Filter Design Techniques

• Filter
– Filter is a system that passes certain
frequency components and totally
rejects all others
• Stages of the design filter
– Specification of the desired properties
of the system
– Approximation of the specification using
a causal discrete-time system
– Realization of the system
 Review of discrete-
time systems
Frequency response :
• periodic : period = 2π
• for a real impulse response h[k]
Magnitude response H ( e ) is even function
jω

Phase response ∠H (e jω ) is odd function

• example :

1
Nyquist frequency
e jπk = ...,1,−1,1,−1,1,...
0 .5

−π π
0
-4 -2 0 2 4

-5
-4 -2 0 2 4
 Review of discrete-
time systems
`Popular’ frequency responses for filter design :
low-pass (LP) high-pass (HP) band-pass (BP)

π π π
band-stop multi-band …

π π π
 Review of discrete-
time systems
“FIR filters” (finite impulse response):
B(z) −1 −N
H (z) = N
= b 0 + b1 z + ... + b N z
z
• “Moving average filters” (MA filters)
• N poles at the origin z=0 (hence guaranteed stability)
• N zeros (zeros of B(z)), “all zero” filters
• corresponds to difference equation
y [ k ] = b 0 u [ k ] + b1 u [ k − 1 ] + ... + b N u [ k − N ]
• impulse response
h [ 0 ] = b 0 , h [1] = b1 ,..., h [ N ] = b N , h [ N + 1] = 0 ,...
 Linear Phase FIR Filters
Non-causal zero-phase filters :
example: symmetric impulse response
h[-L],….h[-1],h[0],h[1],...,h[L]
h[k]=h[-k], k=1..L

k
frequency response is L
+L L
jω
H (e ) = ∑ h[k ]e
k =− L
− jω . k
= ... = ∑ ak cos(ωk )
k =0

- i.e. real-valued (=zero-phase) transfer function

- causal implementation by introducing (group) delay
 Linear Phase FIR Filters
• Causal linear-phase filters = non-causal zero-phase +
delay
example: symmetric impulse response & N even
h[0],h[1],….,h[N]
N=2L (even)
h[k]=h[N-k], k=0..L k
0 N
frequency response is
N L
H (e ) = ∑ h[k ]e
jω − jω . k
= ... = e − jω L
∑a k cos(ωk )
k =0 k =0

= i.e. causal implementation of zero-phase filter, by

introducing (group) delay z − L = e − jωL
z = e jω
 Linear Phase FIR Filters
Type-1 Type-2 Type-3 Type-4
N=2L=even N=2L+1=odd N=2L=even N=2L+1=odd
symmetric symmetric anti-symmetric anti-symmetric
h[k]=h[N-k] h[k]=h[N-k] h[k]=-h[N-k] h[k]=-h[N-k]
L ω L
cos( )∑ ak cos(ωk ) je− jωN / 2 sin(ω) a cos(ωk)
L−1
∑ ak cos(ωk ) e
− jωN / 2 − jωN / 2
ω L
e
k =0
2 k =0 ∑k
k =0
j.e − jωN / 2
sin( )∑ ak cos(ωk )
2k =0

zero at ω =π zero at ω = 0, π zero at ω =0

LP/HP/BP LP/BP HP
 Linear Phase FIR Filters
u[k]
• efficient direct-form realization. Δ Δ Δ Δ
example: Δ Δ Δ Δ

+ + + + +

bo b1 b2 b3 b4
x x x x x

y[k] + + + +
• PS: IIR filters can NEVER have linear-phase property !
 Filter Specification
Ex: Low-pass 1.2
Passband Ripple
1

1−δP 0.8
ωP ωS
0.6 Passband Cutoff -> <- Stopband Cutoff

0.4 Stopband Ripple δS

0.2

0
0 0.5 1 1.5 2 2.5 3
 Filter Design Problem
• Design of filters is a problem of
function approximation

