabal accuracy but, on the other hand, it reduces the efficence.
This example helps to explain this trick.
Assume to have the data set composed by 4
points; we want to perform the parabolic
interpolation of a point y(x) where 0.6 x 1
and y(0.6) = y2 and y(1) = y3

1
2
3
4

0.2
0.6
1
1.4

0.163746
0.329287
0.367879
0.345236

Note that the given points belong to the function y ( x ) = x e

197

Xnumbers Tutorial

Being the degree = 2, we can build

two interpolation polynomials using
the first 3 points and the last 3
points.
We have thus the left interpolation
polynomial PL(x), covering the
range [0.2, 1], and the right
interpolation polynomial PR(x)
covering the range [0.6, 1.4]
The intersection range is [0.6, 1]
that is just the range to interpolate

0.5
0.45

y = -0.1914x 2 + 0.4027x + 0.1566

0.4
0.35
0.3
0.25
0.2

y = -0.3967x 2 + 0.7312x + 0.0334

0.15
0.1
0

If we take the average

0.2

0.4

0.6

0.8

1.2

1.4

1.6

0.2

0.4

0.6

0.8

1.2

1.4

1.6

0.5

PM(x) = [PL(x)+ PR(x)] / 2

0.45

We obtain a new 2nd degree

polynomial interpolating the points
between [ 0.6 , 1]
Observe that it satisfys the nodes
constraints:
PM(0.6) = y2 , PM(1) = y3

0.35

0.4
0.3
0.25
0.2
0.15
0.1
0

If we compute the absolute errors |y(x) - PR(x)| , |y(x) - PL(x)|, |y(x) - PM(x)| in the range

0.6 x 1, we observe a significant error reduction for the average polynomial, as shown in

the following graph.

0.006

average

0.005

left

0.004

right

0.003
0.002
0.001
0
0.4

0.6

0.8

1.2

In this case the average-parabolic polynomial accuracy is comparable with the cubic
polynomial.
cubic

0.0016

average
0.0012
0.0008
0.0004
0
0.4

0.6

0.8

1.2

But we have to consider that the simple cubic interpolation schema is more efficient then the
average-parabolic. Generally odd degree interpolation schemas are preferable.

198

Xnumbers Tutorial

Interpolation with continued fraction

Fract_Interp_Coef(xi, yi, [DgtMax])
Fract_Interp(x, xi, coeff, [DgtMax])
These functions perform the interpolation with the continued fraction method.
Given, for example, a set of 5 points
xi = [ x0, x1, x2, x3, x4] , yi = [ y0, y1, y2, y3, y4 ]
the function Fract_Interp_Coef returns the coefficients vector [ a0, a1, a2, a3, a4 ] of the
continued fraction expansion given by the following formula:

y a0 +

x x0
x x1
a1 +
x x2
a2 +
x x3
a3 +
a4

The function Fract_Interp returns the interpolate value y at the point x

The optional parameter DgtMax sets the maximum digits of the multiprecision arithmetic. If
omitted or 0, the functions works in faster double precision.
Example: find the continued fraction interpolation for the following (x, y) samples

These points are extracted from the following function

y=

x2 + 2
x 2 0 .9

You can verify that the interpolation with these coefficients are better than 1E-14 for all x values
in the range [0.4 1.6]
Note also that this great precision is reached in spite of the pole (x 0.95) in the interpolation
range. The continued fraction interpolation is just suitable for interpolating rational functions.
The following graph shows very well how interpolate points fits the curve.

199

Xnumbers Tutorial
Interpolation with continued
fraction

50
40

y samples

30

y interp

The pink dots are the given

knots.
The blue line is obtained by
interpolation

20
10
0

There is a pole at x 0.95

(red dotted line)

-10
-20
-30
-40
-50
0.4

0.5

0.6

0.7

0.8

0.9

1.1

1.2

1.3

Note that the points outside

the range [0.5, 1.2] keep
also a good accuracy.
(extrapolation )

Example: Find an interpolation formula for the function tan(x) in the range 0 x 1.5 with no
more than 7 points.
The function tan(x) has a pole at x 1.57 , near to the upper bound 1.5; so its presence
suggests to adopt a fraction interpolation.
Assume to sample the function tan(x) at the
values (0, 0.2, 0.6, 1, 1.25, 1.45, 1.5).
The column A contains the knots of the
interpolations
In colum B we have inserted the
correspondent values of tan(x)
In column C we have computed the
coefficients of the fraction interpolation.

Using the function FracInterp we can interpolate any value between 0 and 1.5
obtaining the graph to the left. The second graph shows the absolute error in the given range.
You can verify that the interpolation is better than 1E-5 for any value x in the given range.

200

Xnumbers Tutorial

Interpolation with Cubic Spline

cspline_interp(Xin , Yin , Xtarget )
cspline_eval(Xin, Yin, Ypp, Xtarget)
These functions1 perform the natural cubic spline interpolation
Xin is the vector containing the x-values.
Yin is the vector containing the y-values..
Xtarget is the x value which we want to compute the interpolation
Xpp is the vector containing the 2nd derivative
The cubic spline interpolation is based on fitting cubic polynomial curves through all the given
set of points, called knots
The cubic spline follows these rules:
the curves pass through all the knots
at each knot, the first and second derivatives of the two curves that meet there are
equal
at the first and last knot, the second derivatives of each curve is equal to 0 (natural
cubic spline constraint).
The natural cubic spline has a continuous second derivative (acceleration). This characteristic
is very important in many applied sciences (Numeric Control, Automation, etc...) when it is
necessary to reduce vibration and noise in electromechanical motions. The drawback is that
cubic spline is less efficient than other interpolation methods.
The function cspline_eval is faster than cspline_interp, because it uses the information of the
2nd derivatives and does not have to calculate them all over again like the cspline_interp does.
The 2nd derivatives can be computed by the function cspline_pre (see next function)
Example:

Cubic Spline 2nd derivatives

cspline_pre(Xin , Yin)
This function Returns the cubic spline 2nd derivatives at a given set of points (knots).
Xin is the vector containing the x-values.
Yin is the vector containing the y-values..
For n knots, it returns an array of n 2nd derivative values. The first and the last values are zero
(natural spline constraint).
The 2nd derivatives depend only by the given set of knots. So they can be evaluate only once
for the whole range of the interpolation. By cspline_eval we can compute faster interpolation.
1

These functions appear thanks to the courtesy of Olgierd Zieba

201

Xnumbers Tutorial

Example. Perform the sub-tabulation with x = 0.1 of the following table

The given table is in the range
A3:B13
In the adjacent column C we have
computed the 2nd derivatives by
the function cspline_pre.
Note that this function returns a
vector of 11 values. It must be
inserted with the ctrl+shift+enter
keys sequence.
At the right we have set the new
table with step 0.1.

the value of F3 has been interpolate by the formula

= cspline_eval($A$3:$A$13; $B$3:$B$13; $C$3:$C$13; E3)
The other values are computed simply by dragging down the cell F3.
The following plot shows the knots and the cubic spline fitting

The points of the original table was extracted from the function y = [cos(x)]4 .
You can verify that the interpolation accuracy is better than 1% over the entire range.

202

Xnumbers Tutorial

Cubic Spline Coefficients

cspline_coeff(Xin , Yin)
This function1 returns the coefficients of the cubic spline polynomials
Xin is the vector containing the x-values.
Yin is the vector containing the y-values..
It returns an (n-1 x 4 ) array where n is the number of knots. Each row contains the coefficients
of the cubic polynomial of each segment s [as,3 as,2 as,1 as,0]

ys = as ,3 ( x xs )3 + as , 2 ( x xs ) 2 + as ,1 ( x xs ) + as ,0
where s = 1, 2, (n-1)
Example. Find the cubic spline polynomials that fit the given knots

These functions appear thanks to the courtesy of Olgierd Zieba

203

Xnumbers Tutorial

2D Mesh Interpolation
=Interp_Mesh(TableXY)
This function performs the linear interpolation of a bivariate function given in a pivot table XY.
The x-values and y-values must be sorted but not necessarily equidistant. This function returns
an array. Let's see how it works.
Regularization
As we can see, the use of this function
is straight. Simply select the area
where you want to insert the new
table, insert the function Interp_mash
and pass the old table as parameter.
The function Interp_mesh returns the
equidistant-linear-interpolated array.
Or, in other words, it returns the
regularized table

Rescaling
We can obtain a sub-tabulated
function in an very fast way
Simply select a larger area
The function Interp_mesh
counts the cells that you have
selected and fill all the cells with
the linear interpolated values
In this case the given table has
5 x 4 = 20 values.
The new table has 9 x 7 = 63
values; therefore, there are 43
new interpolated values

204

Xnumbers Tutorial

Macro Mesh Fill

This macro, accessible from the menu Macro / Regression, is useful for completing a 2D mesh
by interpolation. Practically, given a (n x m) matrix with holes, it tries to recover the missing
data. We could say a special case of bi-dimensional linear regression
Using this macro is very simple. Select the matrix with holes and start the macro.
Example
Original data matrix complete filled.
It was generated by the following linear
formula

aij = 2i 3 j

the same data matrix with same missing

values (holes). Select the matrix (range
B13:H21) and start the macro

The macro will try to rebuild the missing

value (red) by linear interpolation, using
the data of the nearest nodes
Of course, this process gives exactly the
original matrix only in the linear case
For all other case we have an error:
Generally, it is small if also the ratio
between holes and filled nodes is small
and if the original data is smooth

205

Xnumbers Tutorial

Interpolation of tabulated data function

Given a tabulated data (x i , y i), i = 1...N, generally not equidistant, we have to interpolate the
function y for an arbitrary x value, where x1 x xN
The points (x i , y i) are called knots of the interpolation

Cubic Spline interpolation

The goal of cubic spline interpolation is to get a polynomial interpolation formula that is smooth
in the 1st derivative, and continuous in the 2nd derivative, within the interval and at each
boundaries.
This method ensures that the functions y(x), y(x), and y(x) are equal at the interior node
points for adjacent segments. The cubic polynomials Pi(x) satisfie these constrains.
Pi(xi1) = yi1
Pi(xi) = yi
P'i(xi) = P'i+1(xi)
P''i(xi) = P''i+1(xi)

for i = 2...N
for i = 2...N
for i = 2...N1
for i = 2...N1

One form to write the interpolation polynomials is:

P(x) = A Pi + B Pi+1 + C Pi + D Pi+1

, for i = 1...(N-1)

Where:
A = (xi+1 x)/ (xi+1 xi)
B= 1A
C = 1/6 (A3 A) (xi+1 xi)2
D = 1/6 (B3 B) (xi+1 xi)2
The 2nd derivatives can be evaluated by the following linear equations
(xi x i-1) Pi-1 + 2 (xi+1 xi-1) Pi + (xi+1 xi) Pi+1 = Hi

for i = 2...(N-1)

where:
Hi = 6[(Pi+1 Pi)/(xi+1 xi) (Pi Pi-1)/(xi xi-1)] , P1 = 0 , PN = 0
That gives the following tridiagonal matrix system
2(x3 x1)

(x3 x2)

...

P2

H2

(x3 x2)

2(x4 x2)

(x4 x3)

...

P3

H3

(x4 x3)

2(x5 x3)

(x5 x4)

...

P4

(x5 x4)

2(x6 x4)

...

...

P5

...

...

...

...

...

(xN xN-1)

...

...

(xN xN-1)

2(xN xN-2)

PN-1

Another common way to write the interpolation polynomial is:

P(x) = a3i (x xi)3 + a2i (x xi)2 + a1i (x xi) + a0i
for i = 1...(N-1)

xi x < xi+1

Where the coefficients are:

a3i = (Pi+1 Pi)/(xi+1 xi)/6
a2i = Pi / 2
a1i = (Pi+1 Pi)/(xi+1 xi) (xi+1 xi) (2 Pi + Pi+1)/6
a0i = Pi
The matrix of the system is tridiagonal, therefore can be solved in O(N) operations
206

H4
H5

HN-1

Xnumbers Tutorial
We note also that its solution (P1 , P2 , ... PN ) depends only by the given knots, therefore the
2nd derivatives can be evaluated only once for any interpolate.
This example shows how the interpolation spline works.
Y

X
0
1
2.5
3
4
5

Assuming to have to sub-tabulate with a step

x = 0.1 a given function known only in the
following 6 points
Note that these points are unequal spaced

0
2
4
3
4
1

For these 6 knots we obtain 5 cubic polynomials having the following coefficients
Polynomials
1st spline
2nd spline
3rd spline
4th spline
5th spline

a3

a2

a1

Range
0x<1
1 x < 2.5
2.5 x < 3
3x<4
4x<5

a0

0.20148927

1.79851073

-0.8783764

0.60446781

2.40297854

5.54708717

-3.348226

-1.7126588

-3.0718353

4.97240473

-0.9005694

1.41436706

-4.2431012

-0.1712659

In the graphs below we can see the interpolated points (dotted line) fitting the data points and
the cubic polynomials (green line) passing through the nodes of each segment. Each
polynomial interpolates inside the proper segment. That is: the 1st spline works for 0 x < 1,
the 2nd spline for 1 x < 2.5, and so on.
In the graphs below are shown the entire interpolation line (left) and the 1st spline (right).
6

spline Interp

knots

0
-1

-1

spline Interp
knots
1st spline

0
1

-1

-2

-1

-2

In the graphs below are shown the 2nd spline (left) and the 3rd spline (right)
6
5

0
-1

-1

spline Interp
knots
3rd spline

spline Interp
knots
2nd spline

0
1

-1

-2

-1

-2

In the graphs below are shown the 4th spline (left) and the 5th spline (right)

207

Xnumbers Tutorial
spline Interp
knots
4th spline

6
5

1
0

0
-1

spline Interp
knots
5th spline

-1

-1

-1

-2

-2

Lets examine the 1st and 2nd derivatives. We can compute them either analytically or
numerically using for example -the following derivative formulas:
y(xi) (yi+1 2 yi + yi-1)/ x2

y(xi) (yi+1 yi-1)/2x

In both ways, we get the following graphs
Y'

Y''
15

3
2

10

1
0
-1 0

-2

-3

-5

-4
-5

-10

As we can see the 1st derivatives is smooth and the 2nd is continuous. This last feature is
particularly appreciated in many fields of engineering. Although this algorithm is less efficient
than other polynomial interpolation methods, it has the advantage of following the interpolated
curve without the spurious oscillations that other schemes can create

208

Xnumbers Tutorial

Other test functions

In our last example we have found that both methods can provide acceptable interpolation for
all range of x . Thus, there are same case, that the superiority of the spline interpolation is more
evident. Gerald [2] used the bump test case to illustrate problems with other interpolation
methods. Lets see.
Interpolate the following knots
Y = (cos(x))10 , for x= -2 , -1 , -0.5 , 0 , 0.5 , 1 , 2
Plotting the interpolated values with a step of 0.1, we get the following graphs
1.2

1.2

cos(x)^10

1
0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

-1

cubic spline

cubic poly

0
-2

cos(x)^10

-2

-1

-0.2

-0.2

The curves appear acceptable in both graphs. The second shows a closer fit near the points x
= 1 and x = -1 where are "knees" of the curve.
But matching the error plots, we see clearly the better accuracy of the spline interpolation.
0.06

As we can see, the amplitude

error of the cubic polynomial is
much more than the spline.

0.04
0.02

We can show that an higher

order of the interpolation
polynomial, is even worst

0
-0.02
error cubic poly

-0.04

error cubic spline

-2

-1.5

-1

-0.06
-0.5
0

0.5

1.5

209

In this case, the cubic spline is

the better choice

Xnumbers Tutorial

Continued fraction interpolation

Continued fractions are often a powerful ways of interpolation when we work near the functions
poles.
In XNUMBERS the continued fraction coefficients can be obtained by the function
FracInterpCoef and the interpolation value with FracInterp
Example. Interpolate the following dataset
y(x) = 1/(x+0.01)1/2 , for x = 0, 0.05, 01, 0.2, 0.5, 1 , with step x = 0.1
In the graphs below we have plotted the interpolated values obtained with three different
methods: cubic polynomial, cubic spline and continued fraction
Note that y(x) have a poles in x = 0.01 very near to the node x = 0
12

12

10

10

8
y

y
cubic
poly

knots
4

0
0

0.2

0.4

0.6

0.8

1.2

12

0.2

0.4

0.6

0.8

1.2

12
y

10

10

cubic spline

fraction

0
0

0.2

0.4

0.6

0.8

1.2

0.2

0.4

0.6

0.8

1.2

We can see a good general accuracy except for the final part of the polynomial interpolation
method. In this case, the worst accuracy is concentrated where the function is more flat, but,
surprisingly, this perturbation is due to the distant pole in x = -0.01. We note also that both
spline and fraction methods keep a good accuracy also for point external at the interpolation
range ( extrapolation for x >1 )
0

0.2

0.4

0.6

0.8

10
1
0.1
0.01

poly
spline
fraction

0.001
0.0001

210

Method
avg. err.
Cubic spline
0.22
Cubic poly
0.19
Contined fraction
0.003

Xnumbers Tutorial

Differential Equations
Xnumbers contains functions for solving the following differential problem of the 1st order with
initial conditions (Cauchys problem):

y = f (t , y )

, y (t0 ) = y0

and for solving the 1st order ordinary differential system.

y = f (t , y )

, y (t0 ) = y 0

y1 = f1 (t , y1 , y2 ... yn )
y1 (t0 ) = y10
y = f (t , y , y ... y ) y (t ) = y

2
2
1
2
n
20
, 2 0

....
....

yn = f n (t , y1 , y2 ... yn ) yn (t0 ) = yn 0

ODE Runge-Kutta 4
= ODE_RK4(Equations, VarInit, Step, [Par, ])
This function integrates numerically a 1st order ordinary differential equation or a 1st order
differential system, with the Runge-Kutta formula of 4th order

k1 = f (ti , yi )
k 2 = f (ti + 0.5 h, yi + 0.5 h k1 )
k3 = f (ti + 0.5 h, yi + 0.5 h k 2 )
k 4 = f (ti + h, yi + h k3 )
yi +1 = yi +

h
(k1 + 2k 2 + 2k3 + k 4 )
6

"Equations" is a math expression string containing the equation to solve. For a system It is a
vector of equations. Examples of correct equation definition are:
y = -2*y*x

v = 2*x-v^2+v

y1 = -3*y1+y2+sin(10*t)

Each string may contain symbolic functions with variables, operators, parenthesis and other
basic functions.
The parameter "VarInit" is a vector containing the initial values. It has two values for two
variables [ t0, y0 ].
For a system with n+1 variables, "Varinit" is an (n+1) vector [ t0, y10 , y20, .yn0 ].
The parameter "Step" is the integration step.
The optional parameter "Par" contains the values of other extra parameters of the equations.
Lets see how it works with an example
Solve numerically the following Cauchys problem for 0 x 3

y = 2 xy

, y ( 0) = 1

x2

We know that the exact solution is y = e

We can arrange a worksheet like the following

211

Xnumbers Tutorial

As we can see, we have written in cell G5 the differential equation

y = - 2*x*y
In the range A5:B5 we have inserted the starting values of x and y. Note that we have written
the labels just above theirs values. Labels are necessary for the correct variables assignment
Finally, in the range A6:B6 - just below the starting values - we have inserted the ODE_RK4,
that returns the value y(0.2) = 0.9607893 with a good accuracy of about 1E-7 (compare with
the exact solution)
Tip: In order to get all other values,
select the range A6:B6 and simply
drag it down.
The cells below will be filled
automatically
Only remember to fix the constant
cells in the function with the $
symbol
=ODE_RK4($G$5,A5:B5,$F$5)

We have also added the column with the exact values in order to check the approximation
error. Both exact and approximated solutions are plotted in the following graph
1.2
y

y exact

0.8
0.6
0.4
0.2
0
0

0.5

1.5

2.5

The approximate solution (dots) fits very good the exact solution (pink curve).

