O'Connor - Matlab's Floating Point System
O'Connor - Matlab's Floating Point System
O'Connor - Matlab's Floating Point System
NUMERICAL ALGORITHMS
Dr Derek O’Connor
1 PURPOSE
The purpose of this exercise is to familiarize you with Matlab’s floating-point arithmetic sys-
tem which is, by default, IEEE double precision. It is important to know the parameters of the
number system you are using and to be aware of its limitations. Although M ATLAB is quite
good at telling you what these are, other systems or compilers are not—try doing this exercise
in Microsoft’s Excel.
2 EXERCISE
1. Write a M ATLAB function MachEps() that calculates machine epsilon. Although M ATLAB
has the built-in function eps to do this, it is important to write your own because other
languages do not have a built-in MachEps function. If written correctly, an implicit bonus
of this function is that it counts the number of bits of precision in the F.P. system used.
Try to find this once you have the function working properly — requires no extra code
except passing back an extra parameter (argument).
2. M ATLAB has various important constants and functions built in. Some of these are eps,
realmax, realmin, inf, pi. Also available are date, ver, and version which are useful for
annotating output.
Determine what these are on your system.
0 ∞
, , ∞ ∗ 0, 1∞ , ∞ − ∞, 00 , ∞0 .
0 ∞
Do you agree with M ATLAB’s results? Explain.
1
Lab Exercise No. 3 : Matlab’s Floating Point Number System Due : Wed 4 Oct. 2006
5. sin(π ) = 0, cos(π ) = −1, and sin2 (π ) + cos2 (π ) = 1. What does M ATLAB give for
sin(pi), cos(pi), sin(pi)ˆ2+cos(pi)ˆ2 ? Explain.
You can get a deeper understanding of what M ATLAB is doing by looking at the binary form of
the numbers above. Use the function num2bin(x) to see a binary version of the decimal number
x. This is available on the class website.
3 SOLUTION
Machine epsilon (ǫm ) is the spacing between the 1.0 and the next higher floating point number.
Knowing this number allows us to calculate the spacing about any number x as ǫm | x|. For
some reason, students have difficulty understanding the meaning and significance of this (f.p)
number. This is one of the reasons for getting you to calculate ǫm in M ATLAB, even though it
is a built-in function (constant?). The other reason is that many languages or systems do not
provide an eps function and so you will have to provide your own.
The function MachEps uses the fact that ǫm is the smallest floating point number such that fl(1.0 + ǫm ) >
1.0. It starts off with ǫm = epsil = 1.0 and repeatedly halves it (base 2) until fl(1.0 + ǫm ) ≤ 1.0.
Thus fl(1.0 + ǫm ) = 1.0 + ǫm at each iteration, except at the end of the last iteration. Thus, it
generates a sequence of contiguous floating point numbers, except the last. You need to think
about that.
%-------------------------------------------
function [meps, prec] = MachEps()
%
% Determines Machine Epsilon and precision in bits.
% Derek O’Connor, Oct 2004.
%-------------------------------------------
k = 1;
epsil = 1.0;
epone = 1.0 + epsil;
while epone > 1.0
epsil = epsil/2.0;
epone = 1.0 + epsil;
k = k + 1;
end;
% -- gone one iteration too far, so :
meps = 2.0*epsil;
prec = k-1;
%---------- End [meps, prec] = MachEps()-------
Figure 1 shows the last 3 iterations of MachEps. The function has generated the sequence of
floating point numbers epone = {1 + 1, 1 + 2−1 , . . . , 1 + 2k ǫm , . . . , 1 + 22 ǫm , 1 + 21 ǫm , 1 + 20 ǫm }.
In the final iteration it generates the number epone = 1 + 2−1 ǫm . This is halfway between
1 + ǫm and 1, contiguous floating point (representable) numbers. Thus it is not representable
and must, therefore, be rounded.
Now comes the delicate bit. The IEEE standard allows four rounding modes :
k = p-1
k=p
k = p+1
1 1 + ε/2 1+ε 1 + 2ε 1 + 4ε
Mode 4, the round-to-nearest mode, is almost always used in practice. This rule is : round x to the
nearer of the two floating point numbers, x− and x+ , adjacent to x. In the case of a tie, choose
the one whose least significant bit is zero.
