FIR Digital Filter Design

MUS421/EE367B Lecture 6
FIR Digital Filter Design In a previous lecture, we looked at many windows, and
examined their properties. Today, we are going to see
Julius O. Smith III (jos@ccrma.stanford.edu) how these windows can be used to design Finite Impulse
Center for Computer Research in Music and Acoustics (CCRMA) Response (FIR) digital filters.
Department of Music, Stanford University
Stanford, California 94305 FFT processors implement long FIR filters more
efficiently than any other method (using Overlap-Add).
March 28, 2005
We need flexible ways to design all kinds of FIR filters for
use in FFT processors.
Outline

• Ideal Lowpass Filter

• Optimal Least Squares in Time Domain
• Window Method
• Optimal Methods
• Case Study: FIR Hilbert-Transform Design

1 2

Ideal Lowpass Filter Impulse Response of Ideal LPF

Amplitude Response:

∆
hideal(n) = DTFT−1
n Hideal
Z ωc
∆ 1
1
= ejωndω
2π −ωc
sin(ωcn)
Gain

=
πn
0
= 2fcsinc(2fcn), n∈Z
0 Frequency fc fs/2
Problems:
• Infinitely long
• Gain = 1 for f < fc • Non-causal
• Gain = 0 for f > fc • Cannot be shifted to make it causal
• Cannot be implemented in practice
Conclusion:
• We must accept some compromise(s) in the design of
any practical lowpass filter.
• Managing such trade-offs is the topic of digital filter
design.
3 4
 Optimal (but Poor) Least-Squares Length L = 30 Least-Squares LPF Design
Impulse Response Design
Length 30 FIR Amplitude Response
Let 1.5

Magnitude (Linear)
• h(n) = ideal filter impulse response. 1

• ĥ(n) = length L causal FIR filter (to be designed). 0.5

Sum of squared errors: 0

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Normalized Frequency (cycles/sample)
0.4 0.45 0.5

∞ 
X 2 XL−1  2
∆    
J2(ĥ) =  h(n) − ĥ(n)  =  h(n) − ĥ(n)  + c2 −10
0

n=−∞

Magnitude (dB)
n=0 −20

∆ P0 2 P∞ 2 −30
where c2 = n=−∞ |h(n)| + n=L |h(n)| does not −40

depend on ĥ. Note that J2(ĥ) ≥ c2. −50

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45

Result: The error is minimized (in the least-squares Normalized Frequency (cycles/sample)

sense) by simply matching the first L terms in the desired

impulse response. • Desired cut-off frequency is fc = 1/4
Optimal least-squares FIR filter: • Stopband attenuation only around 10 dB
(
∆ h(n), 0 ≤ n ≤ L − 1 • Passband gain has a “resonance” near cut-off
ĥ(n) =
0, otherwise • Filter is quite poor for audio use
p • Ĥ = H ∗ asincL
Also optimal under any L norm with any error weighting:
L−1
X  p
 
Jp(ĥ) = w(n) h(n) − ĥ(n) + cp ≥ cp
n=0
5 6

Length L = 71 Least-Squares LPF Design Length L = 71 Optimal Chebyshev LPF Design

Length 71 FIR Amplitude Response

1.5

remez(70,[0 0.5 0.6 1],[1 1 0 0]);

Magnitude (Linear)

Length 71 FIR Amplitude Response

1.5
0.5
Magnitude (Linear)

1
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalized Frequency (cycles/sample)
0.5

0
−10
0
−20 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Magnitude (dB)

Normalized Frequency (cycles/sample)

−30
−40
0
−50
−10
−60
−20
Magnitude (dB)

−70
−30
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 −40
Normalized Frequency (cycles/sample)
−50
−60

• Stopband attenuation still only around 10 dB −70

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45

Normalized Frequency (cycles/sample)
• Corner-frequency resonance is higher, though narrower
• “Gibb’s phenomenon” in the frequency domain • Stopband attenuation better than 60 dB
What we’re doing wrong: • No corner-frequency resonance
• Desired filter specification asks for too much • Transition region from passband to stopband is
(e.g, an infinite roll-off rate). critical
• Time-domain-least-squares is a poor choice of error • Error is “equiripple” in both stopband (visible) and
criterion for audio work. passband (not visible)
7 8
 • Impulse response is slightly “impulsive” at the Lowpass Filter Specifications
endpoints (not shown).