212

Xnumbers Tutorial
If you need, you can include parameters inside the differential equation
Example. Solve the following differential problem

y ' = k x n y
y (0 ) = 1
where k = 2 and n = 1
Note that we have added the
labels "k" and "n" above the
cells D2 and E2. In this way,
the parser will correctly
substitute the value 2 for the
variable "k" and 1 for the
variable "n". in the differential
equation
Do not forget the labels "x"
and "y" in the cells A5 and B5

Example: Solve the following linear differential equation

y '+

1
y = a x n , y (1) = 0
x

For n = 3 and a = 1. Rearranging, we get

y' = a x n

y
, y (1) = 0
x
Note the labels "a" and "n"
above the cells D2 and E2. In
this way, the parser will
substitute the value 1 for the
variable "a" and 3 for the
variable "n". in the differential
equation
Do not forget the labels "x"
and "y" in the cells A5 and B5

With the step h = 0.1, we have a numerical solution with a very good approximation comparing
with the exact solution y = ( x 1) /(5 x ) , (better than 1E-6)
5

Relative Error
8.E-07
7.E-07
6.E-07
5.E-07
4.E-07
3.E-07
2.E-07
1.E-07
0.E+00
0

0.5

1.5

213

2.5

3.5

Xnumbers Tutorial
This function can be used to solve ordinary differential systems.
Example: Solve numerically the following differential system, where v(t) and i(t) are the voltage
and the current of an electric network

v' = i 7 v

i ' = 5 i + 15 v

v(0) = 10

i (0) = 0
The computation can be arranged as
following.
Write the variables labels in the row 6. The
labels v and i must be the same that
you have written in the equations. Just
one row below, insert the starting values in
the same order.
Select the range A8:C8 and insert the
function ODE_RK4. The first step will be
returned.
Now select this row and drag it down for
evaluating all the steps that you need

The graph below shows the transient of v(t) and i(t) with good accuracy
Note that you can change the
step h in order to re-compute
the transient in a very fast and
quick way.

12
v

10

8
6
4
2
0
0

0.4

0.8

1.2

Optional constant parameters can be arranged. For example if you want to add a parameter R,
independent from the time t, write:
Constant parameters can be
written in any part of the
worksheet. You need only to add
the labels with the same symbols
with they appear in the differential
equations. In this case, we have
added the label "R" in the cell C1,
upon its values.
You can add as many optional
parameters that you like

214

Xnumbers Tutorial

ODE Multi-Steps
Another very popular method for integrating ordinary differential equations adopts the multistep Adams formulas. Even if a little formally complicated, they are very fast, and adapted to
build a large family of ODE integration methods
The multi-step Adams formulas can be generally written as:

yi +1 = yi +

h
M

yi +1 = yi +

h
M

where

k =1

N k

y 'i k +1 = yi +

h
( N 1 y'i + N 2 y 'i 1 +... 0 y'i N +1 )
M

N k

y 'i k + 2 = yi +

h
( N 1 y 'i +1 + N 2 y 'i +... 0 y 'i N + 2 )
M

k =1

y 'i = f (ti , yi )

ti = t 0 + h i

The first formula generates the explicit formulas also called predictor formulas.
The second formula generates the implicit formulas also called corrector formulas.
The number N is the order of the formula. A formula of N order requires N starting steps. Of
course, formulas with high N are more accurate.
For N = 1 we get the popular Euler integration formulas

yi +1 = yi + h y 'i

Eulers predictor (1 step)

h
yi +1 = yi + ( y 'i +1 + y 'i )
2

Trapezoid formula corrector (1 step)

Theirs errors are given by

1 2 ( 2)
h y
2
1
e h 3 y ( 3)
12
e

Error predictor 1st order

Error corrector 2st order

For N = 4 we get the popular Adams-Bashfort-Moulton predictor-corrector formulas

h
(55 y 'i 59 y 'i 1 +37 y 'i 2 9 y 'i 3 )
24
h
yi +1 = yi + (9 y 'i +1 +19 y 'i 5 y 'i 1 + y 'i 2 )
24

yi +1 = yi +

Predictor (4 step)
Corrector (4 step)

Theirs errors are given by

251 5 ( 5)
h y
720
19 5 ( 5)
e
h y
720

Error predictor 4th order

Error corrector 4th order

There are a large set of predictor-corrector formulas

215

Xnumbers Tutorial

Multi-step coefficents tables

The following tables list the coefficents for the Adams predictor-corrector formulas up to the 9th
order and relative errors
Multi-step Predictor coefficients
Ord.

10

12

24

720

1440

60480

120960

3628800

7257600

-1

-9

251

-475

19087

-36799

1070017

-2082753

-16

37

-1274

2877

-134472

295767

-9664106

20884811

23

-59

2616

-7298

407139

-1041723

38833486

-94307320

55

-2774

9982

-688256

2102243

-91172642

252618224

1901

-7923

705549

-2664477

137968480

-444772162

4277

-447288

2183877

-139855262

538363838

198721

-1152169

95476786

-454661776

434241

-43125206

265932680

14097247

-104995189

1
2
3
4
5
6
7
8
9

30277247

Multi-step Corrector coefficients

Ord.

10

12

24

720

1440

60480

120960

3628800

7257600

-1

-19

27

-863

1375

-33953

57281

-5

106

-173

6312

-11351

312874

-583435

19

-264

482

-20211

41499

-1291214

2687864

646

-798

37504

-88547

3146338

-7394032

251

1427

-46461

123133

-5033120

13510082

475

65112

-121797

5595358

-17283646

19087

139849

-4604594

16002320

36799

4467094

-11271304

1070017

9449717

2
3
4
5
6
7
8
9

2082753

Error coefficient
The general error is e -khOrd+1y(Ord+1) where k is given by the following table
Ord.
predictor
corrector

0.5

0.41667

0.375

-0.0833

-0.0417

10

0.34861 0.32986 0.31559 0.30422 0.29487 0.28698 0.28019

-0.0264

-0.0188

-0.0143

-0.0114

-0.0094

-0.0079

-0.0068

The predictor-corrector algorithm

Usually the multi-step formulas, implicit and explicit, are used together to build a PredictorCorrector algorithm . Here is how to build the 2nd order PEC algorithm (Prediction-EvaluationCorrection).
It uses the Eulers formula as predictor and the trapezoidal formula as corrector
Prediction
yp1 = y0 +h f(t0, y0)
yp2 = y1 +h f(t1, yp1)
yp3 =.

Evaluation

Correction

f(t1, yp1)

y1 = y0 +h/2 [f(t0, y0) +f(t1, yp1)]

f(t2, yp2)

y2 = y1 +h/2 [f(t1, y1) +f(t2, yp2)]

The value y1 can be reused to evaluate again the function f(t1, y1), that can be used in the
corrector formula to obtain a more accurate value for y1.
216

Xnumbers Tutorial
If we indicate the first value obtained by the corrector with y1(1) and the second value with y1(2)
we can arrange a new following schema
Prediction

Evaluation

yp1

Correction

f(t1, yp1)

y1(1)

Evaluation

Correction

f(t1, y1(1))

y1(2)

This is the so called PECEC or P(EC)2 schema.

The group EC can also be repeated m-times or even iterated still the convergence. In these
cases we have the schemas P(EC)m and P(EC) respectively.
Note that, for m >> 1 the final accuracy depends mainly by the corrector.
Let's come back to the PEC schema.
We note that, at the step, we use the value f(t1, yp1) to predict the new value yp2
We could increase the accuracy if we take the better approximation f(t1, y1).
The new schema becomes:
Prediction
yp1

Evaluation

Correction

f(t1, yp1)

y1

Evaluation
f(t1, y1)

This schema is called PECE and it is used very often being a reasonable compromise between
the accuracy and the computation effort.
Using different schemas with different predictor-corrector formulas we can build a wide set of
algorithms for the ODE integration. Of course they are not equivalent at all. Same of them have
a high accuracy, others show a better efficiency and others have a better stability. This last
characteristic may be very important for long integration intervals. In fact, the most algorithms,
especially those with higher order, become unstable when the integration step grows over a
limit. Algorithms that are stable for any integration step (so called A-stable algorithms) are
much appreciated, but unfortunately they require implicit formulas that, generally, can be
solved only by iterative algorithms.
The 2nd order trapezoid formula is one of the most popular A-stable formula.

217

Xnumbers Tutorial

Predictor- Corrector
= ODE_PRE(yn, f, h)
= ODE_COR(yn, fp, f, h)
These functions perform the integration of the ordinary differential equations with the popular
multi-step predictor-corrector Adams formulas

y = f (t , y )

, y (t0 ) = y0

The first function returns the predictor value yn+1,p while the second function returns the
corrector yn+1.
The parameter yn is the last point of the function y(t).
The parameter f is a vector containing the last N values of the derivative of y(t). That are the
last N-1 values of the corrector.
The parameter fp, only for the corrector, is the best approximation of the derivative of y(t) at
the step n+1. Usually it is provided by a predictor formula
The parameter h sets the integration step

PECE algorithm of 2nd order

Now we see how arrange a PECE algorithm of 2nd order to solve a the following differential
problem.

y ' = xy 2

, y (0) = 2

Let's set in a cell that we like the integration step h and then the heading of the data table. We
set separate columns for predictor and corrector values
Build the first row.
Begin to insert the starting
values (x0, y0) in the cells A6
and B6 respectively, and the
formula evaluations of f(x,y) in
the cell C6 and E6. The
corrector value is set equal to
the starting value B6

The second row is a bit more

complicated. Lets see.
Select the first row A6:E6 and
drag it down one row. This will
copy the formula for fp and fc
Insert in the cell A7 the
increment formula
xi+1 = xi+h

Now we have to add the predictor and corrector function

218

Xnumbers Tutorial
Insert in the cell B7
=ODE_PRE(yn, f, h)
yn is the last value of y(x).
contained in D6. f is the last
value of f(x,y) contained in E6.
h is the step B3.

Insert in the cell D7

=ODE_COR(yn, fp, f, h)
Where yn is the last value of
y(x). In that case is D6. f is
the last value of f(x,y), E6.
fp is the predicted. value of
f(x,y), C7 in this case.
h is the constan step.

Now the setting of the

PECE algorithm of 2nd order
is completed. Select the
second row A7:E7 and drag
it down in order to calculate
the steps that you want.

The yP and yC values can be compared with the ones of the exact solution.

y=

2
1+ x2

0.1

err predictor

0.08

The differences:

0.06

d ip = yip y ( xi )

0.04

d ic = yic y ( xi )

0.02

2nd PECE

are plotted in the graph at the right

We note clearly the characteristic
behavior of the predictor-corrector
algorithm.

-0.02

-0.04

The second formula refines the approximation of the first one.

The final accuracy of PECE algorithm is practically the accuracy of the corrector

219

Xnumbers Tutorial

PECE algorithm of 4th order

Now we solve the above differential equation with a 4th order PECE algorithm using the 4 steps
Adams-Bashforth-Moulton formulas

y ' = xy 2

, y (0) = 2

To start this algorithm needs 4 steps. A good set of starting steps is:
x

y(x)
0
0.2
0.4
0.6

2
1.9230769231
1.7241379310
1.4705882353

We do not investigate here how to get the extra 3 values (they

could comes by Runge-Kutta method or by Taylor series
approximation). The only thing that we have to point out is
that these values must be sufficiently accurate in order to
preserve the global accuracy of the algorithm

The first 4 rows of the PECE algorithm are built as shown in the previous example.
The first 4 values of yp and yc are
the same.
Now let's insert in the cell B10
=ODE_PRE(yn, f, h)
where yn is the last value of
y(x), D9 in that case.
f is a vector of the the last four
values of f(x,y), E6:E9 in this
case.
h is the step B3.

Insert in the cell D10

=ODE_COR(yn, fp, f, h)
where yn is the last value of
y(x). In that case D9.
f is a vector of the last 3 values
of f(x,y), E7:E9.
fp is the predicted value of
f(x,y), C10 in this case.
h is the step B3

T Now the setting of the PECE

algorithm of 4th order is completed.
Select the 5th row and drag it down in
order to calculate the steps you want.
he predictor-corrector error curves are
shown in the following graph

0.015

0.01
0.005

0
0
-0.005

0.5

1.5

2.5
ep
4th PECE

-0.01

220

Xnumbers Tutorial

In order to compare the accuracy of the solutions of the this algorithm with the 2nd order
algorithm of the previous example let's draw both the error curves in a same graph
0.01
0.008
0.006
0.004
0.002
0
-0.002

0.5

1.5

2.5

4th PECE

-0.004

2nd PECE

-0.006

As we can see, the 4th order algorithm is evidently more accurated then the 2nd order. On the
other hand, the first one requires an extra work for providing 3 starting points.

ODE Predictor-Corrector 4
= ODE_PC4(Equations, VarInit, Step, [Par, ])
This function integrates numerically an ordinary differential equation of 1st order or differential
system of 1st order, with the 4th order P(EC)^2 schema predictor-corrector of AdamsBashforth-Moulton.
The input arguments are identical to the function ODE_RK4
The function can return all the solution points in an array. Before inserting the function, select
as many rows that you want to fill. For example if you select a (p x n+1) array , the function will
return p solution points of a n x n differential system.
Lets see how it works with an example

v' = v(1 v)(v 1 2) w

w' = v w

v(0) = 0

w(0) = 1

(FitzHug-Naguno model)

Insert the equations in the cell A2 and A3. Insert the starting point [t0, v0, w0 ] = [0, 0, -1] in the
range A7:C7
Select the range A8:C167 in order to get simultaneously 160 integration points and insert the
function ODE_PC4 by the CTRL+SHIFT+ENTER sequence.

221

Xnumbers Tutorial
The following plot shows the obtained functions v(t) and w(t) for 0 <= t <= 8

The 4th order P(EC)2 schema predictor-corrector of Adams-Bashforth-Moulton is detailed in the

following flow-graph

Compute the next step

by 4th order by the 4th
order Predictor

Starting
x0, y0 , h

yp

Compute the first 3

steps by the 4th order
Runge-Kutta algorithm

Evaluate the function

fp = f (x4, yp)

fp
Starting points

Refine the step by the

th
4 order Corrector

x0, y0
x1, y1
x2, y2
x3, y3

yc
Evaluate the function
fc = f (x4, yc)

P (EC)2
Predictor-Corrector
schema of 4th order

right-shift the
points and repeat

x1, y1
x2, y2
x3, y3
x4, y4

fc
Refine the step by the
4th order Corrector

y4

The initial extra 3 steps, necessary for starting the predictor-corrector algorithm, are calculated
by the function ODE_RK4 requiring 4 function evaluations for step. After that, the PC schema
require only 2 function evaluations for step. Conventionally we measure the effort (EF) counting
the number of function evaluation for n steps.
EF = 12+2*n (for PC schema),

EF = 4*n (for RK algorithm)

For n >> 1 the effort reduction of the P(EC)2 schema is about 50%

222

Xnumbers Tutorial

ODE Implicit Predictor-Corrector

ODE_PC2I(Equations, VarInit, Step, [Par, ])
This function integrates numerically an ordinary differential equation of 1st order or a differential
system of 1st order, with the 2th order implicit Predictor-Corrector method. Thanks to its large
stability it is suitable for solving "stiff" problems.
The input arguments are identical to the function ODE_RK4
The function can return all the solution points in an array. Before inserting the function, select
as many rows that you want to fill. For example if you select a (p x n+1) array , the function will
return p solution points of a n x n differential system.
This function uses the 1st order Euler formula as predictor and the 2nd order trapezoidal formula
as the corrector.

yn +1 = yn + h f (t n , yn )

Predictor (Euler formula)

yn +1 = yn + h / 2 ( f (t n , yn ) + f (t n +1 , yn +1 ) )

Corrector (Trapezoidal formula)

The second non-linear equation is started with the predicted value and then solved respect to
the variable yn+1 till the convergence.
This algorithm, despite its low order, exibits a large stability and, thus, it is suitable for solving
"stiff" problems
Lets see how it works with an example
Solve numerically the following 2nd order Cauchys problem for 1 x 7 and k = 100

d2y
dy
+ (k + 1) + k y = 0
2
dx
dx

, y (1) = 1 , y ' (1) = 0.5

First of all, we transform the problem in a 1st order differential equations system taking

y1 = y , y2 = y '
y1 ' = y2

y2 ' = k y1 (k + 1) y2

y1 (1) = 1

y2 (1) = 0.5

A possible arrangement may be the following

We have written in cell A2 and A3 the
differential equations
In cell D3 we have inserted the k parameter
and in the cell E3 the h step
In the range A8:C8 we have put the starting
values for x and y1 and y2.
note that we have also added the labels "k" ,
"x", "y1", "y2" just above theirs values.
Labels are necessary for the correct
assignment in the symbolic equation.
Now select the range A9:C40 and insert the
function ODE_PC2I passing its arguments

The scatter plot of the 30 points (x, y1) is shown in the following graph
223

Xnumbers Tutorial

Compare the approximate solution (dotted line) with the exact one (pink continue line)
The fitting, even with only 30 points, looks good. If we try to solve this problem with the 4th order
Runge-Kutta algorithm we have to choose a step less then 0.025 for avoiding the instability
and, thus, we need more than 250 points for reaching the same integration interval.