• For FIR filter, it implies polynomial

approximation
• For IIR filter, it implies
approximation by a rational function
of z
 Filter Design by Optimization
(I) Weighted Least Squares Design :
• select one of the basic forms that yield linear phase
e.g. Type-1
L
jω − jω N / 2
H (e ) = e ∑a
k =0
− jω N / 2
k cos(ωk ) = e A(ω )
• specify desired frequency response (LP,HP,BP,…)
H d (ω ) = e − jωN / 2 Ad (ω )
• optimization criterion is
+π +π

a0 ,..., a L ∫
) − H d (ω ) dω = min ∫ W (ω ) A(ω ) − Ad (ω ) dω
jω 2
ω
2
min W ( ) H ( e
a0 ,..., a L
−π −π
1 4444244443
F ( a0 ,..., a L )

where W (ω ) ≥ 0 is a weighting function

 Filter Design by Optimization
F ( a 0 ,..., a L )
644 47444 8
• …this is equivalent to min{ x T Qx − 2 x T p + μ }
x

x T = [a0 a1 ... a L ]
π
Q = ∫ W (ω )c (ω )c T (ω ) dω
0
π
p = ∫ W (ω ) Ad (ω )c (ω ) dω
0

c T (ω ) = [1 cos(ω ) ... cos( Lω ) ]

μ = ...
=standard ‘Quadratic Optimization’ problem
xOPT = Q−1 p
 Filter Design by
Optimization 1.2

• Example: Low-pass design 1

Passband Ripple

Ad (ω ) = 1, ω < ω P
0.8
(pass - band)
Passband Cutoff -> <- Stopband Cutoff
Ad (ω ) = 0, ω S ≤ ω ≤ π
0.6
(stop - band)
0.4 Stopband Ripple

optimization function is 0.2

ωP +π 0
0 0.5 1 1.5 2 2.5 3
F ( a0 ,..., aL ) = ∫ A(ω ) − 1 dω + γ . ∫ A (ω )dω = ...
2 2

1
0
442443 1ω4 S
4244 3
pass - band stop - band
i.e.
W (ω ) = ...
Filter Design by Optimization
• a simpler problem is obtained by replacing the F(..) by…
2
⎧ ⎡ a0 ⎤ ⎫
⎪ ⎪
F (a0 ,..., aL ) = ∑ W (ωi ) A(ωi ) − Ad (ωi ) = ∑ W (ωi )⎨cT (ωi ) ⎢⎢ : ⎥⎥ − Ad (ωi )⎬
2

i i ⎪ ⎢⎣aL ⎥⎦ ⎪
⎩ ⎭
where the wi’s are a set of n sample frequencies
The quadratic optimization problem is then equivalent to a least-squares
problem

min Ax − b = min { x T x − 2 xT {
A b + b{
2 T T T
2
A A
{ b}
x x
∑ W (ω i ) c (ω i ) c T (ω i ) ∑ ... ∑ ...
i i i

xLS = (A A)−1 A b
T T Compare to p.12
+++ : simple
--- : unpredictable behavior in between sample frequencies.
 Filter Design by Optimization
• …then all this is often supplemented with
additional constraints

Example: Low-pass (LP) design (continued)

pass-band ripple control :
A(ω ) − 1 ≤ δ P , ω < ω P (δ P is pass - band ripple)

stop-band ripple control :

A(ω ) ≤ δ S , ω S ≤ ω ≤ π (δ S is stop - band ripple)
 Filter Design by Optimization
Example: Low-pass (LP) design (continued)
a realistic way to implement these constraints, is to impose
the constraints (only) on a set of sample frequencies
ω P1 , ω P 2 ,..., ω Pm in the pass-band
and ω S 1 , ω S 2 ,..., ω Sn in the stop-band
The resulting optimization problem is :
minimize : F ( a ,..., a ) = ...
0 L

x T = [a0 a1 ... aL ]
subject to AP x ≤ bP (pass-band constraints)
AS x ≤ bS (stop-band constraints)
= `Quadratic Linear Programming’ problem
 Filter Design by Optimization
(II) `Minimax’ Design :
• select one of the basic forms that yield linear phase
e.g. Type-1 H (e jω ) = e − jωN / 2 a[ k ] cos(ωk ) = e − jωN / 2 A(ω )
L

