Digital Filter Banks • The digital filter bank is set of bandpass filters with either a common input or a summed output • An M-band analysis filter bank is shown below

1

Copyright © 2001, S. K. Mitra

Digital Filter Banks • The subfilters H k (z ) in the analysis filter bank are known as analysis filters • The analysis filter bank is used to decompose the input signal x[n] into a set of subband signals vk [n] with each subband signal occupying a portion of the original frequency band 2

Copyright © 2001, S. K. Mitra

Digital Filter Banks • An L-band synthesis filter bank is shown below

• It performs the dual operation to that of the analysis filter bank 3

Copyright © 2001, S. K. Mitra

Digital Filter Banks • The subfilters Fk (z ) in the synthesis filter bank are known as synthesis filters • The synthesis filter bank is used to combine a set of subband signals v^k [n] (typically belonging to contiguous frequency bands) into one signal y[n] at its output

4

Copyright © 2001, S. K. Mitra

Uniform Digital Filter Banks • A simple technique to design a class of filter banks with equal passband widths is outlined next • Let H 0 ( z ) represent a causal lowpass digital filter with a real impulse response h0[n] : ∞ H 0 ( z ) = ∑n = −∞ h0[n]z − n • The filter H 0 ( z ) is assumed to be an IIR filter without any loss of generality 5

Copyright © 2001, S. K. Mitra

Uniform Digital Filter Banks • Assume that H 0 ( z ) has its passband edge ωp and stopband edge ωs around π/M, where M is some arbitrary integer, as indicated below

0

ωp π 6

ωs

π

2π

ω

M Copyright © 2001, S. K. Mitra

Uniform Digital Filter Banks • Now, consider the transfer function H k (z ) whose impulse response hk [n] is given by hk [n] = h0[n] e j 2 πk n/ M = h0[n]WM− kn , 0 ≤ k ≤ M −1 where we have used the notation WM = e − j 2π / M • Thus, −n ∞ ∞ k − n H k ( z ) = ∑n = −∞ hk [n]z = ∑n = −∞ h 0[n] zWM ,

(

7

)

0 ≤ k ≤ M −1

Copyright © 2001, S. K. Mitra

Uniform Digital Filter Banks • i.e., H k ( z ) = H 0 ( zWMk ), 0 ≤ k ≤ M − 1

• The corresponding frequency response is given by H k (e jω ) = H 0 (e j (ω− 2 π k /M ) ), 0 ≤ k ≤ M − 1 • Thus, the frequency response of H k ( z ) is obtained by shifting the response of H 0 ( z ) to the right by an amount 2πk/M 8

Copyright © 2001, S. K. Mitra

Uniform Digital Filter Banks • The responses of H k ( z ) , H k ( z ) , . . . , H k ( z ) are shown below

9

Copyright © 2001, S. K. Mitra

Uniform Digital Filter Banks • Note: The impulse responses hk [n] are, in general complex, and hence | H k (e jω )| does not necessarily exhibit symmetry with respect to ω = 0 • The responses shown in the figure of the previous slide can be seen to be uniformly shifted version of the response of the basic prototype filter H 0 ( z ) 10

Copyright © 2001, S. K. Mitra

Uniform Digital Filter Banks • The M filters defined by H k ( z ) = H 0 ( zWMk ), 0 ≤ k ≤ M − 1 could be used as the analysis filters in the analysis filter bank or as the synthesis filters in the synthesis filter bank • Since the magnitude responses of all M filters are uniformly shifted version of that of the prototype filter, the filter bank obtained is called a uniform filter bank 11

Copyright © 2001, S. K. Mitra

Uniform DFT Filter Banks Polyphase Implementation • Let the prototype lowpass transfer function be represented in its M-band polyphase form: M −1 −l H 0 ( z ) = ∑l =0 z El ( z M )

where El ( z ) is the l-th polyphase component of H 0 ( z ): El ( z ) = ∑ 12

∞ −n e [ n ] z n =0 l

=∑

∞ −n h [ l + nM ] z , n =0 0

0 ≤ l ≤ M −1

Copyright © 2001, S. K. Mitra

Uniform DFT Filter Banks • Substituting z with zWMk in the expression for H 0 ( z ) we arrive at the M-band polyphase decomposition of H k ( z ): M −1 −l − kl M kM z W E ( z WM ) M l l =0 M −1 −l − kl M z W E ( z ), 0 ≤ k M l l =0

H k ( z) = ∑ =∑

13

≤ M −1

• In deriving the last expression we have used the identity WMkM = 1 Copyright © 2001, S. K. Mitra

Uniform DFT Filter Banks • The equation on the previous slide can be written in matrix form as H k ( z ) = [1

−k WM

−2 k WM

E0 ( z M )    z −1E ( z M )  .... WM−( M −1) k ]  −2 1 M  z E2 ( z )   .. .   −( M −1) EM −1 ( z M )   z