The Lotcka-Volterra Model

The following two-dimensional differential system

dx
dt = a1 x a2 x y

dy = a y + a x y
3
4
dt
is called the Lockta-Volterra model or also "prey-predator" model. It is very useful in biology,
chemistry and many other fields. The numerical integration of this ODE family requires a stable
algorithm otherwise the result are not acceptable
The function x(t) and y(t), depending from the time, may represent different things. In a
biological models, for example, they simulate respectively the population of pray (rabbits) and
predator (foxes) at time t. The proportionality constants a1 , a2 , a3 , and a4 are positive. It is
known that its solution leads to oscillate steady-state.
. Let's see with a practical example

x' = 2 x 2 xy

y ' = 5 x + 2 xy

x(0) = 1

y (0) = 1
Insert the equation defintion in the cells
A1 and A2 , the step in the cell C2 = 0.02
and the starting values [0, 1, 1] in the
range A6:C6
Do not miss the labels "t", "x", "y"
Then, select the range A7:C407 and
insert the function ODE_PC2I with the
ctrl-shift-enter keys sequence

The following graph at the left shows the result of the function x(t) and y(t). By performing the
scatter plot of x-y variable we have the steady-state plot at the right. It shows a closed loop, a
limit cycle.
224

Xnumbers Tutorial

6
5
4

1
0
0

10

12

We can easily change the zoom of the graph by changing the integration step h
Note that the oscillations are quite deep. The above system has two equilibrium points where
dx/dt = 0 and dy/dt = 0 simultaneously, that can be found solving the algebric system

2 x 2 xy = 0

5 x + 2 xy = 0

x = 0
y = 0

x = 2.5

y =1

The non trivial solutions means that if the initial populations start from x0 = 2.5 and y0 = 1, then
there will be no oscillation and the population fraction x/y will be always constant in the time.
On the contrary, the populations become progressively to oscillate when the initial populations
are different from the equilibrium point [2.5 , 1].
Here same examples:
x(0) = 2.45 and y(0) = 1

x(0) = 2.4 and y(0) = 1

2.5

2.5

1.5

1.5

0.5

0.5

0
0

10

12

225

10

12

Xnumbers Tutorial

Differential Systems
OD Linear System
= ODE_SYSL(A, y0, step, [b]))
This function integrates numerically a system of ordinary linear differential equations of 1st
order with constant coefficients, starting from an initial value. For example, the general form of
a 3x3 linear differential system is:

y1 ' = a11 y1 + a12 y1 + a13 y1 + b1

y2 ' = a21 y1 + a22 y1 + a23 y1 + b2

y '= a y + a y + a y + b
31 1
32 1
33 1
3
3
where all the coefficients aij and bi are constant.
Such system can be put in the following handy matrix form.

y ' = [ A] y + b
where

a11
A = a21
a31

a12
a22
a32

and with the initial condition:

y = [ y1 , y2 , y3 ]

a13
a23
a33

b = [b1 , b2 , b3 ]

(0)

(0)

y0 = y1 , y2 , y3

(0) T

The constant term b is optional. If omitted the system is called "homogeneous"

This function uses the exponential expansion method that, for this kind of differential systems,
is both accurate and efficient.
The function returns an (n x m ) array containing all the nodes of the integration: m is the
number of equations; n is the numbers of the nodes. The function automatically sets n equal to
the rows of the range that you have selected before inserting it.
Let's see how it works practically
Solve the following homogeneous differential system with constant coefficients

y ' = [ A] y
where

20 10 19
6
16
A = 16
2 2
1

y 0 = [1 , 0 , 0]

For 0 x 4 and h = 0.05

The numerical solution can be arranged as in the following worksheet
The initial values are in the first row (range B7:D7). The column "x" was added only for clarity
but it is not indispensable at all. Select the range B8:D87, than insert the function ODE_SYSL
giving the suitable parameters A, y0, h. Then press CTRL+SHIFT+ENTER

226

Xnumbers Tutorial

All the 240 cells will be filled with the nodal solutions of y1, y2, y3.
The scale can be esily arranged simply by changing the parameter h
1.5

1.5

0.5

0.5

y1
y2
y3

0
y1

-0.5

-0.5

y2
y3

-1

-1
0

0.5

1.5

2.5

3.5

h = 0.05

0.25

0.5

0.75

h = 0.012

Compare with the exact solutions

y1 = 2e x + e 2 x + 2e 10 x

2 x
10 x
y2 = 2e 2e

x
2 x
y3 = 2e + 2e

High order linear ODE

The function ODE_SYSL can also be used to solve high order linear ODE with constant
coefficients, that in general can be written as

y ( n ) + an 1 y ( n 1) + ....a2 y ' '+ a1 y '+ a0 y = b

with the initial conditions

y ( x0 ) = y0 , y ' ( x0 ) = y0 ' , y ' ' ( x0 ) = y0 ' ' , , y ( n 1) ( x0 ) = y0( n 1)

As known, such ODE can be transformed into a linear differential system of 1st order.

y1 ' 0
y ' 0
2
L = K

y n 1 ' 0
y n ' a 0

1
0

0
1

K
0

K
0

a1

a2

0 y1 0
0 y 2 0

O
K L + L

1 y n 1 0
M
M a n 1 y n b
M
M

having the initial conditions

y 0 = [ y1 (0) , y 2 (0) , y 3 (0) , y n (0) ] = y (0) , y ' (0) , y ' ' (0) , y n 1 (0)
T

227

Xnumbers Tutorial
For example assume tha you have to calculate the solution of following IVP problem:

y ' ' '+5 y ' '+9 y '+5 y = 0 , with y (0) = 2 , y ' (0) = 3 , y ' ' (0) = 4
Introducing the axiliary variables y1 = y , y2 = y' , y3 = y'' , we get the following equivalent
differential system

1
0 y1
y1 ' 0
y
y ' = 0
0
1
2
2
y3 ' 5 9 5 y3

y1 (0) = 2

y2 (0) = 3
y (0) = 4
3

with

Observe that the last row of the matrix contains the coefficients of the given ODE with the
opposite sign; besides of that, it has only all "1" in the upper subdiagonal.
Insert the initial values in the first row (range B10:D10). The column "x" was added only for
clarity but it is not indispensable at all.Select the range B11:D40, where you want to otput the
results and insert the function ODE_SYSL giving the suitable parameters: A, y0, h.
Then press CTRL+SHIFT+ENTER
The selected area will be filled with the numerical solution of the given system
2.5
2
y1
Exact

1.5
1
0.5
0
0

Observe that the solution y1(x) is also the solution y(x) of the given ODE. Compare with the
exact solution

y ( x) = e x + e 2 x cos x
Of course the above differential system can be also solved with other methods.
In order to show the accuracy of the
exponential method we put in a graph the
average relative error
e(x) = |yi - y(xi)| / | y(xi)| ,
obtained in the same condition, with 3
different methods: Exponential, Euler,
and Runge-Kutta 4.
The graph is eloquent. The error of the
Exponetial method is several times more
accurate then the others

100
1
0.01
0.0001

Euler

1E-06

RK 4

1E-08

Expo

1E-10
1E-12
1E-14
1E-16
0

Clearly it takes advantages using dedicated methods for linear differential equations.
Another important feature of the exponential method is its high stability
Let' s try the following test stiff system

y1 ' = y2

y2 = 20 y1 21 y2

y1 (0.4) = 0.6706555

y2 (0.4) = 0.6770293

228

Xnumbers Tutorial
The exact solution is

y1 = e x + e 20 x

y2 = e x 20e 20 x
In the following graph we shows the numerical result obtained with three different integration
methods, Exponential, Runge Kutta of 4th order and Euler, using different integration steps
The Expo method reaches the integration length of about x = 19 with a step h = 0.4
1
0.8

Expo method
h = 0.42

0.6
0.4
0.2
0
-0.2 0

10

15

20

-0.4
-0.6
-0.8
-1

The Runge-Kutta method, using a step h = 0.14 gives accurate only for x < 3; after that the
solution begins to diverge.
1
0.8

Runge Kutta 4
h = 0.14

0.6
0.4
0.2
0
-0.2 0

10

15

20

-0.4
-0.6
-0.8
-1

The Euler method, with a step h = 0.11, begins to oscillate even from the first steps. At this step
the Euler method is completely unstable.
1
0.8

Euler method
h = 0.11

0.6
0.4
0.2
0
-0.2 0

10

15

20

-0.4
-0.6
-0.8
-1

As we can see the stability step of the Exponential method is here about 3 times greater than
the others methods.

229

Xnumbers Tutorial

ODE for integral function solving

The integral function y(x) that we want to
calculate is defined as

Taking the derivatives of both sides and

remembering Leibniz's rule, we have

y ( x) =

f ( x)dx

y ' ( x) = f ( x)

x0

y ( x0 ) = 0

As we can see, the computing of any integral function is equivalent of solving an ODE.
Example. Plot the integral function of f ( x ) = x e

Insert in the cell A1 the expression

y' = x*exp(-x)
Choose a suitable step, for example h = 0.15.
Select the range A7: B54 and insert the
function ODE_RK4 with crtl+shift+enter
In the adjacent column C we insert the
integration function f(x) = x*exp(-x)
Now select all the range A7:C54 and plot
together.
The result is shown in the following graph

1.2
1
0.8
0.6

f(x)

0.4

f(x)

0.2
0
0

Comparing with the exact integral function y ( x ) = 1 ( x + 1) e , we have obtained a good

solution with an average error of about 5E-7. For this example we have taken about 50 points.
Remark. This method works fine when the function f(x) is smooth together with its derivative
and has no singularity in the integration range. Otherwise we have to choose some quadrature
algorithm as the double exponential method (see chap Integration for further details)

230

Xnumbers Tutorial
Example. Solve the following linear differential equation of 3rd order.

y ' ' '+5 y ' '+33 y '+29 y = 0 with y (0) = 2 , y ' (0) = 3 , y ' ' (0) = 20
This linear equation with constant terms can be transformed into the equivalent differential
system:

1
0 y1
y1 ' 0
y ' = 0
0
1 y2
2
y3 ' 29 33 5 y3

with

y1 (0) = 2

y 2 ( 0) = 3
y (0) = 20
3

Solving numerically this differential

system is very rapid.
Simply write the system matrix A in the
worksheet, for example, in the range
B5:D7. Insert the row starting values,
[2,-3,-20] in the range B10:D10, where
you want to begin the functions
tabulation. Choose a suitable step that
you like, for example h = 0.05
Then, select the range B11:D60 and
insert the function
=ODE_SYSL(B5:D7;B10:D10;B3)
And give the ctrl+shift+enter key
sequence.
The selected cells will be automatically filled with the calculated solutions.
Compare with the exact solution

y = e x + e 2 x cos(5 x)
The following graph shows the exact solutions (continue lines) compared with the calculated
solutions (dotted lines). We note a global good fitting.

5
4
3
2
1
0
-1
y1

-2

y2

-3

y3

-4
-5
0

0.5

1.5

231

2.5

Xnumbers Tutorial

Macro ODE Solver

This macro performs the numerical integration of ordinary differential equations systems (ODE)
using the Runge-Kutta-Fehlberg method of 4th order.
It is useful for solving ODE with initial values (Cauchy problem)

y ' = f ( x, y ) , y ( x0 ) = y0

x0 x xmax

y ' = f ( x, y ) , y ( x0 ) = y 0

x0 x xmax

as well an ODE system

where y = y1 , y2 ... yn

]T

is the vector of depend variables

The macro works with equation definited by worksheet formulas (on site) or even with equation
definited by symbolic string. The macro implements a control of the local error and changes the
integration step consequently Therefore the precision can be set independently from the
tabulating step. The independence of the integration step from the tabulation step is the main
feature of this macro.
The macro implement a predefined simple input schema. Let's see how to use it with a simple
example
Example 2. Find a numerical solution of this simple differential system

y1 ' = y 2

y 2 ' = 4 y1 5 y 2

with

y1 (0) = 0

y 2 ( 0) = 2

and with 0 x 5

Arrange the worksheet reserving two columns for [y1 , y2 ] and the two adjacent columns for the
derivatives [y1', y2']. In the first column set a x-grid with 50 step of increment x = 0.1.
Then select the area A4:E55 and start the macro ODE Solver.

Differential Equations links to the cells where are the equation formulas: = C5, = (4*B5+5*C5) .
Variables are the cells (x, y1, y2) referenced by the formulas; in that case A4:C5
Starting values indicate the starting cells of the table (they may or not may coincide with the
variables cells).
If we have followed the above schema, the input fields are correctly filled.
We have only to set the required precision: Err. Max = 1E-6
Optionally, if we want the macro to fill also the derivative columns, activate the check box

Press "Run" to start the RKF integration algorithm. After few seconds and about 1400 function
evaluations, the macro output its result. Each solution node has a relative error less then 1E-6.
Compare it with the exact solution
232

Xnumbers Tutorial

y1 = e x e 4 x

y2 = e x + 4e 4 x

The following graphs shows the exact solution (pink line) and the calculated solution (dotted
line)

Solution y1(x)

Solution y2(x)

Note the general good fitting. It is a remarkable result overall if we consider the relative few
points in the initial range where both functions have high swinging.
For further details see " ODE tutorial" by Foxes Team, 2006

Macro ODE - Slope Grid

This macro generate and visualize the slope y'(x) of a 1st ODE solution over a rectangular
domain. It is didactically useful for studying the 1st order differential equation y ' = f ( x, y )
For example we have to study the following equation

y '+ y = e x

y ' = xe x y

Therefore we have to plot the slope given by the bivariate function f ( x, y ) = e

suitable rectangular domain, for example in D = {-2 x 2 , {-2 y 2 }

y in a

For using this macro, set the variables x and y in

adjacent cells, for example I4 and J4. Then, in the
adjacent right cell K4, insert the formula defining
y'. thus: = EXP(-I4)-J4

Select the range I4:K4 and start the macro ODE Slope Field

Option Tab

The Grid number, from 4 to 24, set the

density of the grid. The dots check box
adds the grid points to the plot. The axes
check box adds the x-y axes

233

Xnumbers Tutorial
In the Axes tab we can set the domain D. Click the "Draw" button to generate the plot
The macro generates a graphical object
showing the slope y' of each point of the
grid.
Every solution y(x) of the differential
equation follows the direction of the slope
field. Therefore the field gives a global
view of the solutions crossing the given
domain.
For example, the solution crossing the
x

point (0,0) is y = e x ( red line)

We can see that this solution evolves
following the slope directions. In this case,
it never crosses the slope in any point.

Other interesting examples

y ' = 2 xy

y' = x 1 y 2

234

Xnumbers Tutorial

Nonlinear Equations

Bisection
=Zero_bisec(a, b, func, [step], [param])
Approximates the zero of a monovariable function f(x) with the bisection method

f ( x) = 0
This function needs two starting points [a, b] bracketing the zero.
Parameter "func" is a math expression string containing the symbolic function f(x)
Examples of correct function definitions are:
-2*ln(x)

2*cos(x)-x

3*x^2-10*exp(-4*x) , etc.

The optional parameter "step" sets the maximum number of steps allowed. If omitted the
function iterates still the convergence. Step = 1 is useful to study the method step-by-step
Param contains labels and values for parameters substitution (if there are)
At the first step, the function returns a new segment
[ a1, b1] where a1 < x0 < b1
At the second step, the function return a new segment
[ a2, b2] where

a1< a2 < x0 < b2 < b1.

After several iterations, the interval [an , bn ], with n>>1, will be very closed to the value x0
Example: Find the approximated zero of the following equation and show the first steps of the
bisection method.

3 log10 ( x) +

2
log10 ( x)

5 = 0

The plot indicates two zeros: one trivial

x = 10 and another in the interval 2 < x < 9

Starting the algorithm with a = 2

and b = 9 we get
x0 = 4.64158883361278
The root approximates the exact
zero x0 = 1001/3
with an error < 1E-14

235

Xnumbers Tutorial
We can also solve this equation step-by-step in order to investigate how this algorithm works

As we can see, the convergence is quite low but very robust because the zero always remains
bracketed between the interval limits [a, b]. The error estimation is also very quick. Simply
taking the difference |b-a|

Secant
=Zero_sec(a, b, func, [step], [param], [DgtMax])
Approximates the zero of a monovariable function f(x) with the secant method

f ( x) = 0
This function needs two starting points [a, b] bracketing the zero.
This version accepts also a worksheet function.
The parameter "Func" is a math string containing the symbolic function f(x) or a range of two
adjacent cells containing the worksheet functions [ f(a) , f(b) ] (see example).
Examples of correct symbolic function definitions are:
-2*ln(x)

2*cos(x)-x

3*x^2-10*exp(-4*x) , etc.

The optional parameter "Step" sets the maximum number of iterations. If omitted, the function
iterates still the convergence. Step = 1 is useful for investigation about the algorithm.
"Param" contains labels and values for parameters substitution (if there are)
Both parameters "step" and "param" are ignored if funct is not a symbolic function.
The optional parameter "DgtMax" set the number of the precision digits. If omitted, the function
works in the faster double precision.
At the first step, the function returns a new segment
[ a1, b1] where a1 < x0 < b1
At the second step, the function return a new segment
[ a2, b2] where

a1< a2 < x0 < b2 < b1.

After several iterations, the interval [an , bn ], with n>>1, will be very closed to the value x0
Use the CTRL+SHIFT+ENTER sequence to paste this function
Example: Find the approximated zero of the following equation and show the first steps of the
secant method.

exp(3x ln(2)) sin( x) = 0

The plot indicates one zeros into the interval 0 < x < 0.5

236

Xnumbers Tutorial
exp(-6*x*Ln(2))-sin(pi*x)

1.5
1
0.5
0
-0.5

0.2

0.4

0.6

0.8

-1

Starting the algorithm with a = 0 and b = 0.5 we get x0 = 0.166666666666667

The root approximates the
exact zero x0 = 1/6 with
error < 1E-15

Lets see now the iteration trace setting the parameter step = 1

As we can see this convergence is much faster than the one of the bisection method. On the
other hand, it is no guaranteed that the zero always remains bracketed into the interval.

Zero finder for worksheet function

Example we want to find the zero of the function f(x) = x cos(x) , using the standard
worksheet function = COS(x). Arrange a worksheet like the following

237

Xnumbers Tutorial
Insert in A3:B3 the array function zero_sec(a, b, func) using the sequence Ctrl+Shift+Enter
In this case, none parameter is necessary because the function always returns one step of the
secant algorithm.
Now insert the functions in the range C3:D3 by simply dragging-down the range C2:D2

Insert the function =ABS(A3-B3) in E3. This is not necessary but it is usefull for checking the
convergence. Now, let's select the last row A3:E3 and drag it down until the convergence
shows the minimal error

We can compute this result also in multiprecision. In this case we have use, of course, the
equivalent
x-functions; in cell D2 = xsub(A4, xcos(A4)), E2 = xsub(B4, xcos(B4)) and in cell E3 =
ABS(xsub(A5, B5)).
In the range A3:B3 insert the array function {Zero_sec(A4, B4, C4:D4, , , 30)}.
The parameter DgtMax = 30 switches the function in multiprecision mode
Following the other steps we get the result below

As we can see, the final value is

x = 0.739085133215160641655312087673
that is the zero of the function f(x) = x cos(x) with 30 significant digits

238

Xnumbers Tutorial

Derivatives

First Derivative
=Diff1(x, func, [lim], [param])
Approximates the first derivative of an univariate function f(x) at the given point x

f ' ( x) =

d
f ( x)
dx

The parameter func is a math expression string containing the symbolic function f(x)
Examples of function definition are:
-2*Ln(x) ,

2*cos(x) ,

3*x^2-10*exp(-4*x) , x^2+4*x+1 , etc.