∑
k =0

• specify desired frequency response (LP,HP,BP,…)

H d (ω ) = e − jωN / 2 Ad (ω )
• optimization criterion is
min max W (ω ) H (e jω ) − H d (ω ) = min max W (ω ) A(ω ) − Ad (ω )
a0 ,...,a L 0≤ω ≤π a0 ,...,a L 0≤ω ≤π

where W (ω ) ≥ 0 is a weighting function

 Filter Design by Optimization
• Conclusion:
(I) weighted least squares design
(II) minimax design
provide general `framework’, procedures to
translate filter design problems into standard
optimization problems
• In practice (and in textbooks):
emphasis on specific (ad-hoc) procedures :
- filter design based on ‘windows’
- equi-ripple design
 Filter Design using ‘Windows’
Example : Low-pass filter design
• ideal low-pass filter is
⎧1 ω < ωC
H d (ω ) = ⎨
⎩0 ω C < ω < π
• hence ideal time-domain impulse response is π
+π
1 jω sin(ωc k )
∫π H
jω
hd [k ] = (e )e .k
dω = ... = α
2π ωc k
d
−

• truncate hd[k] to N+1 samples : Non-causal and infinitely long

⎧ hd [ k ] − N /2 < k < N /2
h[ k ] = ⎨
⎩ 0 otherwise
• add (group) delay to turn into causal filter
 Filter Design using ‘Windows’
Example : Low-pass filter design (continued)
• note : it can be shown that time-domain truncation corresponds to
solving a weighted least-squares optimization problem with the
given Hd, and weighting function W (ω ) = 1

• truncation corresponds to applying a ‘rectangular window’ :

h[k ] = hd [k ]w[k ]
⎧1 − N /2 < k < N /2
w[ k ] = ⎨
⎩0 otherwise
• +++: simple procedure (also for HP,BP,…)
• --- : truncation in the time-domain results in ‘Gibbs effect’ in the
frequency domain, i.e. large ripple in pass-band and stop-band,
which cannot be reduced by increasing the filter order N.
 Filter Design using ‘Windows’
Remedy : apply windows other than rectangular window:
• time-domain multiplication with a window function w[k]
corresponds to frequency domain convolution with W(z) :

h[k ] = hd [k ]w[k ]
H ( z ) = H d ( z ) ∗W ( z )
• candidate windows : Han, Hamming, Blackman, Kaiser,…. (see
textbooks)
• window choice/design = trade-off between side-lobe levels
(define peak pass-/stop-band ripple) and width main-lobe
(defines transition bandwidth)
Windowing Effect

Gibbs phenomenon
Windowing
 Equiripple Design
• Starting point is minimax criterion, e.g.
min max W (ω ) A(ω ) − Ad (ω ) = min max E (ω )
a0 ,..., a L 0 ≤ω ≤π a0 ,..., a L 0 ≤ω ≤π

• Based on theory of Chebyshev approximation and the ‘alternation

theorem’, which (roughly) states that the optimal ai’s are such that
the ‘max’ (maximum weighted approximation error) is obtained at
L+2 extremal frequencies…

max E (ω ) = E (ωi ) for i = 1,.., L + 2

0 ≤ω ≤π

…that hence will exhibit the same maximum ripple (‘equiripple’)

• Iterative procedure for computing extremal frequencies, etc.
(Remez exchange algorithm, Parks-McClellan algorithm)
• Very flexible, etc., available in many software packages
• Details omitted here (see textbooks)
 Software
• FIR Filter design abundantly available in
commercial software
• Matlab:
b=fir1(n,Wn,type,window), windowed linear-phase FIR design,
n is filter order, Wn defines band-edges, type is
`high’,`stop’,…
b=fir2(n,f,m,window), windowed FIR design based on inverse
Fourier transform with frequency points f and
corresponding magnitude response m
b=remez(n,f,m), equiripple linear-phase FIR design with
Parks-McClellan (Remez exchange) algorithm