The last number generated by MachEps is epone = 1 + 2−1 ǫm = x, and this is equi-distant from
x− = 1 and x+ = 1 + ǫm . Which of these two has a zero least significant bit? The floating point
number x− = 1 has its last bit zero, and so epone = 1 + 2−1 ǫm is rounded to 1. See Figure 1.
Once epone becomes 1, the while –loop ends, with k = p + 1 and epsil = ǫm /2. Finally, the ad-
justment meps = 2.0*epsil; prec = k-1 is made. Table 1 shows the binary numbers generated
by MachEps.
M ATLAB 6.5 uses IEEE double precision for all calculations. The M ATLAB floating point pa-
rameters can be determined from the built-in functions realmax, realmin, eps, which give the
values shown in Table 2.
The number 1/realmin= 4.494232837155790e+307 is in the Floating Point number range but the
number 1.0/realmax= 5.562684646268004e-309 is not but is displayed in M ATLAB. Why? Be-
cause IEEE FP uses gradual underflow which allows numbers below the underflow threshold
realmin to exist as subnormals.
The table below gives M ATLAB’s response to various mathematical indeterminates. These are
the correct results according to the IEEE standard.
∞
1
0
1
∞
0
0 ∞ ∞× 0 1∞ ∞−∞ 00 ∞0
inf 0 NaN NaN NaN NaN NaN 1 1
The expressions in the Table 3 are mathematically meaningless, but computationally they are
important. This is why the IEEE standard has carefully specified the results of such floating
point operations.
Example 1 (Parallel Resistors). Here is a simple example that shows why operations with ∞
are necessary. The circuit below comprises a V −volt battery and two resistors of R1 ohms and
R2 ohms connected in parallel across its terminals. We wish to calculate the currents i, i1 , and
i2 and hence the power dissipated by the circuit, P = R1 i12 + R2 i22 watts.
The state of the circuit is determined by the following equations, which are derived from Ohm’s
and Kirchoff’s laws:
1
V = R1 i1 = R2 i2 = Ri, where R = 1 1
, and i = i1 + i2 .
R1 + R2
So, given V, R1 , and R2 , we wish to calculate R, i1 ,i2 , and i. The calculation of R can cause
floating point exceptions if one or both or the resistors is (i) 0 (short circuit) or (ii) ∞ (open
i1
R1
i2
R2
i V
1 1 1 1
Rsc = 1 1
= 1
= 0, Roc = 1 1
= = R2
0 + R2 ∞+ R2 ∞ + R2 0 + R12
These calculations give the correct physical results but many compilers and mathematical sys-
tems will either give an error or crash.
Although the example above may be of interest to electrical engineers only, the following ex-
ample shows the widespread need for proper exception handling
Example 2 (Calculating the Norm of a Vector.). This is a common problem in Numerical Linear
Algebra : the length or norm of the vector x = ( x1 , x2 , . . . , xn ) is
s
n
k x k2 = ∑ x2i .
i=1
This piece of code is not reliable. If any x(i ) is of the order 10200 then x(i ) ∗ x(i ) will cause
overflow, i.e., it will not fit in a computer word and a fatal error may occur. If the x(i ) values
are small then there is a danger of underflow or negative overflow, i.e., when a result becomes
too small it is set to zero. For example, if n = 10000 and xi = 10−200 then the length of this
vector is 10−180 . All of these numbers are perfectly valid computer numbers but the code above
will give 0 as the result, because x2i = 10−400 will be set to 0 and usually without warning that
underflow has occurred. We can overcome these range violation problems if we scale the data
before we calculate the norm. Here is a M ATLAB function NormS that scales small numbers up
and large numbers down.
Many linear algebra packages use scaling to avoid range violations. This code is time-consuming
and very often the extra work spent in scaling is not necessary.