The optional parameter lim (default = 0) sets the way how the limit approach to x. If lim = 1, it
approaches from the right; if lim = -1, it approaches from the left;
if lim = 0 , it approaches centrally. That is, it returns the following derivatives

lim f ( x) = f ' ( x )
h0

f ' ( x) = lim f ( x) = f ' ( x)

h 0

f ( x) = f ' ( x + )
hlim
0+
"Param" contains labels and values for parameters substitution (if there are)
This function uses the following formulas to approximate each derivative
f '(x )

1
(25 f ( x) 48 f ( x h) + 36 f ( x 2h) 16 f ( x 3h) + 3 f ( x 4h) )
12h

1
( f ( x 2 h) 8 f ( x h) + 8 f ( x + h ) f ( x + 2 h) )
12 h
1
(25 f ( x) 48 f ( x + h) + 36 f ( x + 2h) 16 f ( x + 3h) + 3 f ( x + 4h) )
f ' (x+ )
12h
f ' ( x)

Example. Evaluate numerically the left, right and central derivatives of the given function at the
point x = 0, and check if the given function is differentiable in that point

f ( x) =

x
x + | x | +1
2

As we can see all derivatives are equal, so the function is differentiable in x = 0

239

Xnumbers Tutorial

Second Derivative
=Diff2(x, func, [param])
It approximates the second derivative of an univariate function f(x) at the given point

f ' ' ( x) =

d2
f ( x)
dx 2

The parameter "func" is a math expression string containing the symbolic function f(x)
Examples of function definition are:
-2*Ln(x) ,

2*cos(x) ,

3*x^2-10*exp(-4*x) , x^2+4*x+1 , etc.

"param" contains labels and values for parameters substitution (if there are)
Example: Evaluate the first and second
derivatives at the point x = 2 for the
following function

f ( x) =

x+3
x2 +1

Gradient
=Grad(p, func, [param])
Approximates the gradient of a multivariate function f(x, y, z) at the given point

f f f
f ( x , y , z ) =
,
,

x y z
The parameter "p" is the vector of the variables [x, y, z]
The parameter "func" is an expression string containing the function f(x, y, z). Examples of
function definition are:
-2*ln(x+3y), 2*exp(-x)*cos(3*t), 3*x^2-y^2+z^2, (x^2+y^2)^(1/3),etc.
For performance problem, the number of variables is restricted to 4, x, y, z, t. The
variables values must be always passed in this order.
"param" contains labels and values for parameters substitution (if there are)
Example. Evaluate the gradient of
the following function at the point
P(1, 1)

f ( x, y ) =

1
x + 5y2
2

240

Xnumbers Tutorial

Jacobian matrix
=Jacobian (p, func, [param])
Approximates the Jacobians matrix of a multivariate vector-function F(x, y, z) at the given point
(x, y, z)

f1 ( x, y, z )
F ( x, y, z ) = f 2 ( x, y, z )
f 3 ( x, y, z )

f1 f1 f1
x , y , z

f 2 f 2 f 2
,
,
J ( x, y , z ) =

y
z
x
f 3 f 3 f 3
,
,

y
z
x

The parameter "p" is the vector of the variables [x, y, z]

The parameter "func" is an expression string containing the function f(x, y, z). Examples of
function definition are:
-2*ln(x+3y), 2*exp(-x)*cos(3*t), 3*x^2-y^2+z^2, (x^2+y^2)^(1/3),etc.
For performance problem, the number of variables is restricted to 4, x, y, z, t. The
variables values must be always passed in this order.
"param" contains labels and values for parameters substitution (if there are)
Example. Evaluate the Jacobians matrix of the following vector-function at the point P(1, 1)

f1 ( x, y, z ) =

1
x2 + 5 y2 + z 2

f 2 ( x, y, z ) = z ln( x + 2 y )

f 3 ( x, y, z ) = 4 xyz

Hessian matrix
=Hessian (p, func, [param])
Approximates the Hessian matrix of a multivariate function f(x, y) at the given point p(x, y)

2 f
2 f
,
2

x
xy

H ( x, y ) = 2
f
2 f
,

2
yx y
The parameter "p" is the vector of the variables [x, y]
The parameter "Func" is an expression string containing the function f(x, y, z). Examples of
function definition are:
-2*ln(x+3y), 2*exp(-x)*cos(3*t), 3*x^2-y^2+z^2, (x^2+y^2)^(1/3), etc.
241

Xnumbers Tutorial
For performance problem, the number of variables is restricted to 4, x, y, z, t. The
variables values must be always passed in this order.
"Param" contains labels and values for parameters substitution (if there are)
This function returns a square a matrix (n x n) of the second derivatives
The accuracy of this function is about 1E-10
Example. Approx. the Hessians matrix of the following function at the point (2,1,1)

f ( x, y , z ) =

1
x + 5y2 + z2
2

Parameters. All the differential functions accept parameters into the string definition. They
must be different from the canonical variables "x", "y", "z", "t" but apart this, can have any name
that you like.
Example: Calculate the Hessian matrix of the function

f ( x, y , z ) =

kz
x + y2 + a
2

for (x, y, z) = (1, -3, 4) and for parameters k = 1.5 and a = -1

Note the "param" range G2:H3: the parameters must be always passed with theirs labels "k"
and "a". On the contrary the variables do not need labels but must always be passed in the
order x, y, z.
Using the trick of the scaling for 243, the inverse of the determinant, we get an integer matrix,
that represent the numerator of the Hessian itself. Then, simplifying each element for 243, we
obtain the Hessian matrix in the following exact fractional form

242

Xnumbers Tutorial

Non-linear equation solving with derivatives

Derivatives play a strategic role in solving non-linear equation and non-linear system
The most efficient algorithms use the derivatives information in order to speed up the
convergence to the final solution. From the point of view of numeric calculus, derivatives are
rarely used because they tend to magnify the truncation error. This is general true and a many
rootfinding algorithm avoid the derivatives. In solving non-linear problem, however, the
derivatives can be very useful because they can greatly improve the convergence without
influence the final result accuracy, that depends only by the evaluation function f(x)
Let's see and example.
Solve the following equation

x 2 = sin x with an accuracy better then 1e-25.

First of all we build the function

f ( x) = x 2 sin x
and draw its plot. The point x where f ( x) = 0 is
the solution of the given equation We see that the
zero exists and it is near the point 1.
We note also that in the interval [0.5 1.5] the
function is monotonic
In this interval the Newton-Raphson iterative algorithm, starting from x = 1.5, should work fine.

xn +1 = xn

f ( xn )
f ' ( xn )

n = 0, 1, 2 ...

To implement this algorithm we need the evaluation function f(x) with about 30 significant digits.
For that, it comes in handy the multiprecision function xeval. For the derivative we have to way:
computing the function f'(x) by hand and evaluating it by xeval or approximating the derivative
by the function diff1 in standard precision. Because we are a bit lazy and the derivatives is not
so immediate, we chose the second way. A simple spreadsheet arrangement may be the
following

As we can see the convergence is superb!. After few iteration the solution is
x 0.9496166887146629471509830317 with |f(x)| < 1e-28
This excellent result has been obtained in spite of the limited precision (1e-13) of the derivative.
The reason is simple: the accuracy of the derivative does not influence the final accuracy of the
root. We note that the derivative, after very few iterations, remains constant: we might
substitute this value with an even more approximated values, i.e. f' = 1.57, for all iterations. The
final accuracy will not change. We will need only more few steps, at the most.
243

Xnumbers Tutorial
But this method show its power overall for non-linear systems. For a 2 variables problem the
Newton-Raphson method becomes

f ( x, y ) = 0

g ( x, y ) = 0

fx
x
x
=
y n +1 y n +1 g x

fy
g y

f

g n

The (2 x 2) matrix is the Jacobian calculated at the point (xn, yn). In Xnumbers it can be
evaluated by the function Jacobian
Example. Solve the following system

x 5 + y 2 x 6 = 0
x y
e + x + y 1 = 0
setting:

f ( x, y ) = x 5 + y 2 x 6
g ( x, y ) = e x y + x + y 1
the contour plots of the functions f = 0 and g = 0 show two
intersection points: one near the point (-1 2) and another
near (-1, -2)
The function f(x,y) and g(x,y) are
evaluated and converted in double
precision by the nested functions
=xcdbl(xeval(B4,B7:B8))
=xcdbl(xeval(B5,B7:B8))
At the begin, insert the starting point
(-1, 2) in the cells B7:B8.
The new point is calculated in the
cells E7:E8. Copy this range and reinsert in the range B7:B8. At each
iteration the increments dx,dy of the
range E10:E11 becomes more and
more small.
Starting from (-1, 2) and (-1, -2) the iteration algorithm converges to the correct solutions
x

-1

-1

-2

-1.0201151219

1.9698273171

-0.7964138633

-2.3053792051

-1.0196483063

1.9693084022

-0.8079928505

-2.2042117521

-1.0196480758

1.9693081215

-0.8107932120

-2.1997452584

-1.0196480758

1.9693081215

-0.8108021826

-2.1997248438

244

Xnumbers Tutorial

Conversions
Decibel
=dBel(A, [MinLevel])
Converts a positive number A into decibel

AdB = 20 log10 ( A)
If zero, A is substituted with the value contained in the parameter "MinLevel" (default 1E-15)
Example
A

A dB

1
0.5
0.1
0.05
0.01
0.001
0.0001
0

0
-6.0206
-20
-26.021
-40
-60
-80
-300

Base conversion
cvDecBin(DecNum)

base 10 base 2

cvBinDec(BinNum)

base 2 base 10

cvDecBase(DecNum, Base)

base 10 any base (2-16)

cvBaseDec(BaseNum, Base)

any base (2-16) base 10

baseChange(number, old_base, new_base)

xBaseChange 2

any base (2 - 36) any base (2 - 36)

Any base (2 .. >1 million)

These functions perform the number conversion between different bases.

Example: Converts the decimal number n = 902023485 into bases 2 and 3.
cvDecBin(902023485) = 110101110000111100100100111101
cvDecBase(902023485, 3 )= 2022212022112121020

(base 2)

(base 3)

Example: Converts the hexadecimal number n = 35CFFF3D into decimal

cvBaseDec(35CFFF3D) = 902823741

(base 10)

You can also convert directly base-to-base, nesting two functions.

Example convert n = 35CFFF3D from base 16 into 8
cvDecBase(cvBaseDec(35CFFF3D,16),8) = 6563777475
For this scope you can also use the baseChange function
1
2

The function baseChange appears thanks to Richard Huxtable

The function xBaseChange appears thanks to Ton Jeursen

245

(base 8)

Xnumbers Tutorial
In spite of its digits limitation (15), this function has several interesting features
It converts any number into many different bases (up to 36). The digits greater then 9 are
indicated as A, B, C, D E, F, G, H, etc. It converts also decimal numbers. It formats the result
consistently with the source cell. Let's see how it works
The cell C4 is formatted with 7
digits and consequently its results
have the same format; the cell D4
is formatted with 2 decimals and its
result has the same format.

Multiprecision Base Conversion

= xBaseChange(xBC_OldNum, [xBC_NewBase], [deep])
This function1 converts a number from any base (2 .. >1 million)
xBC_OldNum is the number to convert in the following formats
3642455[7]

(integer number in base 7)

10C2AE[16] (integer number in base 16)

1098414

(integer number in base 10)

213.02[4]

(decimal number in base 4)

The oldnumber is assumed to be base 10; otherwise oldbase is provided between square
brackets [].
xBC_NewBase is the new base. If omitted it is assumed 10 as default.
If the base > 35 the letters are no more sufficients and the digits are represented as group of
digits separate with colon :
9:34:9:23[36]

(integer number in base 36)

6:54:122:42:65[200]

(integer number in base 200)

100:368.356:980[1000] (decimal number in base 1000)

deep is an optional parameter (0 to 50) setting the max number of decimals showed (default
50)
Example:
Results in:

xBaseChange("9:34:9:23[36]", 7)
3642455[7]

Example:
Results in:

xBaseChange(1098414 , 16)
10C2AE[16]

The function xBaseChange appears thanks to Ton Jeursen.

246

Xnumbers Tutorial
Example:
Results in:

xBaseChange(1098414 , 46)
11:13:4:26[46]

Example:
Results in:

xBaseChange("1F0C2AE[16]")
32555694[10]

Example
Results in

xBaseChange(10092.102 , 2048 , 6)
4:1900.208:1835:16:786:884:1507...[2048]

Note that, as in the last case, the digits groups can have different lenght: one , two, or more
digits
If we wan to force all the groups having the same number of digits simply insert the new base
with prefix "0"
Example
Results in

xBaseChange(10092.102 , "02048" , 6)
0004:1900.0208:1835:0016:0786:0884:1507...[02048]

Log Relative Error

= mjkLRE(q, c, NoSD)
= xLRE(q, c, NoSD, [DgtMax])
This function1 returns the log relative error (LRE) for an estimated value (q) and a certified
value (c), which has a specified number of significant digits (NoSD). The LRE is a measure of
the number of correct significant digits only when the estimated value is close to the exact
value. Therefore, each estimated quantity must be compared to its certified value to make sure
that they differ by a factor of less than two, otherwise the LRE for the estimated quantity is
zero.
Definition. The base-10 logarithm of the relative error is defined as:

if | q | > 1
LRE = 0
if c = 0
LRE = min( log(q ) , NoSD )

| q |< 1

LRE = min( log(| q c | / | c |) , NoSD)

if c 0 LRE = NoSD if c = q
LRE = 0
if else

if c q and 1/2 | q/c | 2

Example: Assume that you want to compare an approximate value with a 15 digits certified
value of . LRE metric can show this in a easy way
Certified value C = 3.14159265358979
Approx. value Q = 3.14159265300001
mjkLRE(C, Q, 15) = 9.7
This means that two values are close for about 10 significant digits. LRE metric rejects non
significant digits. Look at this example:
Certified value C = 0.000133333333333333
1

These functions appear by courtesy of Michael J. Kozluk. This algorithm was first programmed into an Excel user
function, by Michael, in standard 32 bit precision. As it works fine also for comparing long extended numbers (NoSD>
15), we have now developed its multiprecision version xLRE().

247

Xnumbers Tutorial
Approx. value Q = 0.000133333333333311
mjkLRE(C, Q, 15) = 12.8
The two numbers appear exact up to the 17th digit, but the relative error is about 1E-13
LRE is very useful when you work with long string of extended numbers.
For example, compare this approximation of "e" (Napiers number)
Certified value C = 2.71828182845904523536028747111
Approx. value Q = 2.71828182845904523536028747135
xLRE(C, Q, 30) = 28.1
At the first sight it is hard to say, but the LRE function shows immediately a precision of about
28 digits

248

Xnumbers Tutorial

Special Functions
The computation of special functions is a fundamental aspect of numerical analysis in virtually
all areas of engineering and the physical sciences.
All these special functions have a high-fixed-precision. Because most of these special functions
are in the form of infinite series or infinite integrals, their numeric solutions are quite
complicated, and we have spent many times for selecting and testing many different algorithms
in order to achieve the highest possible accuracy in 32 bit arithmetic.

Error Function Erf(x)

errfun(x)
Returns the error function

erf ( x) =

e t dt

Accuracy: about 1E-14 per x > 0

Exponential integral Ei(x)

exp_integr(x)
Returns the exponential integral

et
dt
t
x

Ei ( x) =
Accuracy: about 1E-14 for any x 0
This version1 accept also negative argument
For definition is

Ei ( x) = E1 ( x) x > 0

Exponential integral En(x)

expn_integr(x, n)
Returns the exponential integral of n-th order

e xt
En ( x) = n dt
t
1
Accuracy: about 1E-14 for x> 0 and n > 0

This extension appears thanks to the courtesy of Hans Gnter

249

Xnumbers Tutorial

Euler-Mascheroni Constant

xGm([Digit_Max])
Returns the Euler-Mascheroni gamma constant.
The optional parameter Digit_Max sets the maximum digits (default 30, max 415)

1 1
2 3

1
n

= lim 1 + + + ... log(n)

Example: compute the gamma constant with 40 significant digits
xGm(40) = 0.5772156649015328606084804798767149086546

Gamma function (x)

xGamma(x, [Prec])
Returns the gamma function with precision up to 30 digits.
Definition:

( x) = t x 1 e t dt
0

Optional parameter Prec sets the maximun precision from 15 (default) to 30 digits
When Prec = 15 the routine works in faster double arithmetic.
This routine uses the excellent Lanczos series approximation1

2 x + g + 12
( x) g

e
e

x + 12

N
c

c0 + i
i =1 x + i

where: g = 607/128, ci are the Lanczos' coefficients and 14 < N < 22.
Relative accuracy in double precision is better than 10^-14, (except very near to the poles x =
0, -1. -2, -3, etc.). In extended precision the accuracy is better then 1E-30
xGamma(5.1)
= 27.9317537383684
xGamma(5.1, 30) = 27.9317537383683833586731052773
This function works also with large argument because it uses the multiprecision format to avoid
the overflow for arguments greater than 170.
Example,
x
0.0001
0.001
0.01
0.1
1
10
100
1

xgamma(x)

Rel. Error

9.99942288323162E+3
9.99423772484596E+2
9.94325851191507E+1
9.51350769866874
1
3.6288E+5
9.33262154439441E+155

1.00E-15
1.02E-15
1.00E-15
9.33E-16
0
0
5.64E-16

This high accuracy algorithm has been extracted form a very good note by Paul Godfrey, Intersil , C.2001

250

Xnumbers Tutorial
1,000
10,000
100,000
1,000,000

4.02387260077093E+2564
2.84625968091705E+35655
2.82422940796034E+456568
8.26393168833122E+5565702

1.92E-15
1.58E-15
2.75E-15
2.54E-15

Note that relative accuracy is better than 5E-15 in any case.

You can convert in double only the values with x 170, otherwise you will get #VALUE! (error).
You can manipulate these large values only by the "x-functions", or, separating mantissa and
exponent (see xsplit())
FACTORIAL: Thanks to its efficence and accuracy, this function can also be used to calculate
the factorial of a big integer number, using the relation
n! = (n+1)
Example:
xfact(10002) =

2.84711361574652325360317551421E+35667

30 digits, slower

xgamma(10003) =

2.84711361574651E+35667

15 digits, faster

Log Gamma function

xGammaln(x)
xGammalog(x)
These function return the natural and decimal logarithm of the gamma function.
xgammaln(100000)

1051287.7089736568948

xgammalog(100000) =

456568.45089997090835

Relative accuracy is better than 10^-(14+|log(x)|)

for x>0
These functions are added only for compatibility with Excel and other math packages. They are
useful to avoid overflow in standard precision arithmetic for large arguments of gamma
function. However if you use directly the xgamma() and multiprecision arithmetic, you need no
more to use these functions.