Lowpass Filter Specifications

Amplitude

1 + δp
Ideal lowpass
δp
1 filter response

1 − δp

Pass-band Stop-band

TW

δs
0
0 Frequency ωp Transition ωs
band

• δs : stopband ripple (≤ 0.001 = −60 dB typical)

• δp : passband ripple (≤ 0.1 dB typical)
• ωs: stopband edge frequency
• ωp: passband edge frequency
• TW: transition width = ωs − ωp
• SBA: stop-band attenuation = −20 log(δs)

9 10

Example Ripple Calculations in Matlab So, that’s about 60 dB stop-band rejection, which is not
too bad. Now let’s set the stopband ripple to 1/100
Let’s first consider the passband ripple spec, ±0.1 dB. times the passband ripple:
Converting that to linear ripple amplitude gives, in
Matlab, >> ds=dp/100;
>> 20*log10(ds)
format long; ans =
dp=10^(0.1/20)-1 -78.72623816882052
dp =
0.01157945425990 So now we’re close to 80dB SBA, which is getting into
the ”high fidelity” zone.
Let’s check it:
>> 20*log10(1+dp) Ideal Lowpass Filter
ans =
0.10000000000000 The ideal lowpass filter is defined by the following
>> 20*log10(1-dp) specifications:
ans =
-0.10116471483635 ∆
• ωs = ωp = ωc = 2πfc ⇒ TW = 0
Ok, close enough. Now let’s set the stopband ripple to • δp = δs = 0 ⇒ SBA = ∞
1/10 times the passband ripple and see where we are:

>> ds=dp/10;
>> 20*log10(ds)
ans =
-58.72623816882052
11 12
 The Window Method for FIR Filter
Design
• In practical FFT processors, we are limited to a finite
duration impulse response.
• An “obvious” method for FIR filter design is to simply
Example
A typical windowed ideal lowpass filter response is
depicted in the following diagram:
Length 101 Lowpass FIR Filter - Hamming Window
1

window the ideal impulse response, just like we 0.8

window ideal sinusoids in spectrum analysis: 0.6

∆ 0.4
ĥw (n) = w(n) · hideal(n)

Amplitude
0.2

where w is a zero phase window of length M (odd). 0

• We saw that the rectangular window is optimal in the -0.2

time-domain least-squares sense, but a poor choice -0.4

for typical audio filter design. What about other -0.6

-60 -40 -20 0 20 40 60
windows? time (samples)

• Windowing is multiplication in the time domain ↔

convolution in the frequency domain.
• Thus, in the window method, the ideal frequency
response is convolved with the window transform:
Ĥw (ω) = (W ∗ Hideal)(ω)
The convolution ‘smears’ the response, introducing
deviations in both the pass-band and stop-band.

13 14

• At this point in the design, we have a non-causal Examples

(zero phase) filter.
• In order to implement this in a real time situation, we
• length 25 FIR filter
need to shift the filter by an amount equal to half its
length. • ωc = π/4
  sin(ωc n)
M −1 • hideal(n) = n∈Z
ĥcausal(n) = ĥ n − πn
2 • We will examine several windows:
• This shift in the time domain results in a linear phase
Rectangular Hanning
term in the frequency domain: 0.3 0.3

M −1
Ĥcausal(ω) = e−jω
0.2 0.2
2 Ĥ(ω)
Amplitude

Amplitude

0.1 0.1

0 0

-0.1 -0.1
0 10 20 0 10 20
time (samples) time (samples)
Hamming Kaiser
0.3 0.3

0.2 0.2
Amplitude

Amplitude

0.1 0.1

0 0

-0.1 -0.1
0 10 20 0 10 20
time (samples) time (samples)

15 16
 Amplitude Response of Examples Effect of Changing Filter Length, M = 11, 15, 41

Hamming Window
Rectangular Hanning
0 0 0
Amplitude - dB

Amplitude - dB
-20 -20 −10

-40 -40
−20

-60 -60

Amplitude − dB
−30
-2 0 2 -2 0 2
Frequency (radians/sample) Frequency (radians/sample)
Hamming Kaiser −40

0 0
Amplitude - dB

Amplitude - dB

−50
-20 -20

−60
-40 -40

-60 -60 −70

−3 −2 −1 0 1 2 3
Frequency (radians/sample)
-2 0 2 -2 0 2
Frequency (radians/sample) Frequency (radians/sample)

17 18

Other Types of Filters Summarizing:

• To design a high-pass filter, first design a finite

Ideal highpass, bandpass, and bandstop filters can all be length, causal lowpass filter
considered as modifications of an ideal lowpass filter.
• Use a lowpass cutoff frequency equal to
A highpass filter can be constructed by shifting a lowpass π−(highpass cutoff)
filter by an amount equal to half the sampling rate.
• Design the lowpass filter as in the previous section
|H|
• Modulate the impulse reponse by (−1)n to exchange
ag replacements
DC and fs/2
ω Alternatively, if the passband of the prototype lowpass
−2π
i
−π 0 π 2π filter Ĥ(ω) is very flat about unity gain, a highpass filter
|H|
can be designed by simply subtracting from 1:
∆
ω Hhp(ω) = 1 − Hlp(ω)
−2π −π 0 π 2π ↔ δ(n) − hlp(n)
i

This is equivalent to a modulation by e−jπn in the time A bandpass can be derived by designing the appropriate
domain. Hence, lowpass filter (with cutoff frequency equal to half the
width of the bandpass), and then modulating it to the
desired center frequency ωc. In the frequency domain,
hhp(n) = hlp(n)e−jπn this means shifting two copies of the lowpass prototype,
one to the left by −ωc, and the other to the right by ωc.
= hlp(n)(−1)n
The two shifted copies are added together to give a
bandpass filter having a real impulse response.
19 20
 |H|
Summary of the Window Method