The M ATLAB function NormCS above uses conditional scaling : it performs the ordinary unscaled
sum first and then uses floating point exceptions to determine if scaling is necessary. This
change can give dramatic speed-ups in large simulations that use the norm operation on many
different vectors.
Question 1. In the M ATLAB functions NormS and NormCS above, how should the values for
LARGE,SMALL,LScale,SScale be chosen? What do they depend on?
Mathematically we have
4 4 1 1 1 1
a= , b= −1 = , c= + + = 1, e = 1 − 1 = 0.
3 3 3 3 3 3
Performing these statements in F (b, p, −, −), where b is not a multiple (power?) of 3, we get
C===========================================================================
double precision function epslon (x)
C===========================================================================
double precision x
c
c estimate unit roundoff in quantities of size x.
c
double precision a,b,c,eps
c
c this program should function properly on all systems
c satisfying the following two assumptions,
c 1. the base used in representing dfloating point
c numbers is not a power of three.
c 2. the quantity a in statement 10 is represented to
c the accuracy used in dfloating point variables
c that are stored in memory.
c the statement number 10 and the go to 10 are intended to
c force optimizing compilers to generate code satisfying
c assumption 2.
c under these assumptions, it should be true that,
c a is not exactly equal to four-thirds,
c b has a zero for its last bit or digit,
c c is not exactly equal to one,
c eps measures the separation of 1.0 from
c the next larger dfloating point number.
c the developers of eispack would appreciate being informed ⋆
c about any systems where these assumptions do not hold. ⋆
c
c *****************************************************************
c this routine is one of the auxiliary routines used by eispack iii
c to avoid machine dependencies.
c *****************************************************************
c
c this version dated 4/6/83.
c
a = 4.0d0/3.0d0
10 b = a - 1.0d0
c = b + b + b
eps = dabs(c-1.0d0)
if (eps .eq. 0.0d0) go to 10
epslon = eps*dabs(x)
return
end
We know that sin(π ) = 0, cos(π ) = −1, and sin2 (π ) + cos2 (π ) = 1. But M ATLAB gives
As we can see, 2 and 3 are correct but 1 is not. There are two sources of error here : (i) the error
in representing π, a transcendental number, and (ii) the error in computing sin( x) or any other
function.
Representing π.
Let us consider the representation of π. For reference, here are the first 100 decimal digits of π
along with the binary representation :
3.141592653589793238462643383279502884197169399375105820974944592307816406286208998628034825342117067
11.0010010000111111011010101000100010000101101000110000100011010011000100110001100110001010001011100000
0011011100000111001101000100101001000000100100111000001000100010100110011111001100011101000000001000
0010111011111010100110001110110001001110011011001000100101000101001010000010000111100110001110001101
0000000100110111011110111110010101000110011011001111001101001110100100001100011011001100000010101100
00101001101101111100100101111100010100001101110011
M ATLAB gives the value of pi= 3.141592653589793 (2384626433832795028) and so all 16 digits
are correct. Here is a table of the floating point numbers just below and above pi. Notice that
each adjacent pair differs by 1 bit in the last place.
Table 4: Matlab’s pi 6= π
We see that the true value of π lies about 1/4 ≈ (.2384626433832795028) the way between pi
and the next higher floating point number pi+2*eps. Hence the error in pi is
If we assume that sin( x) and cos( x) are calculated correctly then what effect does the error in
pi have on the results? The Taylor series expansions of sin( x ) and cos( x ) are
x3
sin( x) = x − + · · · ≈ x for small x
6
x2
cos( x) = 1 − + · · · ≈ 1 for small x
2
Using the series approximations above we get
These are very close to the answers M ATLAB gives. See 4.1 for more on calculations with sin( x)
and cos( x).
Exercise 1. Work through the the function MachEps by hand for the floating point system F (2, 5, −6, +6).
Exercise 2. Modify the function MachEps so that it calculates realmin, the smallest floating point
number. This f.p. number defines the underflow threshold of the floating point system in use.
Exercise 3. Modify the function MachEps so that it calculates realmax, the largest floating point
number. This f.p. number defines the overflow threshold of the floating point system in use.