0≤ k ≤ M − 1 14

Copyright © 2001, S. K. Mitra

Uniform DFT Filter Banks • All M equations on the previous slide can be combined into one matrix equation as M E ( z )  0  1 1 1 ... 1  H0 ( z)   −1 M −1 −2 −( M −1)   ... 1 W W W ( ) z E z  H ( z)   M M M 1  −4 − 2 ( M −1)    H 1 ( z )  = 1 W − 2 −2 M ... W W M M z E ( z ) 2 M 2  .  . . . .   .. .. .. ... ..  ..  ..  .  H M −1 ( z )   −( M −1) − 2 ( M −1)... −( M −1) 2   1 WM

WM

WM

     M   z −( M −1) E ( z )  M −1

−1

MD

• In the above D is the M × M DFT matrix 15

Copyright © 2001, S. K. Mitra

Uniform DFT Filter Banks • An efficient implementation of the M-band uniform analysis filter bank, more commonly known as the uniform DFT analysis filter bank, is then as shown below

16

Copyright © 2001, S. K. Mitra

Uniform DFT Filter Banks • The computational complexity of an M-band uniform DFT filter bank is much smaller than that of a direct implementation as shown below

17

Copyright © 2001, S. K. Mitra

Uniform DFT Filter Banks • For example, an M-band uniform DFT analysis filter bank based on an N-tap prototype lowpass filter requires a total of M log M + N multipliers 2 2 • On the other hand, a direct implementation requires NM multipliers

18

Copyright © 2001, S. K. Mitra

Uniform DFT Filter Banks • Following a similar development, we can derive the structure for a uniform DFT synthesis filter bank as shown below

19

Type I uniform DFT synthesis filter bank

Type II uniform DFT synthesis filter bank Copyright © 2001, S. K. Mitra

Uniform DFT Filter Banks M

• Now Ei ( z ) can be expressed in terms of E0 ( z M )   H 0( z )    z −1E ( z M )  H 1( z )   1  = 1 D H2 ( z)  −2 M z E2 ( z )  .  M   .. ..   .   H M −1 ( z )  −( M −1) M EM −1 ( z )  z

• The above equation can be used to determine the polyphase components of an IIR transfer function H 0 ( z ) 20

Copyright © 2001, S. K. Mitra

Nyquist Filtrs • Under certain conditions, a lowpass filter can be designed to have a number of zerovalued coefficients • When used as interpolation filters these filters preserve the nonzero samples of the up-sampler output at the interpolator output • Moreover, due to the presence of these zero-valued coefficients, these filters are computationally more efficient than other lowpass filters of same order 21

Copyright © 2001, S. K. Mitra

Lth-Band Filters • These filters, called the Nyquist filters or Lth-band filters, are often used in single-rate and multi-rate signal processing • Consider the factor-of-L interpolator shown below x[n ]

L

xu [n]

H (z)

y[n ]

• The input-output relation of the interpolator in the z-domain is given by Y ( z) = H ( z) X ( z L ) 22

Copyright © 2001, S. K. Mitra

Lth-Band Filters • If H(z) is realized in the L-band polyphase form, then we have L −1 −i H ( z ) = ∑i =0 z Ei ( z L ) • Assume that the k-th polyphase component of H(z) is a constant, i.e., Ek ( z ) = α:

H ( z ) = E0 ( z L ) + z −1E1 ( z L ) + ... + z −( k −1) Ek −1 ( z L )

+ z −( k +1) Ek +1 ( z L ) + ... + z −( L −1) EL −1 ( z L ) 23

Copyright © 2001, S. K. Mitra

Lth-Band Filters • Then we can express Y(z) as

Y ( z ) = α z −k X ( z L ) + • As a result,

L −1

−l E ( z L ) X ( z L ) z ∑ l

l =0 l≠k

y[ Ln + k ] = α x[ n]

• Thus, the input samples appear at the output without any distortion for all values of n, whereas, in-between ( L − 1) output samples are determined by interpolation 24

Copyright © 2001, S. K. Mitra

Lth-Band Filters

25

• A filter with the above property is called a Nyquist filter or an Lth-band filter • Its impulse response has many zero-valued samples, making it computationally attractive • For example, the impulse response of an Lth-band filter for k = 0 satisfies the following condition α, n=0  h[Ln] =   0, otherwise Copyright © 2001, S. K. Mitra

Lth-Band Filters • Figure below shows a typical impulse response of a third-band filter (L = 3)

• Lth-band filters can be either FIR or IIR filters 26

Copyright © 2001, S. K. Mitra

Lth-Band Filters • If the 0-th polyphase component of H(z) is a constant, i.e., E0 ( z ) = α then it can be shown that L −1 k ) = Lα = 1 (assuming α = 1/L) H ( zW ∑ k =0 L • Since the frequency response of H ( zWLk ) is the shifted version H (e j (ω− 2 πk / L ) ) of H (e jω) , the sum of all of these L uniformly shifted versions of H (e jω) add up to a constant 27