Gamma quotient
xGammaQ(x1, x2)
Performs the division of two gamma functions.
q = (x1) / (x2)
Relative accuracy is better than 1E-14, for x1 > 0 and x2 > 0
Example: suppose you have to calculate for v =1,000,000 the following quotient

q=
Taking

( v 2+1)
( 2v )

x1 = 500,000.5 and x2 = 500,000 , we have easily

xgammaq(500000.5 , 500000) = 707.106604409874

(rel error = 5.96E-16 )

Note that if you have used the standard GAMMALN() function, you should have:

251

Xnumbers Tutorial
EXP(GAMMALEN(500000.5) - GAMMALEN(500000)) = 707.106604681849
(with a rel. error = 3.846E-10)
As we can see, In this case, the error is more than 500,000 times bigger that the previous one!

Gamma F-factor
xGammaF(x1, x2)
Returns the gamma factor of the Fischer distribution.

x +x
1 2
2
k=
x x
1 2
2 2
Relative accuracy is better than 1E-14, for x1 > 0 and x2 > 0

Digamma function
digamma(x)
Returns the logarithmic derivative of the gamma function

( x) =

d
' ( x)
ln(( x) ) =
dx
( x)

Relative accuracy is better than 1E-14, for x > 0

Example
digamma(x)
0.01
0.1
1
10
100
1000

value

rel. error

-100.560885457869
-10.4237549404111
-0.577215664901532
2.25175258906672
4.60016185273809
6.90725519564881

3.24E-15
2.23E-15
1.49E-15
4.92E-16
5.65E-16
2.97E-16

Note that (1) = (Eulero- constant)

Beta function
xBeta(a, b)
Returns the beta function
1

B(a, b) = t a 1 (1 t )b 1 dt
0

Relative accuracy is better than 1E-14, for a > 0 and b > 0

252

Xnumbers Tutorial

Incomplete Gamma function

xGammaI(a, x, [sel])
Returns one of the following Incomplete Gamma functions specified by the "sel" parameter
(default 1)
Sel =1

Sel =2

Sel = 3

(a, x) = t a 1e t dt

(a, x) = t a 1 e t dt

P ( a, x ) =

Sel = 4 (Tricomi)

( a, x )

* ( a, x ) =

(a )

( a, x )
x a (a )

Relative accuracy is better than 1E-14 for : a > 0, x 0

Incomplete Beta function

xBetaI(x, a, b, [sel])
Returns one of the following Incomplete Beta functions specified by the "sel" parameter
(default 1)
Sel = 1

Sel = 2

Bx (a, b) = t a 1 (1 t )b 1 dt

I x ( a , b) =

Bx ( a , b )
B ( a, b )

where B(a, b) is the Beta function

Relative accuracy is better than 1E-14 for a > 0, b > 0. 0 x 1 .

Combinations function
xcomb_big(n, k)
Returns the combination, or binomial coefficients, for large integer numbers

n
n!
Cn, k = =
k
(
n

k )!n!

Relative accuracy is better than 1E-14, for n >> 0 and k >> 0
This function uses the gamma function to calculate the factorials. It is much faster than xcomb
function. For this reason is adapted for large integer values (10,000 - 1,000,000)
xcomb(5000,2493) =

1.5627920156854189438574778889E+1503

(30 digits, slow)

xcomb_big(5000,2493) =

1.56279201568542E+1503

(15 digits , fast)

253

Xnumbers Tutorial

Bessel functions of integer order

BesselJ (x, [n])

Bessel function of 1 kind, order n: Jn(x)

BesselY (x, [n])

Bessel function of 2kind, order n: Yn(x)

BesseldJ (x, [n])

First derivative of Bessel functions of 1 kind, order n: J'n(x)

BesseldY (x, [n])

First derivative of Bessel functions of 2 kind, order n: Y'n(x)

BesselI (x, [n])

Modified Bessel function of 1 kind, order n: In(x)

BesselK (x, [n])

Modified Bessel function of 2kind, order n: Kn(x)

BesseldI (x, [n])

First derivative of mod. Bessel functions of 1 kind, order n: I'n(x)

BesseldK (x, [n])

First derivative of mod. Bessel functions of 2 kind, order n: K'n(x)

Relative accuracy is better than 1E-13, for x > 0 and n any integer
These routines1 have a high general accuracy. Look at the following example. We have
compared results obtained from our BesselJ with the standard Excel similar function
x
0.1
0.5
1
5
10
50

J0(x) (BesselJ)

Rel. Error

0.997501562066040
0.938469807240813
0.765197686557967
-0.177596771314338
-0.245935764451374
0.055812327669252

1.11E-16
1.06E-15
7.25E-16
2.66E-15
1.06E-13
3.98E-15

J0(x) (Excel standard)

0.997501564770017
0.938469807423541
0.765197683754859
-0.177596774112343
-0.245935764384446
0.055812327598901

Rel. Error
2.71E-09
1.95E-10
3.66E-09
1.58E-08
2.72E-10
1.26E-09

As we can se, the general accuracy is better than 200,000 times!

Cosine Integral Ci(x)

CosIntegral(x)
Returns the Cosine integral defined as:

ci( x ) =
x

cos(t )
dt
t

Relative accuracy is better than 1E-13, for x > 0

Sine Integral Si(x)

SinIntegral(x)
Returns the sine integral defined as:
x

si( x) =
0

sin(t )
dt
t

Relative accuracy is better than 1E-13, for x > 0

1

All these special functions are provided thanks to the FORTRAN 77 Routines Library for Computation of Special
Functions developed by Shanjie Zhang and Jianming Jin . The programs and subroutines contained in this library are
copyrighted. However, authors kindly gave permission to the user to incorporate any of these routines into his
programs.

254

Xnumbers Tutorial

Fresnel sine Integral

Fresnel_sin(x)
Returns the Fresnel's sine integral defined as:
x

S ( x ) = sin ( 12 t 2 ) dt
0

Relative accuracy is better than 1E-13, for x > 0

Remember also the following relation
x

k sin (t 2 ) dt = S (k z )

where:

k=

Fresnel cosine Integral

Fresnel_cos(x)
Returns the Fresnel's cosine integral defined as:
x

C ( x) = cos( 12 t 2 ) dt
0

Relative accuracy is better than 1E-13, for x > 0

Remember also the following relation
x

k cos(t 2 )dt = C (k z )

where:

k=

Fibonacci numbers
xFib(n, [DgtMax])
Returns the Fibonacci's numbers defined by the following recurrent formula:

F1 = 1 , F2 = 2 ,

Fn = Fn 1 + Fn 2

Example:
xFib(136)

= 11825896447871834976429068427

xFib(4000) = 3.99094734350044227920812480949E+835

255

Xnumbers Tutorial

Hypergeometric function
Hypgeom(a, b, c, x)
Returns the Hypergeometric function
The parameter "a" is real, "b" is real, "c" is real and different form 0, -1, -2, -3 ...
The variable "x" is real with |x| < 1
Relative accuracy is better than 1E-14, for -1 < x < 1
The hypergeometric function is the solution of the so called Gaussian-hypergeometric
differential equation

x(1 x ) y + (c (a + b + 1)x ) y + ab y = 0
An integral form of the hypergeometric function is
1

F (a, b, c.x) =

(c)
c b 1
(1 t x )a dt
t b1 (1 t )

(b)(c b) 0

More known is the series expansion that converges for !x| < 1

F (a, b, c.x) = 1 +

ab x a(a + 1)b(b + 1) x 2 a(a + 1)(a + 2)b(b + 1)(b + 2) x 3

+
+
+ ...
c 1!
c(c + 1)
2!
c(c + 1)(c + 2)
6!

Special result are:

F ( p,1,1, x) = (1 x) p
ln(1 + x)
x

F (1,1,2, x) =

F ( 13 , 23 , 56 , 27
32 ) = 1.6

Zeta function (s)

Zeta(s)
The Riemann zeta function (s) is an important special function of mathematics and physics
which is intimately related with results surrounding the prime number, series, integrals, etc.
Relative accuracy is better than 1E-14, for any s 1
It uses the fast Borwein formula.
Definition. For |s|>1 the function is defined as:

(s) =
k =1

1
ks

Analytic continuation. It can be defined for 0< s <1 by the following analytic continuation:

1
(s) =
1 21 s

k =1

(1) k 1
ks

For s< 0 the function is defined by the following relation:

(1 s ) = 2(2 ) s cos( 12 s ) ( s) ( s)

256

Xnumbers Tutorial
Same known exact results are: (2) = (2) /6 , (4) = (4)/90
Zeta function is very useful in computing series. Look at this example:

k =0

1
1
1
1
5
=
=
1 2 = ( 2)

2
2
2
( k + 2)
2
4
k =2 k
k =1 k

So, the final result is (2)/6 - 5/4

Airy functions
= AiryA(x)
= AiryB(x)
= AiryAD(x)
= AiryBD(x)
Return the Airy functions A(x), B(x) and theirs derivatives A'(x), B'(x)

A(x) =

B(x) =

1
32 / 3
1
31 / 6

n=0

n=0

(13 (n + 1)) 1 / 3 n 2(n + 1)

3 x sin

n!
3

( 13 (n + 1)) 1 / 3 n 2( n + 1)
3 x sin

n!
3

Variable x is real. Relative accuracy is better than 1E-14

Elliptic Integrals
= IElliptic1(, k)
= IElliptic2(, k)
Return the Jacobian Elliptic Integrals of 1st and 2nd kind

IElliptic1 =

F ( , k ) =
0

d
1 k 2 sin 2

IElliptic2 =

E ( , k ) = 1 k 2 sin 2 d
0

Variable is real, 0 < k < 1. Relative accuracy is better than 1E-14

257

Xnumbers Tutorial

Kummer confluent hypergeometric functions

= Kummer1(a, b, z)
= Kummer2(a, b, z)
They return the Kummer conflent hypergeometric functions of 1st and 2nd kind, also called
M(a, b, z), and U(a, b, z) respectively.
The confluent hypergeometric functions give the complete solutions of the confluent
hypergeometric differential equation.

d 2w
dw
+ (b a )
aw = 0
2
dz
dz

having the complete solution

w = q1M (a, b, z ) + q2U (a, b, z )

The parameters "a" and "b" can be positive or negative but always different from 0, -1, -2, -3 ...
The variable z must be positive for Kummer2
For a > 0 and b > 0, the confluent hypergeometrix functions have an integral representation
(Abramowitz and Stegun 1972, p. 505)
1

(b)
M ( a , b, z ) =
e zt t a 1 (1 t ) b a 1 dt

(b a )(b) 0

1
U ( a , b, z ) =
e zt t a 1 (1 + t ) b a 1 dt
(a ) 0
a

0.5

0.2

M(a,b,z)

0.5

0.2

601.287114305

0.40255172241

0.5

0.2

10

112016.644576

0.29849708947

0.5

0.2

20

3063046479.13

0.21689337800

0.5

0.2

30

7.6396125E+13

0.17882728338

8.2

1.4

41.1797940694

1.39280467E-06

-0.9

0.4

10

-345.535973205

7.72952303674

5.85803416181

U(a,b,z)
0.70928623024

258

Xnumbers Tutorial

Integral of sine-cosine power

= IntPowSin(n, x)
= IntPowCos(n, x)
These functions return the integral of sine and cosine raised to the n integer power
x

CPI ( x, n) = (cos t ) n dt

SPI ( x, n) = (sin t ) n dt

Variable x 0 is real, n = 0, 1, 2, 3 ... is integer. Relative accuracy is better than 1E-15

2

1.4
1.2

CPI(x)

SPI(x)

1.5

n=8

1
0.5

0.8
0.6

0.4

-0.5

0.2

-1

-1.5

-0.2

0.5

1.5

2.5

3.5

n= 8
CPI(x)
SPI(x)

-2
-6

-4

-2

These functions use the fast recursive algorithms

cos x(sin x) n 1 n 1
+
SPI ( x, n 2) ,
n
n

SPI ( x, 1) = 1 cos x , SPI ( x, 0) = x

sin x(cos x ) n 1 n 1
+
CPI ( x, n 2) ,
n
n

CPI ( x, 1) = sin x , CPI ( x, 0) = x

SPI ( x, n) =

CPI ( x, n) =

The recursive formula of CPI(x, n) is very stable and accurate for all values x 0 , n 0, n N
On the contrary, the recursive formula of SPI(x, n) degrades precision for small value of x
In that case, the function automatically switches on the power expansion series
x

(sin t )
0

ak
1

dt = y n +1
y 2k
+
n + 1 k =1 2k + n + 1

where
k
2i 1 1 3 5...( 2k 1)
ak =
=
2 4 6...( 2k )
i =1 2i

259

y = sin x

x <1

Xnumbers Tutorial

Spherical Bessel functions of integer order

= BesselSphJ(x, [n])
= BesselSphY(x, [n])
These functions return the Spherical Bessel functions of 1st and 2nd kind of integer order n = 0,
1, 2... (default n = 0), and for any real x > 0, definited as:

jn ( x ) =

yn ( x) =

2x

2x

J n + 1 ( x)

Spherical Bessel functions of 1st kind

Yn + 1 ( x)

Spherical Bessel functions of 2nd kind (also Neuman function)

The Spherical Bessel functions are the solutions of the Helmholtz radial equation in spherical
coordinates

x2

d2y
dy
+ 2 x + ( x 2 n 2 n) y = 0
2
dx
dx

yn(x)
n
x
jn(x)
-------------------------------------------------------------------------------0
1
2

0.5
1
3

0.958851077208406
0.301168678939757
0.298637497075734

-1.755165123780750
-1.381773290676040
-0.267038335264499

260

Xnumbers Tutorial

Formulas Evaluation
Multiprecision Expression Evaluation
These functions realize a little math shell, putting together the power of multiprecision numeric
computation with the ease of symbolic calculus. Sometime we may want to perform the
computation using symbolic formulas. We would pass these strings to a routine for evaluation,
returning the numerical results with a given accuracy. These functions perform this useful task.

xeval( Formula, [Var], [DgtMax], [Angle],)

xevall( Formula, [Var1, Var2 ] )
These functions return the evaluation of a math expression in multiprecision arithmetic. They
use the same algorithm1 and have the same variable accuracy. They differ only for the input
parameters.
The parameter "Formula" is a math expression string containing variables, operators,
parenthesis and other basic functions. Examples.
3+1/(x^2+y^2), sin(2*pi*t)+4*cos(2*pi*t), (x^4+2x^3+6x^2-12x-10)^(1/2)

The optional parameter "Var" is an array containing one or more value for variables
substitution. Before computing, the parser substitutes each symbolic variable with its
correspondent value. It can be a single value, an array of values or, even an array of values +
labels (see examples).
The optional parameter "Var1", "Var2"... are single values or array as "Var" but without labels,
because the function xevall automatic finds by itself the appropriate labels. (See example)
The optional parameter "DgtMax" from 1 to 200 - sets the maximum number of precision
digits (default=30). Setting DgtMax = 0 will force the function to evaluate in faster standard
precision.
The optional parameter "Angle" sets the angle unit "RAD (default) DEG, GRAD.of for
trigonometric computation:
Example:
xeval("(1+sqr(2))/2+5^(1/3))") = 2.91708272786324451375395323463
xeval("(1+cos(x))/2+x^y" , {5, 1.2}) = 7.5404794000376872941836369067
xeval("(a+b)*(a-b)", {2, 3})= (2+3)*(2-3) = -5

All the function parameters can also be passed by cell references, like A1, $B$2, etc.
Example. Tabulate the following function for x = 1, 1.5, 2, with 30 significant digits

f ( x) =

1+ x
1+ x2

The algorithm is divided into two steps: parsing and evaluation. The first step is performed by the MathParser class.
The evaluation is performed with the x-functions of XNUMBERS.

261

Xnumbers Tutorial

Note how the use of this function is simple and straight comparing with the correspondent
nested formulas
=xdiv(xadd(1,A6),xsqr(xadd(1,xpow(A6,2))))
Calculating functions with more than one variable a bit complication arises, because we have to
pay attention which values are assigned to the variables. Lets see this example
Calculate the following bivariate function for x = 2.4, y = 5.5

f ( x, y ) =

ln( y ) + xy
1+ x2

In order to pass to the parameter "Var" the correct value for each variable we select the
variables range B2:C3 including the labels "x" and "y" (header). The labels must contain the
same symbols contained into the formula string

Note. If we pass the range B3:C3 without labels, the function assigns the values to the
variables in the same order that they appear in the formula, from left to right. In our example
the first variables is y and the second is x, so the function assignes the first value 2.4 to y
and the second value 5.5 to x
To by-pass the variable order rule, the function uses the trick of the variables labels. On the
contrary, for one or none variable it is impossible to make confusion so the header can be
omitted.
Variables order. The function returns the variables order in the Excel function insertion panel

In our example we see the string y x , that means you have to pass the first value for the
variable y and the second value for the variables x
262

Xnumbers Tutorial
The variables order is by-passed by labels. Using labels you do not need to worry about the
variables order
Lets see another example. Calculat,e with 30 digits precision, the following function

f ( x, a, b) = a sin( x) b cos(2 x)
for x = 30 deg , a = 1 , b = -2

Note that we have set the angle unit = DEG

Note also that in this case the variable order would be a, x, b, but with the aid of the labels the
function can associate the exact parameters in the right way.
Sometime it is not possible to add a label near its value (in the middle of a table, for example).
Neither all parameters are located adjacent each others. For these cases we can use the
second evaluator function xevall
This function accepts separate parameters or separate array. We do not need to pass variables
labels. The function automatically finds all labels present on the active worksheet
Of course all this has a cost. the function xevall is about 10 times slower then xeval.
Lets see how it works with an example. Tabulate the given trigonometric function, from t = 0 to
0.5, with step = 0.1 and an error less then 1E-20

f (t ) = a cos( t ) + b cos( t )
where a = 0.5 and b = -2

The above sheet shows a possible arrangement. If we look the last cell B8 we discover that the
parameters are:
Var1 the cell A8 containing the value of the independent variable t
Var2 the range C4:D4, containing the values of the parameters a and b
Var3 range C7:D7, containing the internal parameter DgtMax and AngleSet
The internal parameters DgtMax and AngleSet are reserved word and must write as is.
Note also that the cell A8 has no label, but the function performs the correct assignment to the
t variable.

263

Xnumbers Tutorial
Label Rules. Labels must stay always at the
top or at the left of the corresponding values.
Labels can have any alphanumeric name
starting with any letter and not containing any
blank.
In the example:
t = 0.1, a = 0.5 , DgtMax = 30
The function xevall only assigns the cell value to the variable on top to the correspondent
column. Othewise it assigns the value to the variable on the left to the correspondent row.

