Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
748 views

Finite Word Length Effects in FFT Algorithm

Finite word length effects cause round-off errors in FFT algorithms. Quantization introduces errors when converting continuous signals to discrete, finite length representations. Product quantization error arises from multiplying finite-bit data and coefficients that produce outputs larger than the register size. For an 8-point DIT FFT algorithm, each stage forms a new array from the previous using basic operations, introducing round-off noise modeled as additive noise. The variance of output noise is proportional to the number of butterflies affecting each DFT computation node, which is N-1 for each of the N output nodes.

Uploaded by

anu
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
748 views

Finite Word Length Effects in FFT Algorithm

Finite word length effects cause round-off errors in FFT algorithms. Quantization introduces errors when converting continuous signals to discrete, finite length representations. Product quantization error arises from multiplying finite-bit data and coefficients that produce outputs larger than the register size. For an 8-point DIT FFT algorithm, each stage forms a new array from the previous using basic operations, introducing round-off noise modeled as additive noise. The variance of output noise is proportional to the number of butterflies affecting each DFT computation node, which is N-1 for each of the N output nodes.

Uploaded by

anu
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 12

FINITE WORD LENGTH

EFFECTS IN FFT
ALGORITHM :
ROUND OFF ERRORS
Finite Word Length Effect
• Most of the signals in real world are continuous.
• In DSP,

Continuous Signal Quantized Signal


Discrete Signal
(infinite length) (Finite Length)

 To convert into finite length we either truncate / round off the sequence
Product Quantization Error

Arise at the output of a multiplier

 Multiplication of b bit data and b bit coefficient will produce 2b bit output

 But only b bit registers are available, hence output should be truncated/ rounded off to b bits

 This produces an error


x[n]
x(t) Sampler Quantizer xq[n]

Quantization error,
e(n) = x[n] - xq[n]
Assume a sinusoidal signal between -1 and 1

If ADC is used to convert, it employs (b+1) bits, including the sign bit

 Then number of levels available for quantizing x(n) is

Thus interval between two successive levels are

, q -> Quantization step size


Let us consider the computation of DFT using Radix-2 DIT FFT algorithm for N=8:
The DFT is computed in

at each stage a new array of N numbers are formed from the previous array by using the
basic operations of DIT FFT algorithm,

, here m and (m+1) refers to mth and (m+1)th array


p and q represent location of number in each array
Here product of causes ROUND-OFF ERROR

 Therefore, we shall model the round off noise by associating an additive noise generator

 Here e(m,q) represents a complex error introduced in multiplication.


We make the following assumptions:
1. Round off noise due to each multiplication is uniformly distributed in amplitude
between -2-b /2 to 2-b /2
2. All noise sources are uncorrelated to each other
3. All noise sources are uncorrelated with input and output
As input data and twiddle factor are complex, 4 real multiplications are associated for each
complex multiplication.
 Variance of the Round off Error,

To calculate the variance of any output noise at any output node, we consider butterflies
that affect the computation of dft at that node.

 Eg – the output node X(1) is connected to N-1 = 7 butterflies.

i.e, 7 noise sources affect dft computation of X(1)


This is applicable for all Output nodes

 Hence, (N-1) noise sources propagate to each output node, which results in an
output noise variance given by,

For large values of N,

 This is the same result that we obtained from direct computation of the DFT
Thank you

You might also like