Exercise 4. The crucial statement in MyEps is if 1.0 + epsil > 1.0. This apparently works as
intended in M ATLAB, but it it is dangerous. If you use this in other programming languages
then the compiler may ‘optimize’ the statement to epsil > 0.0. Mathematically this is the
same as the original statement. Is it the same computationally? Check it out. Re-run your ma-
chine epsilon program with the ‘optimized’ statement and see what happens. General warning:
Problem 1. Write the M ATLAB functions NormS and NormCS in Section 3 above, and test them
on the Sea Surface Height data whose link is on the class webpage after the Introductory Lecture
slides.
Most of you did a good job on this exercise and your LATEX is improving. I urge those who still
balk at LATEX to persevere.
1. The crucial statement in MyEps is if 1.0 + epsil > 1.0. This apparently works as in-
tended in M ATLAB, but it is dangerous. If you use this in other programming languages
then the compiler may ‘optimize’ the statement to epsil > 0.0. Mathematically this is
the same as the original statement. Is it the same computationally? Check it out. Re-run
your machine epsilon program with the ‘optimized’ statement and see what happens.
General warning : Be very wary of optimizing compilers.
2. You need to check that Matlab’s results for eps, etc., agree with the floating point ‘theory’.
For example, does Matlab’s eps = b1− p ?
3. Indeterminates (1∞ etc.). The IEEE committee thought long and hard about these and
other floating point exceptions and finally agreed to what Prof W. Kahan told them to do
in the first place.
4. The a = 4/3; b = a-1; c = b+b+b; e = 1-c Problem. Most students did not explain the
result obtained by M ATLAB. Those who did try, got into a muddle. The problem is that
what you see (on the screen) is a decimal conversion of internal binary floating point
numbers and these are not the same.
5. Programming style need to be improved. Remember that you are, in general, writing
short pieces of mathematical software. Variable names need to be chosen that reflect
the problem at hand. Long, tedious, ‘stating-the-obvious’ comments obscure the code
(maybe this is your intention?)
6. Functions for calculating eps etc., should not have I/O statements. I/O statements should
be confined to special I/O functions, not strewn willy-nilly throughout a program.
7. Functions should do one job and do it well. A function that calculates eps, realmin,
realmax, and the the date of Easter Sunday in the year 2020, is not a good idea.
8. Those of you who are not using the MikTeX - WinEdt combination are going out of their
way to make life difficult for themselves. For example, how do you typeset 3 equations
whose ‘=’ signs are aligned? WinEdt has a template for this and it is chosen from a drop-
down menu.
9. When writing any report, you are trying to convince (‘con’ for short) the reader that you
understand what you are writing. In other words, write the report (or program com-
ments) for the reader, not yourself.
10. Test programs should state the version of M ATLAB that you are using. The following is
an example that writes out the version and date. Always use these where appropriate.
Remember that M ATLAB’s function disp([s1 s2 s3]) prints out strings. Hence all num-
bers must be converted to strings. M ATLAB should re-design all its I/O functions to be
easy and consistent, instead of a mixture of C and its own awkward I/O functions.
Throughout this and similar courses on numerical analysis we use relative error rather than
absolute error to measure the difference between a number x and an approximation to x. Let
x̂ = x + Ex be an approximation to x. Then2
| x̂ − x| E
Absolute Error Ex = | x̂ − x|, and Relative Error ex = = x, x 6= 0.
| x| | x|
To see why we use relative error, consider an algorithm A( x) that calculates x3 , using 3-digit
decimal arithmetic. For x = 1 we get A( x) = 0.333, the best we can do with 3-digit arithmetic.
For x = 100 we get A( x) = 33.3, the best we can do with 3-digit arithmetic. Now let us compare
the errors in these two calculations.
For x = 1 we have
1 333 1 999 − 1000 1 E1 1
E1 = 0.333 − = − = = , and e1 = 1
= .
3 1000 3 3000 3000 3
1000
We can see that E100 >> E1 which might lead the naïve to conclude that the algorithm A( x)
performs badly for larger values of x. This is obviously a false conclusion because A( x) is
doing the best it can with 3-digit arithmetic. The fact that the relative errors are equal shows
that A( x) performs properly in both cases.