264

Xnumbers Tutorial

Complex Expression Evaluation

=cplxeval( Formula, [Var1, Var2 ] )
This function1 evaluates a math expression in complex arithmetic.
The parameter "Formula" is a math expression string containing variables, operators,
parenthesis and other basic functions.
(3+8j)*(-1-4j) ,

(1+i)*ln(1+3i) ,

((x+3i)/(x+4-2i))^(1-i)

The optional parameter "Var1", "Var2",... can be single or complex value. See How to insert a
complex number for better details
Example: Evaluate the given complex polynomial for z = 2 i

z 2 + (3 + i ) z + (2 + 5i )
Note that we use the complex rectangular format, i.e. "3+2j", only in the symbolic math
formula.
When we pass a complex value to a variable
we must always use the double cell format.
Note also that we can write i or j as well for
imaginary symbol, the parser can recognize
both of them.

For complex numbers labels are not supported. When we have formulas with two or more
variables, we must provide the values for variable substitutions in the exact order that they
appear in the formula, starting from left to right. The formula wizard will easily help you. Look at
this example.
Example. Compute the expression for the given complexes values

e ks
F ( s) =
( s a )( s b)
In the cell B2 we have inserted the string

s = 1 + j , a = 1 4j , b = 3 + 6j ,
k = 0.5
exp(k*s)/((s-a)(s-b)

When we enter the formula, the parser

recognizes the variables symbols and shows
us the exact order in which we have to pass to
the function itself.
In this case: k, s, a, b

This function uses the clsMathParserC class by A. De Grammont and L. Volpi

265

Xnumbers Tutorial

Multiprecision Excel Formula Evaluation

= xCalc(Formula, [DgtMax])
This function1 is useful for performing multiprecision calculation directly in a worksheet
Formulas can be entered just like in Excel and it may also contain references to other cells.
The optional parameter "DgtMax" from 1 to 200 sets the maximum number of precision
digits (default=30).
This smart functions has several advantages: it is easy to use and can handle long formula; the
formula is not changed at all and can keep the references to other cells. The only drawback
remains its speed because the function have to parse the formula and evaluate the result using
the multiprecision arithmetic; therefore the evaluation takes about 5-6 times slower then basic
x-functions.
Of course, not all the Excel functions can be converted. The functions recognized are.
+ - * / ^
ABS ACOS ACOSH ASIN ASINH ATAN
COMBIN COS COSH EXP FACT INT LN LOG MOD PI
ROUND SIGN SIN SINH SQRT TAN TANH

ATANH

Nota. The functions are indicated with their original English names, that usually are different from the local names.

Let's see how it works. Assume to have in a cell the following Excel formula
=(A2*EXP(B2)+A4*EXP(B4))/(A2^2+A4^2)

Note that the formula references the values of other 4 cells: A2, B2, A4, B4
The formula value, calculated in standard precision, is 0.0140820785524786
We can get the value in 30 significant digits applying the xCalc function
=xCalc( (A2*EXP(B2)+A4*EXP(B4))/(A2^2+A4^2) )
For that, simply, select the cell D2 and move the mouse to the input field; insert xCalc( ... ) just
around the original formula.

The result in the cell D4 has now a precision of 30 significat digits

Observe that the the external cells links remain active and the formula is unchanged as in a
standard Excel formula

The function xCalc appears thanks to Ton Jeursen.

266

Xnumbers Tutorial
Note that, using the x-function, the above formula should be translated as
=xdiv(xadd(xmult(A2,xexp(B2)),xmult(A4,xexp(B4))),xadd(xpow(A2,"2"),xpow(A4,"2")))
The result is equivalent but it is evident the great advantage to keep the original formula.

267

Xnumbers Tutorial

Math expression strings

Functions like Integr, Series, xeval, xevall, cplxeval operate with symbolic math expressions
by the aid of clsMathParser and claMathparserC (two internal class modules).
These programs (for real and complex numbers) accept in input any string representing an
arithmetic or algebraic expression with a list of variable values and return a multiprecision
numeric result. Typical math expressions are:
1+(2-5)*3+8/(5+3)^2

sqr(2)+asin(x)

(a+b)*(a-b)

x^2+3*x+1

1.5*exp(-t/12)*cos(pi*t + pi/4)

(1+(2-5)*3+8/(5+3)^2)/sqr(5^2+3^2)

2+3x+2x^2

0.25x + 3.5y + 1

sqr(4^2+3^2)

1/(1+e# ) + Root(x,6)

(-1)^(2n+1)*x^n/n!

|x-2|+|x-5|

And((x<2),(x<=5))

sin(2*pi*x)+cos(2*pi*x)

Variables can be any alphanumeric string and must start with a letter
x, y, a1, a2, time, alpha , beta
Also the symbol "_" is accepted to build variable names in "programming style".
time_1, alpha_b1 , rise_time
Capitals are accepted but ignored. Names such as "Alpha", "alpha", "ALPHA" indicate the
same variable.
Implicit multiplication is not supported because of its intrinsical ambiguity. So "xy" stands for
variable named "xy" and not for x*y. The multiplication symbol "*" generally cannot be omitted.
It can be omitted only for coefficients of the classic math variables x, y, z. It means that string
like 2x and 2*x are equivalent

2x, 3.141y, 338z^2

2*x, 3.141*y, 338*z^2

On the contrary, the following expressions are illegal in this context.

2a, 3(x+1), 334omega
Constant numbers can be integer, decimal, or exponential
2

-3234

1.3333

-0.00025

1.2345E-12

Note: This version support also the comma "," as decimal separator, depending on your system
internation option setting.
1,3333

-0,00025

1,2345E-12

Logical expressions are now supported

x<1

x+2y >= 4

x^2+5x-1>0

t<>0

(0<x<1)

Logical expressions return always 1 (True) or 0 (False). Compact expressions, like 0<x<1 ,
are now supported; you can enter: (0<x<1) as well (0<x)*(x<1)
Numerical range can be inserted using logical symbols and Boolean functions. For example:
For

2<x<5

insert

(2<x)*(x<5)

For x<2 , x>=10

insert

OR(x<2, x>=10)

For -1<x<1

insert

(x>-1)*(x<1)

268

or also (2<x<5)
or also (x<2)+(x>=10)

, or (-1<x<1) , or also

|x|<1

Xnumbers Tutorial
Piecewise Functions. Logical expressions can also be useful for defining a piecewise
function, such as:

f(x) =

2x-1-ln(2)

ln(x)

x/2-1+ln(2)

x 0.5
0.5< x < 2
x2

The above function can be written as:

f(x) = (x<=0.5)*(2*x-1-ln(2))+(0.5<x<2)*ln(x)+(x>=2)*(x/2-1+ln(2))

The parser adopts a new algorithm for evaluating math expressions depending on logical
expressions, which are evaluated only if the logical conditions are true (Conditioned-Branch
algorithm). Thus, the above piecewise expression can be evaluated for any real value x without
any domain error. Note that without this features the formula could be evaluated only for x>0.
Another way to compute piecewise functions is splitting it into several formulas (see example 6)
Math Constants supported are: Pi Greek (), Euler-Napier (e)
pi = 3.14159265358979

or

pi# = 3.14159265358979

e# = 2.71828182845905
Angle expression
This version supports angles in RAD radians, DEG degree, or GRAD degree.
For example if you set the unit "DEG", all angles will be read and converted into degrees
sin(120) => 0.86602540378444
asin(0.86602540378444) => 120
rad(pi/2) => 90
, grad(400) => 360

, deg(360) => 360

Angles can also be write in DMS format like for example 29d 59m 60s
sin(29d 59m 60s)

=>

0.5

29d 59s 60m

=>

30

Note This format is only for sexagesimal degree. It is independent from the unit set

Complex number can be indicated in a formula string as an ordered couple of number

enclosed into parenthesis (..) and divided by a comma , like for example:
(2, 3)

(a, b)

(-1, -0.05)

(-1.4142135623731,

-9.94665E-18)

On the other hand, complex numbers can also be indicate by the common rectangular form:
3+3j

a+bj

-1 0.05j

-1.4142135623731 9.94665E-18j

You note that the second form is suitable for integer numbers, while, on the contrary, for
decimal or exponential number the first one is clearer. The parenthesis form is more suitable
also in nested results like
((2+3*4), (8-1/2)) that gives the complex number (14, 7.5)
Note: Pay attention if you want to use the rectangular convention in nested formulas.
wrong (2+3*4)+(8-1/2)j.

correct (2+3*4)+(8-1/2)*j

Do not omit the product symbol * before j because the parser recognize it as an expression,
not a complex number. The product symbol can be omitted only when before the letter j is a
constant number
Note: You can use both j and i for indicating the imaginary number

Functions
Functions are called by their function-name followed by parentheses. Arguments can be:
numbers, variables, expressions, or even other functions.
sin(x)
max(a,b)

log(x)
root(x,y)

cos(2*pi*t+phi)
comb(n,k)

269

atan(4*sin(x))
beta(x,y)

Xnumbers Tutorial
For functions having more than one argument, the successive arguments are separated by
commas "," or alternatively, by semicolon ";" if your system has comma as decimal separator.
max(a;b)

root(x;y)

comb(n;k)

beta(x;y)

(only if comma "," is decimal sep.)

max(a,b)

root(x,y)

comb(n,k)

beta(x,y)

(only if point "." is decimal sep.)

270

Xnumbers Tutorial

List of basic functions and operators

The following table lists all functions and operators recongize in a math expression string
Use: R = Real, C = Complex, M = Multiprecision
The functions with M can be evaluated either in standard or multiprecision
The functions with C alone must be used with cplxeval only
The functions with R alone must be used with xeval, xevall setting DgtMax = 0 (standard
precision)
Function

Use

Description

RCM

subtraction

Note

RCM

factorial

5!=120

RM

percentage

35% = 0.35 , 100+35% =103.5

RCM

multiplication

RCM

division

35/4 = 8.75

RCM

integer division

35\4 = 8

RCM

raise to power

3^1.8 = 7.22467405584208

||

RCM

absolute value

|-5|=5

RCM

addition

<

RM

less than

return 1 (true) 0 (false)

<=

RM

equal or less than

returns 1 (true) 0 (false)

<>

RM

not equal

returns 1 (true) 0 (false)

RM

equal

returns 1 (true) 0 (false)

>

RM

greater than

returns 1 (true) 0 (false)

>=

RM

equal or greater than

returns 1 (true) 0 (false)

abs(x)

RCM

absolute value

abs(-5)= 5

acos(x)

RCM

inverse cosine

argument -1 x 1

acosh(x)

RCM

inverse hyperbolic cosine

argument x 1

acot(x)

RM

inverse cotangent

acoth(x)

RM

inverse hyperbolic cotangent

acsc(x)

RM

inverse cosecant

acsch(x)

RM

inverse hyperbolic cosecant

alog(z)

complex exponential with base 10

10^z

and(a,b)

logic and

returns 0 (false) if a=0 or b=0

arg(z)

polar angle of complex number

-pi < a <= pi

asec(x)

RM

inverse secant

asech(x)

RM

inverse hyperbolic secant

argument 0 x 1

asin(x)

RCM

inverse sine

argument -1 x 1

asinh(x)

RCM

inverse hyperbolic sine

atanh(x)

RCM

inverse hyperbolic tangent

atn(x), atan(x)

RCM

inverse tangent

beta(x,y)

RC

beta

betaI(x,a,b)

Beta Incomplete function

x >0 , a >0 , b >0

cbr(x)

cube root

clip(x,a,b)

Clipping function

comb(n,k)

RCM

combinations

cbr(2) =1.25992104989487, also 2^(1/3)

return a if x<a , return b if x>b, otherwise
return x.
comb(6,3) = 20

conj(x)

conjugate

cos(x)

RCM

cosine

cosh(x)

RCM

hyperbolic cosine

cot(x)

RM

cotangent

coth(x)

RM

hyperbolic cotangent

csc(x)

RM

cosecant

csch(x)

RM

hyperbolic cosecant

dec(x)

RM

decimal part

dec(-3.8) = -0.8

deg(x)

RM

degree sess. conversion

conv. degree (60) into current unit of

(the same as fact)

(the same as abs)

argument x<-1 or x>1

argument -1 < x < 1

argument x>0 y>0

argument in radians

271

argument (in radians) x k* with k = 0,

1, 2
argument x>2
argument (in radians) x k* with k = 0,
1, 2
argument x>0

Xnumbers Tutorial
angle
digamma(x)

RC

digamma

argument x>0

Ei(x)

RC

exponential integral function

argument x>0

Ein(x,n)

Exponential integral of n order

x >0 , n = 1, 2, 3

Elli1(x)

Elliptic integral of 1st kind

, 0<k<1

Elli2(x)

Elliptic integral of 2st kind

, 0<k<1

erf(x)

RC

argument x>0

erfc(x)

RC

eu

RM

error Gauss's function

error Gauss's function for complex
value
Euler-Mascheroni constant

exp(x)

RCM

exponential

exp(1) = 2.71828182845905

fact(x)

RM

factorial

argument x >0

fix(x)

RM

integer part

fix(-3.8) = 3

FresnelC(x)

Fresnel's cosine integral

FresnelS(x)

Fresnel's sine integral

gamma(x)

RC

gamma

argument x>0

gammaln(x)

RC

logarithm gamma

argument x>0

gcd(a,b)

greatest common divisor

grad(x)

degree cent. conversion

HypGeom(x,a,b,c)

Hypergeometric function

The same as mcd

conv. degree (100) into current unit of
angle
-1 < x <1 a,b >0 c 0, 1, 2

im(z)

immaginary part of complex number

int(x)

integer part

int(-3.8) = 4

integral(f,z,a,b)

Def. integral of complex function f(z)

Integral('1/z^2','z',1-i,1+i)

inv(x)

inverse of a number

1/x

lcm(a,b)

RM

lowest common multiple

The same as mcm

ln(x), log(x)

RCM

logarithm natural

argument x>0

max(a,b,...)

RM

maximum

max(-3.6, 0.5, 0.7, 1.2, 0.99) = 1.2

min(a,b,...)

RM

minimum

min(13.5, 24, 1.6, 25.3) = 1.6

mcd(a,b,...)

RM

maximun common divisor

mcm(a,b,...)

RM

minimun common multiple

mod(a,b)

RM

modulus

mcm(4346,174) = 2
mcm(1440,378,1560,72,1650) =
21621600
mod(29, 6) = 5 mod(-29, 6) = 1

nand(a,b)

logic nand

returns 1 (true) if a=1 or b=1

neg(z)

opposit of complex

-z

nor(a,b)

logic nor

returns 1 (true) only if a=0 and b=0

not(a)

logic not

returns 0 (false) if a 0 , else 1

nxor(a,b)

logic exclusive-nor

returns 1 (true) only if a=b

or(a,b)

logic or

returns 0 (false) only if a=0 and b=0

pi

RM

Pi greek

3.141.

PolyCh(x,n)

Chebycev's polynomials

x , orthog. for -1 x 1

PolyHe(x,n)

Hermite's polynomials

x , orthog. for x +

PolyLa(x,n)

Laguerre's polynomials

x , orthog. for 0 x 1

PolyLe(x,n)

Legendre's polynomials

x , orthog. for -1 x 1

rad(x)

RM

radians conversion

converts radians into current unit of angle

re(x)

real part of complex number

rnd(x)

random

root(x,n)

RCM

n-th root

returns a random number between x and

0
argument x 0 (the same as x^(1/n)

round(x,d)

RM

round a number with d decimal

round(1.35712, 2) = 1.36

sec(x)

RM

secant

sech(x)

RCM

hyperbolic secant

argument (in radians) x k*/2 with k =

1, 2
argument x>1

serie(.)

serie expansion of complex value z

Serie('z/n','n',1,10,'z',1+3i)

sgn(x)

RC

sign

returns 1 if x >0 , 0 if x=0, -1 if x<0

sin(x)

RCM

sin

argument in radians

sinh(x)

RCM

hyperbolic sine

272

Gamma constant 0.577..

Xnumbers Tutorial
sqr(x)

RCM

square root

sqr(2) =1.4142135623731, also 2^(1/2)

tan(x)

RCM

tangent

argument (in radians) x k*/2 with k =

1, 2

tanh(x)

RCM

hyperbolic tangent

xor(a,b)

returns 1 (true) only if a b

BesselI(x,n)

BesselJ(x,n)

BesselK(x,n)

BesselY(x,n)

J0(x)

logic exclusive-or
Bessel's function of 1st kind, nth
order, mod.
Bessel's function of 1st kind, nth
order
Bessel's function of 2nd kind, nth
order, mod.
Bessel's function of 2nd kind, nth
order
Bessel's function of 1st kind

Y0(x)

Bessel's function of 2nd kind

x 0

I0(x)

x >0

K0(x)

Si(x)

Bessel's function of 1st kind, modified

Bessel's function of 2nd kind,
modified
Sine integral

Ci(x)

Cosine integral

x >0

zeta(x)

Riemman's zeta function

argument x<-1 or x>1

AiryA(x)

Airy function Ai(x)

AiryB(x)

Airy function Bi(x)

Mean(a,b,...)

RM

Arithmetic mean

Meanq(a,b,...)

RM

Quadratic mean

meanq(8,9,12,9,7,10) = 9.300

Meang(a,b,...)

RM

Arithmetic mean

meang(8,9,12,9,7,10) = 9.035

Var(a,b,...)

RM

Variance

var(1,2,3,4,5,6,7) = 4.6666

Varp(a,b,...)

RM

Variance pop.

varp(1,2,3,4,5,6,7) = 4

Stdev(a,b,...)

RM

Standard deviation

Stdev(1,2,3,4,5,6,7) = 2.1602

Stdevp(a,b,...)

RM

Standard deviation pop.

