Use of the DFT In linear filtering The product of two DFTs is equivalent to the circular

convolution of the corresponding time-domain sequences.

Circular convolution is of no use if our objective is to

determine the o/p of a linear filter to a given i/p sequence.

Block Convolution

So we seek a frequency domain methodology equivalent

to a linear convolution.

Let finite duration sequence x(n) of length L excite an FIR Let

filter of length M. Ln<0 x(n)=0, for M n<0 h(n)=0 for The o/p sequence y(n) of the FIR filter can be expressed as in the time domain as linear convolution of these two.

Filtering of Long Data Sequences

The result of Linear convolution will be exactly equal to circular convolution if the sequences x(n) and h(n) are extended to a length equal to L+M-1 by zero padding. ie, x(n) is extended to a length of L+M-1, by padding M-1 zeros and h(n) is extended to a length of L+M-1, by paddding L-1 zeros. Thus the response of a linear filter can be computed using circular convolution and in turn by using DFT if the above method follows. In practical applications, involving linear filtering of signals, the i/p sequence x(n) is often a very long sequence. This is especially true in some real time signal processing applications (with signal monitoring and Analysis) Since linear filtering performed via the DFT, involves operations on a block of data, which by necessity must be limited in size due to limited memory of a digital computer, a long i/p signal sequence must be segmented to fixed sized blocks prior to processing. Since the filtering is linear, successive blocks may be processed one at a time via the DFT and the o/p blocks are fitted together to form the overall o/p signal sequence.

This overall o/p sequence is identical to the sequence obtained if the long block had been processed via timedomain convolution.

Overlap-save method:
L Input Data
M-1 x1 (n) M-1

L . .

L L
X2 (n)

Two methods for FIR filtering long data sequence are 1) Overlap-save method and 2) Overlap-add method. For both method we assume that FIR filter has duration M.
Out put Signal

h (n)

L-1

M-1

X3 (n) M-1 M-1

X4(n) y1 (n) Y2 (n)

The i/p data sequence is segmented into blocks of L points, where by assumption, L>>M without loss of generality.

Discard M-1 points.

Y3 (n)

Discard M-1 points.

Y4 (n)

Discard M-1 points.

Discard M-1 points.

Overlap-add Method:
L Input Data
M-1 zeros
x1 (n)

M-1 zeros

L
X2 (n)

M-1 zeros
h (n) X3 (n)

M-1 zeros

X4(n)

Out put Signal

y1 (n) Y2 (n)

M-1 points add together M-1 points add together

Y3 (n) Y4 (n)

M-1 points add together