None-the-less, relative error can be misleading when comparing real measurements, and dan-
gerously so when safety is involved. We now demonstrate this point with two examples.
In May 2006 I was asked to solve the following problem by Eamon Heffernan, Brusselstown
Window Company, Co. Wicklow.
Problem
Given a right-angled triangle with base 2873 mm and height 200 mm, find the angles and the
hypotenuse. The angle θ is between the base and the hypotenuse. Measurements are within
±5mm.
800 mm
θ
2873 mm
2 Generally the sign of the error in numerical analysis is not important, but in real applications it maybe very
important.
Solution.
3. Calculated the hypotenuse as 200/ sin(4◦ ) = 2867. Less than the base and obviously
wrong.
4. Calculated the hypotenuse as 2873/ cos (4◦ ) = 2880. Not obviously wrong, but is it accu-
rate?
√
5. Check with 2002 + 28732 = 2880. Seems ok.
θ3 θ5 θ7
sin θ = θ − + − + · · · ≈ θ for small θ
6 120 5040
θ2 θ4 θ6
cos θ = 1 − + − + · · · ≈ 1 for small θ
2 24 720
This shows that the calculation of cos θ is impervious to small changes about θ = 0, while sin θ
is not. The calculation 2873/ cos (4◦ ) = 2880 has virtually no error due to the rounded θ, but
200/ sin(4◦ ) = 2867 has a significant error (−13 mm) which is outside the ±5 mm measure-
ment tolerance.
Questions.
Because of the simplicity of this problem it was immediately obvious that the 200/ sin(4◦ ) =
2867 calculation was wrong.
• Would the error have been obvious to a computer program or, God forbid, a spreadsheet?
• Imagine if this had been a bridge design problem where the ‘mm’s become ‘m’s.
• Imagine if this had been an aircraft navigation problem where the ‘mm’s become ‘miles’.
An error of 13 miles is still 13 miles, even if the plane has flown 10,000 miles. 3
Moral.
Check your answers. Nature will anyway.
A High-Precision Solution.
Using PariGP 2.2.13 the following calculations were performed in 28-digit precision to show
how the the sin θ and cos θ solutions vary.
» atan(200/2873)
%1 = 0.06950151949217482078904047120
» 200/sin(%1)
%2 = 2879.952951004581949679215264x
» 2873/cos(%1)
%3 = 2879.952951004581949679215264
» sqrt(200ˆ2+2873ˆ2)
%4 = 2879.952951004581949679215264
» for(k = 1,10,print(2873/cos(0.0690+k/10000)));
» for(k = 1,10,print(200/sin(0.0690+k/10000)));
The difference in the lengths of the hypothenuse calculated by these angles is 13mm, which is
very significant: if the width of a window and the concrete wall opening differ by 13mm then
either the wall or the window or both have to be altered(cut). It is no use saying that 13mm is
just 100*13/2890 = 1/2% off. In physical measurements the absolute error is important.
For any θ in the range [0.05 − 0.09] radians, the value of the hypotenuse calculated by
On the morning of February 14th 1900, a cattle train of the Dublin Wicklow & Wexford Railway
Co. departed from Enniscorty at 10.a.m. bound for Harcourt Street Dublin, a distance of about
100 miles. Having loaded cattle at various stations along the way, it ‘arrived’ in Harcourt St.
station. This arrival is shown in Figure 3.
Figure 3: A Small Relative Error at Harcourt St Station, Dublin, February 14, 1900.
We can guess from the picture that the train overshot the platform by about 100 ft. Hence the
relative error was 100/(100 × 5280) ≈ 2 × 10−4 . This is quite a small relative error but, as the
picture shows, it has little meaning when measurements involve safety.
