EEO 401

Digital Signal Processing

Prof. Mark Fowler

Note Set #27

• General Issues for Design of Digital Filters
• Reading: Sect. 10.1 & 10.2.1 of Proakis & Manolakis
Ideal D-T Filters
Here is the so-called “ideal filter” frequency response for a Lowpass Filter
Cut-off frequency = B rad/sample
Ideal Lowpass Filter (LPF)
| H f (ω ) |
1

− 4π − 3π − 2π −π -B B π 2π 3π 4π ω

Linear Phase
∠H (ω )
f

− 4π − 3π − 2π −π -B B π 2π 3π 4π
As always with DT… only need to look here

There are also highpass, bandpass, and bandstop filters;

and there are other more specialized types
2/10
 Why can’t an ideal filter exist in practice?? Same as for CT…. It is non-causal

=
For the ideal LPF H (ω )
f
p2 B (ω )e − jωnd ω ∈ [−π , π ) repeats elsewhere
Now consider applying a delta function as its input: x[n] = δ[n] ↔ Xf(ω) = 1

Then the output has DTFT:

Y f (ω ) X =
= f
(ω ) H f (ω ) p2 B (ω )e − jωnd ω ∈ [ −π , π ) repeats elsewhere
Linear Phase Imparts Delay

∞
From the DTFT Table: B
π sinc [ πB n ] ∑p
k = −∞
2B (ω + 2π k )

=
So the response to a delta (applied at n = 0) is: h[n ] ( B / π ) sinc [ ( B / π )( n − nd ) ]

h[n ]
x[n] = δ[n] Ideal
LPF … …
n nd n
Starts before input starts…
Thus, system is non-causal! 3/10
Paley-Wiener Theorem
So we’ve seen that a causal filter can’t give that perfect rectangular frequency
response of the ideal filter.
But…. What can it give??? The Paley-Wiener Theorem answers that…

If h[n] has finite energy and h[n] = 0 ∀n < 0 then

∫π ( )
π
ln H f (ω ) d ω < ∞
−

Conversely, if |Hf(ω)| is square integrable and satisfies this… then

there exists a phase θ(ω) so that

H f (ω ) = H f (ω ) e jθ (ω )

has an IDTFT h[n] that is causal.

4/10
So what does the Paley-Wiener Theorem actually say about what shapes we
can have for frequency response of a causal filter??

∫π ( )
π
ln H f (ω ) d ω < ∞
−

Clearly if |Hf(ω)| = 0 on any interval then |ln(|Hf (ω)|)| is infinite on that interval
 So can’t have a perfect stop band
But… it is possible for |Hf (ω)| = 0 at finitely many points… just not over intervals.

Rational Transfer Function

Just because the P-W Theorem says that there is a causal filter that exists for
some desired |H(ω)| does not mean it has a rational transfer function.
Recall that our focus is on DT systems with rational transfer functions:
 Standard-Form Difference Eq M M

 Poles & Zeros

∑ bk e − jωk
∑ bk z − k
H f (ω ) = k =0N H z ( z ) = k =0N
 Block Diagram
∑
1 + a e − jωk k ∑
1 + a z −k
k =1
k
k =1

Our Goal: Find rational approximate to desired Hfd(ω) 5/10

Practical Filter Specification
Lowpass Filter Specification
|H(Ω)|
1 − δ p ≤ H (Ω) ≤ 1 + δ p
1 + δp 0 ≤ H (Ω ) ≤ δ s
1 – δp

Passband Ripple = 20 log10 (1 + δ p ) dB

Stopband Attenuation = −20 log10 (δ s ) dB

To make filter “more ideal”:

δp→ 0, δs→ 0, Ωs → Ωp

δs

Ωp Ωs Ω

Pass-band Transition Stop-band

6/10
band
Highpass Filter Specification
To make filter “more ideal”: δp→ 0, δs→ 0, Ωs → Ωp
|H(Ω)|

1 + δp
1 – δp

Passband Ripple = 20 log10 (1 + δ p ) dB

Stopband Attenuation = −20 log10 (δ s ) dB

δs

Ωs Ωp π Ω

Stop-band Transition Pass-band

band 7/10
Bandpass Filter Specification
To make filter “more ideal”: δp→ 0, δsi→ 0, Ωsi → Ωpi
|H(Ω)|

1 + δp
1 – δp

δs2
δs1

Ωs1 Ωp1 Ωp2 Ωs2 π Ω

Stop Trans Pass-band Trans Stop-band

band band band 8/10
Bandstop Filter Specification
To make filter “more ideal”: δp→ 0, δsi→ 0, Ωsi → Ωpi
|H(Ω)|
1 + δp2
1 + δp1
1 – δp1
1 – δp2

δs

Ωp1 Ωs1 Ωs2 Ωp2 π Ω

Pass Trans Stop-band Trans Pass-band

band band band 9/10
Our Goal: Find rational approximate to desired Hfd(ω)

Some Issues to Consider:

• Achieve the specs (band edges and ripples) with the least
complexity
 Complexity does have links to the “delay” through the
filter
• FIR vs IIR filters
 FIR can achieve the desirable linear phase response
 IIR can achieve magnitude specs with lower complexity
• Recursive vs. Non-Recursive
 Connections to FIR vs IIR
 Impact on complexity
• Stability of filter
 FIR filters are inherently stable
 IIR filter design needs attention to stability
10/10