Copyright © 2001, S. K. Mitra

Half-Band Filters • An Lth-band filter for L = 2 is called a halfband filter • The transfer function of a half-band filter is thus given by H ( z ) = α + z −1E1 ( z 2 ) with its impulse response satisfying α, n=0  h[2n] =   0, otherwise 28

Copyright © 2001, S. K. Mitra

Half-Band Filters • The condition H ( z ) = α + z −1E1 ( z 2 ) reduces to H ( z ) + H (− z ) = 1 (assuming α = 0.5) • If H(z) has real coefficients, then H (−e jω ) = H (e j ( π−ω) )

• Hence 29

H (e jω ) + H (e j ( π−ω) ) = 1 Copyright © 2001, S. K. Mitra

Half-Band Filters H (e j ( π / 2−θ) ) and H (e j ( π / 2+ θ) ) add up to 1 for all θ • Or, in other words, H (e jω) exhibits a symmetry with respect to the half-band frequency π/2, hence the name “half-band filter”

•

30

Copyright © 2001, S. K. Mitra

Half-Band Filters • Figure below illustrates this symmetry for a half-band lowpass filter for which passband and stopband ripples are equal, i.e., δ p = δs and passband and stopband edges are symmetric with respect to π/2, i.e., ω p + ωs = π

31

Copyright © 2001, S. K. Mitra

Half-Band Filters • Attractive property: About 50% of the coefficients of h[n] are zero • This reduces the number of multiplications required in its implementation significantly • For example, if N = 101, an arbitrary Type 1 FIR transfer function requires about 50 multipliers, whereas, a Type 1 half-band filter requires only about 25 multipliers 32

Copyright © 2001, S. K. Mitra

Half-Band Filters • An FIR half-band filter can be designed with linear phase • However, there is a constraint on its length • Consider a zero-phase half-band FIR filter for which h[n] = α * h[− n] , with | α | = 1 • Let the highest nonzero coefficient be h[R]

33

Copyright © 2001, S. K. Mitra

Half-Band Filters • Then R is odd as a result of the condition α, n=0  h[2n] =   0, otherwise • Therefore R = 2K+1 for some integer K • Thus the length of h[n] is restricted to be of the form 2R+1 = 4K+3 [unless H(z) is a constant] 34

Copyright © 2001, S. K. Mitra

Design of Linear-Phase Lth-Band Filters

35

• A lowpass linear-phase Lth-band FIR filter can be readily designed via the windowed Fourier series approach • In this approach, the impulse response coefficients of the lowpass filter are chosen as h[n] = hLP [n] ⋅ w[n] where hLP [n] is the impulse response of an ideal lowpass filter with a cutoff at π/L and w[n] is a suitable window function Copyright © 2001, S. K. Mitra

Design of Linear-Phase Lth-Band Filters • Now, the impulse response of an ideal Lthband lowpass filter with a cutoff at ωc = π / L is given by sin( πn / L) hLP [n] = , −∞ ≤ n ≤∞ πn • It can be seen from the above that hLP [n] = 0 for n = ± L, ± 2 L, ... 36

Copyright © 2001, S. K. Mitra

Design of Linear-Phase Lth-Band Filters • Hence, the coefficient condition of the Lthband filter α, n=0  h[Ln] =   0, otherwise

37

is indeed satisfied • Hence, an Lth-band FIR filter can be designed by applying a suitable window w[n] to hLP [n] Copyright © 2001, S. K. Mitra

Design of Linear-Phase Lth-Band Filters • There are many other candidates for Lthband FIR filters • Program 10_8 can be used to design an Lthband FIR filter using the windowed Fourier series approach • The program employs the Hamming window • However, other windows can also be used 38

Copyright © 2001, S. K. Mitra

Design of Linear-Phase Lth-Band Filters • Figure below shows the gain response of a half-band filter of length-23 designed using Program 10_8 0

Gain, dB

-20 -40 -60 -80 0

0.2

0.4

0.6

0.8

1

ω/π

39

Copyright © 2001, S. K. Mitra

Design of Linear-Phase Lth-Band Filters • The filter coefficients are given by h[ −11] = h[11] = − 0.002315; h[ −10] = h[10] = 0; h[ −9] = h[9] = 0.005412; h[ −8] = h[8] = 0; h[ −7 ] = h[ 7 ] = − 0.001586; h[ −6] = h[ 6] = 0; h[ −5] = h[5] = 0.003584; h[ −4] = h[ 4] = 0; h[ −3] = h[3] = − 0.089258; h[ −2] = h[ 2] = 0; h[ −1] = h[1] = 0.3122379; h[ 0] = 0.5;

40

• As expected, h[n] = 0 for n = ± 2, ± 4, ± 6, ± 8, ±10 Copyright © 2001, S. K. Mitra