This is an interesting history of this crash, reproduced from http://www.har ourtstreettrain rash. om/
In 1900 the City of Dublin was served by 6 railway sloped downwards from Ranelagh and this meant
termini, which circled the City. Moving clockwise that locomotive drivers had to exercise care when
approaching the station. On the morning of Febru-
1. The Great Northern Railway Co., operated ary 14th 1900, Bray engine driver William Hyland
from Amiens Street Station (Connolly Sta- and his fireman Peter Jackson, also from Bray, de-
tion). parted from Enniscorty with 0-6-0 locomotive No
2. The London and North Western Railway 17, Wicklow built in 1899 by Dublin Wicklow and
Co., from the North Wall Station (the Point Wexford Railway Co., at 10.a.m. bound for Har-
Depot), closed April 1923. court Street Dublin, collecting wagons of cattle en
route for the Dublin City Market with the major-
3. The Dublin Wicklow and Wexford Rail- ity of them being collected at Arklow, Co.Wicklow,
way Co., from Westland Row (Pearse Sta- where a fair had taken place.
tion) which it leased from the Dublin and
Kingston Railway Co., and from Harcourt When the last wagon was coupled to the train, it
Street which closed in December 31st 1958. numbered 29 wagons, all of which were loose cou-
pled with braking being provided by the locomo-
4. The Great Southern Railway Co., operated tive and the guard’s van. This was the era before
from Kingsbridge Station (Hueston Station). air breaks were in use and locomotive drivers usu-
5. The Midland Great Western Railway Co. op- ally gave a series of steam whistle blasts if they
erated from Broadstone, which closed 16th wanted the guard to apply his brake in a hurry. All
January 1937 and is currently used as a bus went well until the train passed Ranelagh Station
depot. and began its approach to Harcourt Street Station
when Driver William Hyland applied the locomo-
6. The Northside of the Liffey was connected to
tive brakes to slow the train down for a gentle en-
the Southside via the City of Dublin Junction
try into the station after which the wagons would
Railway “The Loop Line” which had been
be shunted to the cattle bank for unloading. To the
opened in 1891 connecting Amiens Street
horror and amazement of Driver Hyland, the loco-
with Westland Row.
motive began to slide on the tracks, propelled in
part by the weight of the unbraked wagons behind
It was not until the turn of the century that lines
the locomotive.
had been laid down linking all Dublin termini to
facilitate the transfer of rolling stock from one ter- Passengers waiting in the station suddenly saw
minus to another. the cattle train enter the station sliding along the
track, at a walking pace, unstoppable, and travel
Railways first came to Dublin in 1834 with the
the full length of the platform before striking the
opening of the Dublin and Kingstown Railway
stationery buffers at the end of the platform, dis-
which ran from the present day Pearse Station to
lodging them. The locomotive then punched a
Salthill near the West Pier of Dun Laoghaire, then
hole in the station wall before coming to a stop
called Kingstown. In 1837 the line was extended to
in a raised position. A drayman passing down
the modern day Dun Laoghaire Station, to that part
Hatch Street had a narrow escape from the falling
of the Station where today Dun Laoghaire only
masonry which fell into the street below with his
DART trains and depart from.
dray sustaining minor damage following the im-
With all this rail traffic, there were no major pact of the locomotive with the rear wall of the sta-
mishaps at any of the Dublin termini until the tion. Fireman Peter Jackson, realizing that the train
evening of February 14th 1900 when a steam loco- was not going to stop as it entered the station and
motive went out of control and broke through the was going to collide with the buffers and end wall,
end wall of Harcourt Street Station, finishing up jumped off the locomotive footplate before the lo-
hanging up over Hatch Street where it remained in comotive struck the buffer bank and escaped in-
this precarious position until it was taken down. jury. The Driver William Hyland was not so for-
It is one if not the main event associated with the tunate as he remained in the locomotive cab right
Station. up to the moment of impact when he was thrown
Uniquely Harcourt Street Station which was clear. But his right arm became trapped between
opened in 1859 was at the foot of a gradient that some metal bars at the side of the locomotive coal
tender and a wall and was unable to move. At tailed examination of his injuries revealed most of
first those in the station though that the locomo- his right arm was severely damaged with much of
tive crew were trapped beneath the locomotive or the forearm being crushed. Amputation was seen
the debris surrounding it but this was found to be as the only option available to save his life, but
untrue. Driver Hyland refused to allow the operation to
proceed. A priest summed to administer the Last
All the cattle wagons remained upright and on the
rails. Immediately the stationmaster sent for med- Rights managed to persuade him that the opera-
tion was necessary to save his life. He urged him
ical assistance and also notified various officials
to follow the advice of Surgeon Ormsby, who ad-
of the company of what had happened. Medical
assistance in the form of Surgeon Ormbsy, Chief vocated the operation as the only option available,
and in the end Driver Hyland consented and the
Consultant Surgeon to the Dublin Wicklow and
Wexford Railway Company arrived on the scene operation then proceeded. During the operation,
and was joined by a Dr. McCausland. Company Surgeon Ormsbsy removed Driver Hyland’s right
arm just below the shoulder and also treated his
officials who arrived on the scene included Mr.