Stdevp(1,2,3,4,5,6,7) = 2

Step(x,a)

Haveside's step function

Returns 1 if x a , 0 otherwise

DSBeta(x, a, b, [j])

Beta distribution (j=1 cumulative)

0 < x < 1 , a > 0, b > 0,

DSBinomial(k, n, p, [j])

k integer, n integer , 0 < p < 1

DSCauchy(x, m, s, n, [j])

DSChi(x, r, [j])

Binomial distribution (j=1 cumulative)

Cauchy (generalized) distribution (j=1
cumul.)
Chi distribution (j=1 cumulative)

DSErlang(x, k, l, [j]))

Erlang distribution (j=1 cumulative)

k integer, x > 0

DSGamma(x, k, l, [j]))

Gamma distribution (j=1 cumulative)

x > 0, k > 0, l > 0

DSLevy(x, l, [j]))

x > 0, l > 0

DSLogNormal(x, m, s, [j]))

DSLogistic(x, m, s, [j]))

DSMaxwell(x, a, [j]))

DSMises(x, k, [j]))

DSNormal(x, m, s, [j]))

Levy distribution (j=1 cumulative)

Log-normal distribution (j=1
cumulative)
Logistic distribution (j=1 cumulative)
Maxwell-Boltzman distribution (j=1
cumulative)
Von Mises distribution (j=1
cumulative)
Normal distribution (j=1 cumulative)

DSPoisson(k, z, [j]))

Poisson distribution (j=1 cumulative)

k integer, z > 0

DSRayleigh(x, s, [j]))

Rayleigh distribution (j=1 cumulative)

x > 0, s > 0

DSRice(x, v, s, [j]))

Rice distribution (j=1 cumulative)

x > 0, v 0 , s > 0

DSStudent(t, v, [j]))

Student distribution (j=1 cumulative)

v integer degree of freedom

DSWeibull(x, k, l, [j]))

Weibull distribution (j=1 cumulative)

x > 0, k integer, l > 0

x >0 , n = 0,1, 2, 3
x 0 , n = 0, 1, 2, 3
x >0 , n = 0,1, 2, 3
x 0 , n = 0,1, 2, 3
x 0

x >0

x, example cbr(2) = 1.2599, cbr(-2) = 1.2599

argument (in radian) x k*/2 with k =
1, 2
mean(8,9,12,9,7,10) = 9.1666

n integer , s > 0
r integer, x > 0

x > 0, m 0, s > 0
x > 0, m 0, s > 0
x > 0, a > 0
k > 0, - < x <
s>0

Symbol "!" is the same as "Fact", symbol "\" is the integer division, symbols |x| is the same as Abs(x)
Logical function and operators returns 1 (true) or 0 (false)
Note: the arguments separator of the functions changes automatically from "," to ";" if your system decimal separator is
comma ",".

273

Xnumbers Tutorial

Function Optimization
Macros for optimization on site
These macros has been ideated for performing the optimization task directly on the worksheet.
This means that you can define any function that you want simply using the standard Excel
built-in functions.
Objective function. For example: if you want to search the minimum of the bivariate function
51
35
) + ( y 100
)
f ( x, y ) = (x 100
2

insert in the cell E4 the formula "=(B4-0.51)^2+(C4-0.35)^2", where the cells B4 and C4
contain the current values of the variables x and y respectively. Changing the values of B4 e/o
C4 the function value E4 also changes consequently.

For optimization, you can choose two different algorithms

Downhill-Simplex 1
The NelderMead downhill simplex algorithm is a popular derivative-free
optimization method. Although there are no theoretical results on the
convergence of this algorithm, it works very well on a wide range of
practical problems. It is a good choice when a one-off solution is wanted
with minimum programming effort. It can also be used to minimize
functions that are not differentiable, or we cannot differentiate.
It shows a very robust behavior and converges for a very large set of
starting points. In our experience is the best general purpose algorithm,
solid as a rock, it's a "jack" for all trades.
Divide-Conquer 1D
For monovariable function only, it is an high robust derivative free
algorithm. It is simply a modified version of the bisection algorithm
Adapt for every function, smooth or discontinue.
It converges for very large segments. Starting point not necessary

For mono and

multivariate
functions

For monovariable
function only. It
needs the segment
where the max or
min is located

Example assume to have to minimize the following function for x > 0

f ( x) = e 3 x sin(3 x) + e x cos(4 x)
1

The Downhill-Simplex of Nelder and Maid routine appears by the courtesy of Luis Isaac Ramos Garcia

274

Xnumbers Tutorial

We try to search the minimum in the range 0 < x < 10

Choose a cell for the variable x , example B6, and insert the function
= SIN(3*B6)*EXP(- 2*B6) + COS(4*B6)*EXP(-B6)

in a cell that you like, for example C6.

After this, add the constrain values into another range, for example B3:C3
The values of the variables at the start are not important

Select the cell of the function C6 and start the macro "1D divide and conquer", filling the input
field as shown

Stopping limit. Set the maximum evaluation points allowed.

Max/Min. The radio buttons switches between the minimization and maximization algorithm
The "Downhill-Simplex" macro is similar except that:
The constrain box is optional.
It accepts up to 9 variables (range form 1 to 9 cells)
The algorithm starts from the point that you give in the variable cells. If the constrain
box is present, the algorithm starts from a random point inside the box
Let's see how it works with some examples1.

The following examples are extracted from "Optimization and Nonlinear Fitting" , Foxes Team, Nov. 2004

275

Xnumbers Tutorial

Example 1 - Rosenbrock's parabolic valley

This family of test functions is well known to be a minimizing problem of high difficult

f ( x, y ) = m ( y x 2 ) + (1 x )
2

The parameter "m" tunes the difficult: high value means high difficult in minimum searching.
The reason is that the minimum is located in a large flat region with a very low slope. The
following 3D plot shows the Rosenbrock's parabolic valley for m = 100

The following contour plot is obtained for m = 10

The function is always positive except in the point (1, 1) where it is 0. it is simple to
demonstrate it, taking the gradient

f = 0

4m x 3 + 2 x(1 2m y ) 2 = 0

2m( y x 2 ) = 0

276

Xnumbers Tutorial
From the second equation, we get

2m( y x 2 ) = 0

y = x2

2x 2 = 0

Substituting in the first equation, we have

4m x 3 + 2 x (1 2m x 2 ) 2 = 0

x =1

So the only extreme is the point (1, 1) that is the absolute minimum of the function
To find numerically the minimum, let's arrange a similar sheet.
We can insert the function and the parameters as we like
Select the cell D4 - containing the objective function - and start the macro "Downhill-Simplex".
The macro fills automatically the variables-field with the cells related to the objective function.
But, In that case, the cell A4 contains the parameter m that the macro must not change. So
insert only the range B4:C4 int the variables field.

The cells B4:C4 will change for minimizing the

objective function in the cell D4
Leave empty the constraint input box and press "Run"
Starting from the point (0, 0) we obtain the following good results
m
10
100

Algorithm
Simplex
Simplex

error

2.16E-13
4.19E-13

where the error is calculated as |x-1|+|y-1|

Example 2 - Constrained minimization

Example: assume to have to minimize the following function

f ( x, y ) = x 2 + 2 xy 4 x + 4 y 2 10 y + 7
with the ranges constrains

0 x 2 , 0 y 0.5
The Excel arrangement can be like the following

277

time
2 sec
2 sec

Xnumbers Tutorial

Compare with the exact solution x = 1.5, y = 0.5

Note that the function has a free minimum at x = 1, y = 1
Repeat the example living empty the constrains box input, for finding those free extremes.

278

Xnumbers Tutorial

Example 3 - Nonlinear Regression with Absolute Sum

This example explains how to perform a nonlinear regression with an objective function
different from the "Least Squared". In this example we adopt the "Absolute Sum".
We choose the exponential model

f ( x, a , k ) = a e k x
The goal of the regression is to find the best couple of parameters (a, k) that minimizes the sum
of the absolute errors between the regression model and the given data set.

AS = | yi f ( xi , a, k ) |
The objective function AS depends only by parameter a, k. Giving in input this function to our
optimization algorithm we hope to solve the regression problem
A possible arrangement of the worksheet may be:

We hope that changing the parameters "a" and "k" int the cells E2 and F3, the objective
function (yellow cell) goes to its minimum value. Note that the objective function depends
indirectly by the parameters a and k.
1.2
1
0.8

0.6

y*

0.4
0.2

The starting condition is the following, where y

indicates the given data and y* is the
regression plot (a flat line at the beginning)
Start the Downhil-Simplex and insert the
appropriate ranges: objective function = G3
and variables = E3:F3

0
0

0.2

0.4

0.6

0.8

1.2

Starting form the point (1, 0) you will see the cells changing quickly until the macro stops itself
leaving the following "best" fitting parameters and the values of the regression y*
Best fitting parameters
a

-2

The plot of the y* function and the samples

y are shown in the graph. As we can see
the regression fits perfectly the given
dataset.

279

Xnumbers Tutorial

Example 4 - Optimization of Integral function

The method of optimization on-site used by Xnumbers macros and by the Excel Solver is very
flexible. This is expecially useful when the optimization function has a complicated form.
Assume, for example, to have to maximize the following integral respect to the parameter
1

x( x)e

dx

The function integr returns the integral in

the cell B5. The integration limits and the
parameter are contained in the cells B2,
B3 and B4 respectively.
Start the optimization macro that you like,
setting the objective cell B5 and changing
the cell B4

Here, we have choose three different algorithms. They seem work very fine for this problem

(max)

Algorithm

error

Divide-conquer

2.210360464

9.8E-09

Downhill-Simplex
Excel Solver

2.210360527
2.210361438

5.4E-08
9.6E-07

Another example. Find the maximum of the following function for 0 < < 2

f ( ) = sin( x)dx
0

Here we use the Excel Solver for finding the

value of the cell D2 (the variable )
maximizing the cell E2 (the integral),
subjected to the constraints 0 < < 2 .
The final result w = 0.74201929194068
has an high precision, better 1E-8.

280

Xnumbers Tutorial

How to call Xnumbers functions from VBA

There are two ways for calling Xnumbers functions from other user VBA macros

by reference method

by Application.Run method

Both methods have advantages and drawbacks. Let's see.

Reference method
Prepare a new worksheet called, for example test.xls
First of all , after we have installed
and loaded Xnumbers.xla in Excel,
turn on the VBA editor by shift+F11
and add a new module.
Our VBA editor should look like this.
In particular we must see the
XNUMBERS (xnumbers.xla) project
and our new project VBAProject
(test.xls)

Now we add a reference link in our

VBAProject to XNUMBERS project
From the menu tool/ references...
Switch on the box of XNUMBERS library and click OK

Note that a references object has been added to our project

From now on all functions (except the functions declared private) of XNUMBERS are available
in our macros. Of course all functions that you use in worksheet are public. But not all the
functions that we can use in a worksheet can be called from VBA and vice versa.
The link that we have added is permanent. This means that when we load our test.xls file, the
Xnumbers.xla will also be automatically loaded.
If you want to remove the link to Xnumbers.xla simply return to menu tools/references and
deselect the box
Now prepare the following simple macro that reads two numbers from the sheet and returns the
division in multiprecision

281

Xnumbers Tutorial

Sub Test()
Dim a, b, c, DgtMax
DgtMax = Range("B2")
a = Range("B3")
b = Range("B4")
c = XNUMBERS.xdiv(a, b,
DgtMax)
Range("B5") = "'" & c
End Sub
Note that the above code works also if we write the compact statement

c = xdiv(a, b, DgtMax)
The XNUMBERS prefix is not necessary; it is useful only - at the developing time - to generate
automatically the list of the functions when we write the dot "." after XNUMBERS.
The result of the Test macro will be:

Of course the x-functions can be nested as any other function.

For example assume now to have to compute the expression = a * Ln(b)
The code will be
Sub Test1()
Dim a, b, c, DgtMax
DgtMax = Range("B2")
a = Range("B3")
b = Range("B4")
c = xmult(a, xln(b, DgtMax), DgtMax)
Range("B5") = "'" & c
End Sub
The output of the Test1 macro will be

282

Xnumbers Tutorial
As we have seen calling x-function using the reference method is very easy and straightforward
because it works as we have "virtually" copied all the Xnumbers.xla code in our project
The main drawback is that the link set in our workbook is static. This mean that if we move or
install the Xnumbers.xla addin in another directory, also the link must be removed and then
rebuilt.
Application.Run method
A metod for calling external functions that always works, independently where you have
installed the addin, uses the Excel Application.Run method. Using this method we do not need
to add any reference link. We have only to load the addin in Excel.
Therefore our VBA environment looks like as simple as possible and the code of the macro
Test will modified as the following
Sub test()
Dim a, b, c, DgtMax
DgtMax = Range("B2")
a = Range("B3")
b = Range("B4")
With Application
c = .Run("xdiv", a, b,
DgtMax)
End With
Range("B5") = "'" & c
End Sub
The first argument of the "Run" method is the name of the function that we want to call, that is
"xdiv" in our case (note the quotes around the name).
The other arguments are, in sequence, the parameters that we pass to the function xdiv.
The "Run" method returns in the variable c the result of the xdiv function.
This method is both simple and efficient. The only drawback is that it cannot accept nested
functions. Thus the Test1 macro becomes a bit more tricky.
Sub test1()
Dim a, b, c, DgtMax
DgtMax = Range("B2")
a = Range("B3")
b = Range("B4")
With Application
c = .Run("xLn", b, DgtMax)
c = .Run("xmult", a, c, DgtMax)
End With
Range("B5") = "'" & c
End Sub
The final result, of course, is the same.

283

Xnumbers Tutorial

References & Resources

"MSIEVE: A Library for Factoring Large Integers", Jason Papadopoulos , v1.17, March 2007
"LAPACK -- Linear Algebra PACKage 3.0", Updated: May 31, 2000
"EISPACK Guide, Matrix Eigensystem Routines, Smith B. T. at al. 1976
"Numerical Analysis" F. Sheid, McGraw-Hill Book Company, New-York, 1968
"Handbook for Autom. Computation - Linear Algebra", vol II, Wilkinson, Martin, Peterson, 1970
"Matrix Analysis and Applied Linear Algebra", C. D. Mayer, Siam, 2000
"Linear Algebra", J. Hefferon, Saint Michaels College, Colchester, Vermont, 2001,
"Numerical Methods that usually work", F. S. Acton, The Mathematica Association of America, 1990
"Analysis Nunerical Methods", E. Isaacson, H. B. Keller, Wiles & Sons, 1966
"Calculo Numrico", Neide M. B. Franco, 2002
"Metodos Numericos" Sergio R. De Freitas, 2000
"Numerical Mathematics in Scientific Computation", G. Dahlquist, . Bjrck, vol II.
"Numerical Recipes in FORTRAN 77- The Art of Scientific Computing - 1986-1992 by Cambridge
University Press. Programs Copyright (C) 1986-1992 by Numerical Recipes Software
"Computation of Special Functions", by Shanjie Zhang and Jianming Jin - John Wiley and Sons, Inc
"Lanczos Implementation of the Gamma Function" by Paul Godfrey, Intersil Corp, 2001
"clsMathParser - A Class for Math Expressions Evaluation in Visual Basic", Leonardo Volpi , Michael
Ruder, Thomas Zeutschler, Lieven Dossche, . 3.2 Jan. 2003, by .Volpi
"clsMathParserC -A Class for Complex Math Expressions Evaluation in Visual Basic", Arnaud de
Grammont, Leonardo Volpi, v. 3.2 , Jan. 2003
"F F T (Fast Fourier Transform)", Paul Bourke, June 1993, http://astronomy.swin.edu.au
"2 Dimensional FFT" , Paul Bourke, June 1998, http://astronomy.swin.edu.au
Solutions Numeriques des Equations Algebriques, E., Durand, Tome I, Masson,Paris ,1960.
A modified Newton method for Polynomials , W.,Ehrilich, Comm., ACM, 1967, 10, 107-108.
Ein Gesamschrittverfahren zur Berechnung der Nullstellen eines Polynoms, O., Kerner, Num.Math.,
1966, 8, 290-294.
Iteration methods for finding all the zeros of a polynomial simultaneously, O. Aberth, Math. Comp. ,1973,
27, 339-344.
The Ehrlich-Aberth Method for the nonsymmetric tridiagonal eigenvalue problem, D. A. Bini, L.
Gemignani, F, Tisseur, AMS subject classifications. 65F15
Progress on the implementetion of Aberths method, D. A. Bini, G. Fiorentino, , 2002, The FRISCO
Consortiunm (LTR 21.024)
A Method for Finding Real and Complex Roots of Real Polynomials With Multiple Roots, by C. Bond, July
10, 2003
A Robust Strategy for Finding All Real and Complex Roots of Real Polynomials, by C. Bond, April 24,
2003
"Nonlinear regression", Gordon K. Smyth Vol. 3, pp 1405/1411, in Encyclopedia of Environmetrics (ISBN
0471 899976), Edited by John Wiley & Sons, Ltd, Chichester, 2002
"Optimization", Gordon K. Smyth Vol. 3, pp 1481/1487, in Encyclopedia of Environmetrics (ISBN 0471
899976), Edited by John Wiley & Sons, Ltd, Chichester, 2002
"Process Modeling", The National Institute of Standards and Technology (NIST) website for Statistical
Reference Datasets, (http://www.itl.nist.gov/div898/handbook/pmd/pmd)
"Metodos numericos con Matlab"; J. M Mathewss et al.; Prentice Hall
"Genetic and Nelder-Mead", E. Chelouan, et al, EJOR 148(2003) 335-348
"Convergence properties", J.C. Lagarias, et al, SIAM J Optim. 9(1), 112-147

284

Xnumbers Tutorial
"Optimization for Engineering Systems", Ralph W. Pike, 2001, Louisiana State University
(http://www.mpri.lsu.edu/bookindex)
"Zeros of Orthogonal Polynomials" , Leonardo Volpi, Foxes Team
(http://digilander.libero.it/foxes/Documents)
"How to tabulate the Legendres polynomials coefficients" Leonardo Volpi, Foxes Team
(http://digilander.libero.it/foxes/Documents)
"DE-Quadrature (Numerical Automatic Integrator) Package", by Takuya OOURA, Copyright(C) 1996
Advanced Excel for scientific data analysis", Robert de Levie, 2004, Oxford University Press
"Microsoft Excel 2000 and 2003 Faults, Problems, Workarounds and Fixes", David A. Heiser, web site
http://www.daheiser.info/excel/frontpage.html
"NIST/SEMATECH e-Handbook of Statistical Methods", January 26, 2005
(http://www.itl.nist.gov/div898/handbook)
"DE-Quadrature (Numerical Automatic Integrator) Package", by Takuya OOURA, Copyright(C) 1996,
Research Institute for Mathematical Sciences, Kyoto University, Kyoto 606-01, Japan
(http://momonga.t.u-tokyo.ac.jp/~ooura)
"A Comparison of three high-precision quadrature schemes", David H. Bailey and Xiaoye S. Li, Copyright
2003, University of California (http://repositories.cdlib.org/lbnl/LBNL-53652)
"Asymptotic Bit Cost of Quadrature Formulas Obtained by Variable Transformation" P. Favati, Appl. Math.
Lett. Vol. 10, No. 3, pp. 1-7, 1997, Pergamon Copyright 1997
"Gli algoritmi della crittografia a chiave pubblica", Giovanni Fraterno, settembre 2000.
http://digilander.libero.it/crittazione
Bailey, D. H.; Borwein, J. M.; Calkin, N. J.; Girgensohn, R.; Luke, D. R.; and Moll, V. H. "Integer Relation
Detection." 2.2 in Experimental Mathematics in Action. Wellesley, MA: A K Peters, pp. 29-31, 2006.
Bailey, D. H. and Broadhurst, D. J. "Parallel Integer Relation Detection: Techniques and Applications."
Math. Comput. 70, 1719-1736, 2001.
Bailey, D. H. and Ferguson, H. R. P. "Numerical Results on Relations Between Numerical Constants Using
a New Algorithm." Math. Comput. 53, 649-656, 1989.
Bailey, D. and Plouffe, S. "Recognizing Numerical Constants."
http://www.cecm.sfu.ca/organics/papers/bailey/.
Borwein, J. M. and Lisonek, P. "Applications of Integer Relation Algorithms." To appear in Disc. Math.
http://www.cecm.sfu.ca/preprints/1997pp.html.
Weisstein, Eric W. "Integer Relation." From MathWorld--A Wolfram Web Resource.
http://mathworld.wolfram.com/IntegerRelation.html
Helaman R. P. Ferguson and David H. Bailey , "A Polynomial Time, Numerically Stable Integer Relation
Algorithm" RNR Technical Report RNR-91-032, July 14, 1992