J.W. Pim. Chairman, Mr. Scallon, Deputy Chair- right foot, which had several broken bones in it.
Following the operation Driver Hyland rallied and
man, Mr. Coghlan, Traffic Manager, Mr. Shannon,
made good recovery.
Chief Engineer, Mr. Cronin, Locomotive Superin-
tendent, and Mr. Hickey, Traffic Superintendent. Back to the station, the wall had been shored up
Initially it was feared that the locomotive might ex- and made safe. The coal in the tender was un-
loaded and when this was done a crane was user to
plode due to the build up of steam pressure, but
lift up the tender and place it back on the rails and
the safety valve activated and allowed the excess
steam to blow off and once a breakdown crew ar- it was then moved of the station. In the case of the
locomotive, a special frame was built with which it
rived, they drew the fire, opened valves and re-
leased all the remaining steam. The locomotive was lowered to the ground and it was then placed
tender, full of coal, acted as a break keeping the on a set of temporary rails laid in Hatch Street and
brought along these to the station yard for exami-
locomotive in place. The cattle wagons were un-
coupled from the tender and drawn away from it nation and repair. A large crowd of spectators who
were kept behind a barrier witnessed this part of
by another locomotive leaving just the tender and
the operation.
raised locomotive in place.
A railway accident of this scale was subject to an
Word of the accident spread all over Dublin and
soon a very large crowd gathered in Hatch Street independent investigation by the Board of Trade
with the inquiry being carried out by Colonel von
but were kept away from the accident site by mem-
Donop. He formed the opinion that Driver Hyland
bers of the Dublin Metropolitan Police who cor-
doned off the street to spectators. In the station had miscalculated his speed while approaching
the station (steam locomotives at that time were
itself only bona fide passengers were allowed to
remain in case the locomotive exploded or further not fitted with speedometers) but was also criti-
quantities of masonry fell into the street. Later, cal of the station layout which prevented any train
directly entering the good’s yard. He suggested
the first task of the break down crew was to prop
up the locomotive in a fixed position using timber that pending this being remedied, all trains should
stop at the home signal or previous station. While
baulks.
the directors made all trains approaching Harcourt
While all activity was in progress, attempts were Street station stop at Ranelagh station, a procedure
made to extricate Driver Hyland from his position observed right up to the day the station closed, the
by the medical and engineering personnel there, issue of altering the station layout so as to make di-
and once the metal bars, which had trapped his rect access from the main line to good’s yard pos-
arm, were broken, he was freed. It was then pos- sible, was never remedied.
sible to carry out a medical examination on site,
When he recovered from his ordeal Driver William
which revealed that his forearm was nearly sev-
ered and that he was suffering from severe exhaus- Hyland returned to work with the company and
tion and shock, and required prompt hospitaliza- served until the 1930s as a goods checker in Bray.
The locomotive Wicklow was repaired and returned
tion if his life was to be saved.
to traffic, continuing in service until 1925 when it
The Dublin fire brigade and one of there ambu- was assigned the No, 440 by the Great Southern
lances arrived at the scene and Driver Hyland Railways and was withdrawn 1929.
was immediately rushed to the Meath Hospital for
treatment and surgery. There a through and de-