285

Xnumbers Tutorial

Credits
Software developed
Xnumbers contains code developed by the following authors that kindly contributed to this
collection.
Luis Isaac Ramos Garcia

Orthogonal Polynomials and Downhill Simplex Nelder-Mead routine

Olgierd Zieba

Cubic Spline and documentation

Alfredo lvarez Valdivia

Robust Method fitting routines

Michael J. Kozluk

Log Relative Error function, Linear regression debugging, and

documentation (StRD benchmark)

Ton Jeursen

Format function, Xnumber debugging; Xnumber for Excel95,

xBaseChange - Multiprecision BaseChange
xCalc - multiprecision conversion function

Richard Huxtable

ChangeBase and Prime functions

Michael Ruder

MathParser improvement and debugging

Thomas Zeutschler

MathParser improvement and debugging

Lieven Dossche

Class MathParser development

Arnaud de Grammont

Complex MathParser developement

Rodrigo Farinha

MathParser improvement and debugging

Vladimir Zakharov

Installation and initialization improvement

Hans Gnter

Extension of exponential integral for negative x

David A. Heiser

Weilford's algorithm for statistic standard deviation

John Jones

Modular arithmetic and Xnumber converter parser

Software translated
Xnumbers contains VB code translated from the following packages:
Takuya OOURA

DE-Quadrature (Numerical Automatic Integrator) Package

Shanjie Zhang - Jianming Jin

FORTRAN routines for computation of Special Functions

Software used
This documentation was developed thanks to:
Microsoft Help Workshop 4

Compiler for Help-on line

Microsoft Word and Equation Editor

RTF main document and formulas

Winplot ( Peanut Software by Richard Parris)

Math plots and graphical elaboration

Derive 5 (Texas Instrument)

Math graph and symbolic elaboration

MathGV 3.1 (by Greg VanMullem)

Math graphs and bitmap elaboration

Msieve 1.17

by Jason Papadopoulos. Large number factoring

Many thanks to everybody that have contributed to this document and made it more intelligible.
I am especially grateful for invaluable help on many occasions to Robert de Levie.

286

Xnumbers Tutorial

Analytical Index
Analytical index
Complex Function Integration (Romberg
method); 156
Complex Gamma Function; 104
Complex Hyperbolic Cosine; 103
Complex Hyperbolic Sine; 102
Complex Hyperbolic Tan; 103
Complex inverse; 101
Complex Inverse Hyperbolic Cos; 103
Complex Inverse Hyperbolic Sin; 103
Complex Inverse Hyperbolic Tan; 103
Complex Log; 101
Complex Logarithm Gamma Function; 104
Complex Multiplication; 98
Complex negative; 101
Complex power; 100
Complex Quadratic Equation; 105
Complex Roots; 100
Complex Series Evaluation; 184
Complex Sin; 102
Complex Subtraction; 98
Complex Tangent; 102
Complex Zeta Function; 104
Constant e; 59
Constant Ln(10); 59
Constant Ln(2); 59
Constant pi; 62
Continued Fraction; 108
Continued Fraction of Square Root; 109
Convert extended into double; 34
Convol; 189
corrector; 211
Corrector; 212; 214
Cos; 61
Cosine Integral Ci(x); 249
CosIntegral; 249
cplxabs; 99
cplxacos; 102
cplxacosh; 103
cplxadd; 98
cplxasin; 102
cplxasinh; 103
cplxatan; 102
cplxatanh; 103
cplxconj; 101
cplxcos; 102
cplxcosh; 103
cplxdigamma; 103
cplxdiv; 98
cplxei; 103
cplxEquation2; 105
cplxerf; 104
cplxerfc; 104
cplxeval; 260
cplxExp; 101
cplxgamma; 104
cplxgammaln; 104

A
Absolute; 28
Adams; 211; 212
Adams-Bashfort-Moulton; 211
Addition; 23
Airy functions; 252
AiryA; 252
AiryAD; 252
AiryB; 252
AiryBD; 252
Aitken; 158; 183
Arccos; 62
Arcsine; 62
Arctan; 62
Arithmetic Mean; 40
B
Base conversion; 240
baseChange; 240
Bessel functions; 249
BesselSphJ; 255
Beta function; 247
Bisection; 231
Bivariate Polynomial; 88
Brouncker-Pell Equation; 120
C
Cebychev; 78
Change sign; 28
Check digits; 32
Check Integer; 112
Check odd/even; 112
Check Prime; 109
Cholesky; 131
Coefficients of Orthogonal Polynomials; 96
Combinations; 39
Combinations function; 248
Complement of right angle; 63
Complex absolute; 99
Complex Addition; 98
Complex ArcCos; 102
Complex ArcSin; 102
Complex ArcTan; 102
Complex Complementary Error Function; 104
Complex conjugate; 101
Complex Cos; 102
Complex digamma; 103
Complex Division; 98
Complex Error Function; 104
Complex Exp; 101
Complex Exponential Integral; 103
Complex Expression Evaluation; 260

287

Xnumbers Tutorial

cplxintegr; 156
cplxinv; 101
cplxLn; 101
cplxmult; 98
cplxneg; 101
cplxpolar; 99
cplxpow; 100
cplxrect; 99
cplxroot; 100
cplxserie; 184
cplxsin; 102
cplxsinh; 102
cplxsub; 98
cplxtan; 102
cplxtanh; 103
cplxzeta; 104
Crout; 131
cspline_coeff; 199
cspline_eval; 197
cspline_interp; 197
cspline_pre; 197
Cubic Spline 2nd derivatives; 197
Cubic Spline Coefficients; 199
cvBaseDec; 240
cvBinDec; 240
cvDecBase; 240
cvDecBin; 240

errfun; 244
Error Function Erf(x); 244
Euler; 211
Euler constant ; 60
Euler-Mascheroni Constant; 245
Euler's Totient function; 121
exp_integr; 244
expn_integr; 244
Exponential; 58
Exponential any base; 58
Exponential integral Ei(x); 244
Exponential integral En(x); 244
Extended Number Check; 32
F
Factor; 116
Factorial; 39
Factorial with double-step; 39
Factorize; 113
Factorize function; 116
FFT; 137; 142
FFT_INV; 139
FFT2D; 141
FFT2D_INV; 141
Fibonacci numbers; 250
First Derivative; 234
Flip; 33
Format Extended Number; 32
Fourier; 165
Fourier_cos; 164
Fourier_cos; 166
Fourier_cos; 166
Fourier_sin; 164
fract; 107
Fract_Interp; 195
Fract_Interp_Coef; 195
FractCont; 108
FractContSqr; 109
Fresnel cosine Integral; 250
Fresnel sine Integral; 250
Fresnel_cos; 250
Fresnel_sin; 250
Function Integration (Double Exponential
method); 152
Function Integration (mixed method); 154
Function Integration (Newton-Cotes); 159
Function Integration (Romberg method); 151

D
Data Conditioning; 53
Data Integration (Newton Cotes); 157
dBel; 240
Decibel; 240
Decimal part; 29
DFSP; 140
DFSP_INV; 140
DFT; 137; 142
DFT_INV; 139
Diff1; 234
Diff2; 235
digamma; 247
Digamma function; 247
Digit_Max; 22; 35
Digits count; 31
Digits sum; 33
DigitsAllDiff; 32
Diophantine; 119
Diophantine Equation; 119
DiophEqu; 119
Discrete 2D Fourier Transform; 141
Discrete Convolution; 189
Discrete Fourier Inverse Transform; 139
Discrete Fourier Spectrum; 140
Discrete Fourier Transform; 137
Division; 24
Double Data integration; 179
Double Exponential; 152
Double Integral; 168
Double integration function; 174
Double Series; 185
Downhill; 268
DPOLYN; 74

G
Gamma F-factor; 247
Gamma function); 245
Gamma quotient; 246
GCD; 106
Geometric Mean; 40
Grad; 235
Gradient; 235
Greatest Common Divisor; 106
H
Hermite; 78
Hessian; 236
Hessian matrix; 236
Hyperbolic Arc Cosine; 60
Hyperbolic Arc Sine; 59

E
Elliptic Integrals; 252

288

Xnumbers Tutorial

Hyperbolic Arc Tangent; 60

Hyperbolic Cosine; 60
Hyperbolic Sine; 59
Hyperbolic Tangent; 60
Hypergeometric function; 251
Hypgeom; 251

Linear Regression - Coefficients; 50

Linear Regression - Standard Deviation of
Estimate; 45
Linear Regression Coefficients; 43
Linear Regression Covariance Matrix; 47
Linear Regression Evaluation; 49
Linear Regression Formulas; 46
Linear Regression Min-Max; 56
Linear Regression Statistics; 48
Linear Regression with Robust Method; 55
Log Gamma function; 246
Log Relative Error; 242
Logarithm in any base; 58
Logarithm natural (Napiers); 58

I
I BesselSphY; 255
IElliptic1; 252
IElliptic2; 252
Incomplete Beta function; 248
Incomplete Gamma function; 248
Infinite integral; 177
Infinite Integration of oscillating functions; 165
Integer Division; 24
Integer part; 29
Integer polynomial; 86
Integer relation; 122
Integer Remainder; 25
Integer roots; 70
integr; 166
Integr; 154
Integr_2D; 174
Integr_fcos; 162
Integr_fsin; 162
Integr_nc; 159
Integr_ro; 151
Integr2D; 168
Integr3D; 171
Integral function; 147
Integral of sine-cosine power; 254
Integral_Inf; 177
Integration of oscillating functions (Filon
formulas); 162
Integration of oscillating functions (Fourier
transform); 164
IntegrData; 147
IntegrData2D; 179
IntegrDataC; 157
Interp_Mesh; 200
Interpolation with continue fraction; 195
Interpolation with Cubic Spline; 197
IntPowCos; 254
IntPowSin; 254
Inverse; 24
Inverse 2D Discrete Fourier Transform; 141
Inverse Discrete Fourier Spectrum; 140

M
Macro for Double Integration; 168
Macro for Multiprecision Matrix operations; 135
Macro for Triple Integration; 171
Macro Integer relation finder; 126
Macro Mesh Fill; 201
Macro Poly Regression; 51
Macro Regression; 51
Macro Sampler; 145
Macros for optimization on site; 267
Macros X-Edit; 35
Math expression strings; 263
MathParser; 265
matrix; 135
Matrix Addition; 129
Matrix Determinant; 129
Matrix Inverse; 129
T
Matrix LL decomposition; 131
Matrix LU decomposition; 131
Matrix Modulus; 130
Matrix Multiplication; 129
Matrix Power; 130
Matrix Subtraction; 129
Maximum Common Divisor; 106
MCD; 106
MCM; 106
Mesh Interpolation 2D; 200
Minimum Common Multiple; 106
mjkLRE; 242
Modular Addition; 110
Modular Division; 110
Modular Multiplication; 110
modular power; 110
Modular Power; 110
Modular Subtraction; 110
Multiple roots; 72
Multiplication; 24
Multiprecision Base Conversion; 241
Multiprecision Excel Formula Evaluation; 261
Multiprecision Expression Evaluation; 256

J
Jacobian; 236
Jacobian matrix; 236
K
Kummer confluent hypergeometric functions;
253
Kummer1; 253
Kummer2; 253

N
Next Prime; 109
NextPrime; 109
Non Linear Equation Solving; 238
N-Root; 28
Numbers comparison; 31

L
LCM; 106
Least Common Multiple; 106
Legendre; 78

289

Xnumbers Tutorial

Polynomial shift; 81
Polynomial solving; 86
Polynomial subtraction; 77
Polynomial System of 2nd degree; 87
Polynomial writing; 76
PolyRadius; 82
PolyRem; 77
PolyShift; 81
PolySolve; 86
PolySub; 77
Polyterms; 80
PolyTerms; 75
predictor; 211
Predictor; 212
Predictor; 214
Prime; 109
Prime Numbers Generator; 117
Prime Test; 117
PrimeGenerator; 117
Product; 27
PSLQ; 122

Objective function; 267

OD Linear System; 222
ODE Implicit Predictor-Corrector; 219
ODE Multi-Steps; 211
ODE Predictor-Corrector 4; 217
ODE Runge-Kutta 4; 207
ODE_COR; 214
ODE_PC2I; 219
ODE_PC4; 217
ODE_PRE; 214
ODE_RK4; 207
ODE_SYSL; 222
Orthogonal polynomials; 92
Orthogonal Polynomials; 92
Orthogonal Polynomials evaluation; 93
oscillating; 165
P
Partial; 89
PECE; 214; 216
Perfect Square; 112
permutation; 40
Permutations; 40
Polar Conversion; 99
Poly_ChebychevT; 93
Poly_ChebychevU; 93
Poly_Gegenbauer; 93
Poly_Hermite; 93
Poly_Jacobi; 93
Poly_Laguerre; 93
Poly_Legendre; 93
Poly_Weight_ChebychevT; 95
Poly_Weight_ChebychevU; 95
Poly_Weight_Gegenbauer; 95
Poly_Weight_Hermite; 95
Poly_Weight_Jacobi; 95
Poly_Weight_Laguerre; 95
Poly_Weight_Legendre; 95
PolyAdd; 76
PolyBuild; 83
PolyBuildCfx; 85
PolyCenter; 81
PolyDiv; 77
PolyInt; 86
PolyInterp; 191
PolyInterpCoef; 191
PolyMult; 76
POLYN; 73
POLYN2; 88
Polynomial addition; 76
Polynomial building from roots; 83
Polynomial building with multi-precision; 85
Polynomial center; 81
Polynomial coefficients; 75
Polynomial derivatives; 74
Polynomial division quotient; 77
Polynomial division remainder; 77
Polynomial evaluation; 73
Polynomial interpolation; 191
Polynomial multiplication; 76
Polynomial Regression - Standard Deviation of
Estimates; 50
Polynomial Regression Statistics; 50
Polynomial roots radius; 82

Q
Quadratic Mean; 40
R
Raise to power; 27
Rational Fraction approximation; 107
Rectangular Conversion; 99
RegLinMM; 56
RegLinRM; 55
regression; 272
Relative Rounding; 30
Root Error Estimation; 68
Rounding; 29
S
Scalar Product; 130
Scientific Format; 34
Secant; 232
Second Derivative; 235
Serie_trig; 186
Serie2D_trig; 188
Series acceleration with ^2; 183
Series Evaluation; 182
sign; 31
Significant Digits count; 31
Similarity Transform; 130
Simplex; 268
Sin; 61
Sine Integral Si(x); 249
SinIntegral; 249
Solve Linear Equation System; 132
SortRange; 33
Spherical Bessel functions of integer order; 255
Split scientific format; 34
Square Delta Extrapolation; 133
Square Root; 28
Standard Deviation; 40
Sub-Tabulation; 53
Subtraction; 23
Sum; 27
sumDigits; 33
SYSPOLY2; 87

290

Xnumbers Tutorial

xeu; 60
xeval; 26; 256
xevall; 256
xexp; 58
xfact; 39
xfact2; 39
xFib; 250
xFormat; 32
xfrac; 107
xFractCont; 108
xGammaF; 247
xGammaI; 248
xGammaln; 246
xGammalog; 246
xGammaQ; 246
xGm; 245
xgmean; 40
xint; 29
xinv; 24
xIsInteger; 112
xIsOdd; 112
xIsSquare; 112
xIsXnumber; 32
xLn; 58
xLn10; 59
xLn2; 59
xLog; 58
xLRE; 242
xMatAbs; 130
xMatAdd; 129
xMatBAB; 130
xMatDet; 129
xMatInv; 129
xMatLL; 131
xMatLU; 131
xMatMult; 129
xMatPow; 130
xMatSub; 129
xMCD; 106
xMCM; 106
xmean; 40
xmult; 24
xmultmod; 110
xneg; 23; 28
xpi; 62
xpi2; 62
xpi4; 62
xpow; 27
xpowmod; 110
xProdScal; 130
xProdVect; 132
xqmean; 40
xRegLinCoef; 43
xRegLinCoef; 45
xRegLinEval; 49
xroot; 28
xround; 29
xroundr; 30
xSerie2D; 185
xsin; 61
xsinh; 59
xsplit; 34
xsqr; 28
xstatis; 42
xstdev; 40
xstdevp; 40

Tan; 62
tanh-sinh transformation; 152
Totient; 121
Trigonometric double serie; 188
Trigonometric series; 186
Truncating; 29
U
Univariate Statistic; 42
V
Variance; 41
Vector flip; 33
Vector Product; 132
W
Weight of Orhogonal Polynomials; 95
X
x2pi; 62
xabs; 28
xacos; 62
xacosh; 60
xadd; 23
xaddmod; 110
xanglecompl; 63
xasin; 62
xasinh; 59
xatan; 62
xatanh; 60
xBaseChange; 241
xBeta; 247
xBetai; 248
xCalc; 261
xcdbl; 34
xcomb; 39
xcomb_big; 248
xcomp; 31
xcos; 61
xcosh; 60
xcplxabs; 99
xcplxadd; 98
xcplxconj; 101
xcplxdiv; 98
xcplxExp; 101
xcplxinv; 101
xcplxLn; 101
xcplxmult; 98
xcplxneg; 101
xcplxpolar; 99
xcplxpow; 100
xcplxrect; 99
xcplxroot; 100
xcplxsub; 98
xcvexp; 34
xdec; 29
xDgt; 31
xdiv; 24
xdivint; 24
xdivmod; 110
xdivrem; 25
xe; 59

291

Xnumbers Tutorial

xsub; 23
xsubmod; 110
xsum; 27
xSYSLIN; 132
xtan; 62
xtanh; 60
xtrunc; 29
xUnformat; 32
xvar; 41

xvarp; 41
Z
Zero_bisec; 231
Zero_sec; 232
Zeros of Orthogonal Polynomials; 95
Zeta; 251
Zeta function; 251

292

Xnumbers Tutorial

2007, by Foxes Team

ITALY
Oct 2007

293