−ω0c ωc ω Basic Idea:

−π π
|H|
• Start with Ideal Lowpass-Filter Impulse Response

−π −ω0 0 ωl ω0 ωu πω • Sinc function sin(πt/T )/(πt/T )

To design a bandpass filter:
• Design a lowpass prototype filter of the desired width
(cutoff ωc = ωu − ω0)
• Shift to the left and right by ω0 and sum:
hbp(n) = 2 cos(nω0)hlp(n)
A bandstop filter can be constructed by first designing a
bandpass with similar specifications.
In the frequency domain, we desire a filter with a transfer
function equal to Hbs = 1 − Hbp. Hence, in the time
domain we have:
( are designed by first designing a low pass, and
1 − hbp(0) n = 0
hbs(n) =
−hbp(n) n = ±1, ±2, . . . , ± M2−1
• Infinite duration
• Non-causal
• Window the infinite duration sinc to get a finite
duration filter
• Window length determines transition width
• Window type determines the sidelobe level (SLL)
• Rectangular window yields the optimal filter in the
unweighted least squares sense
• Shift the noncausal (zero phase) filter to get a causal
(linear phase) filter
• The linear phase slope equals half the filter length
• Other types of filters (highpass, bandpass, bandstop)
applying transformations

21 22

Important Points (Window Method) Optimal FIR Digital Filter Design

• Easy to design (see fir1 and fir2 in Octave and
Matlab SPTB)
• Optimal Chebyshev FIR Design
• Can design extremely long FIR filters without
• Lp Norm Minimization
numerical problems
• Least squares problems
• Not usually optimum in any sense
• Min-Max problems
General Remarks (Window-Method) • Least Squares Optimization
• Transition width TW determined by window length M • PseudoInverse
• As window gets longer, its main lobe narrows • FIR Filter Design
• Narrower main lobe ⇒ narrower transition band • Linear Phase
• Tighter specs demand longer filters • Complex Filter Design
• Length M ∝ 1/TW in general • Other Applications
• Pass-band and stop-band ripple are determined by
window side lobes Books including digital filter design:
• We do not have individual control over passband • Boyd and Vandenberghe
and stopband ripple
• Parks and Burrus
(This is a limitation of the window method)
• Ripple determined by window used • Rabiner and Gold
• Ripple normally largest around transition band • Oppenheim and Schafer
• Steiglitz
23 24
 • Strum and Kirk Optimal Chebyshev Filters

See Music 421 References1

• Minimize the maximum of the error
• Stopband and passband errors are equiripple
• Remez Multiple Exchange (Parks-McClellan
algorithm)
• Available in Octave and the Matlab SP Tool Box:
remez(), cremez()

Problems with Remez Exchange

• Convergence unlikely for FIR filters longer than a few

hundred taps or so
• Fundamentally real polynomial approximation on the
unit circle
• Applies only to linear phase filters
• Cannot design individual (non-minimum phase)
“sub-filters” in polyphase filter bank form
• Really need optimal weighted complex approximation
e.g., minimum phase FIR filter design
1
http://ccrma.stanford.edu/˜jos/refs421/
PSfrag replacements
25 26

replacements ω
π
ω 0
π |H|
0 2π
|H|
−π
Example:
2π Chebyshev window and its transform: −2π DFT of Chebyshev Window: M = 11, Ripple = -40 dB
0
−π ω
−2π 1
Chebyshev Window: M = 101, Ripple = -40 dB -10

ω 0.9 |H| -20

0.8 π
dB Magnitude

-30
|H| 0.7 ωc
-40
π 0.6 ω0
Amplitude

ωc 0.5 ωu -50

ω0 0.4
ωl -60
ωu 0.3
−π -70
ωl 0.2
−ωc -3 -2 -1 0 1 2 3
−π 0.1
−ω0 freq

−ωc -100
0 -50 0 50 100
Time (samples)
−ω0

27 28
 Lp norms • L∞-norm
N −1
! p1
∆
X
The Lp norm of an N -dimensional vector x is defined as k x k∞ = lim |xi|p
p→∞
i=0
! p1
N
X −1 In the limit as p → ∞, the Lp norm of x is
∆
k x kp = |xi|p dominated by the maximum element of x.
i=0

Special cases

• L1 norm
N
X −1
∆
k x k1 = |xi|
i=0
• Sum of the absolute values of the elements
• “City block” distance
• L2 norm v
uN −1
uX
∆ t
k x k2 = |xi|2
i=0

• “Euclidean” distance
• Minimized by “Least Squares” techniques

29 30

Filter Design using Lp Norms Least-Squares Linear-Phase FIR Filter Design

Let the FIR filter length be L + 1 samples, with L even,

In terms of Lp norm minimization, the FIR filter design and suppose we’ll initially design it to be centered about
problem becomes: the time origin (“zero-phase”). Then the frequency
response is given on our frequency grid ωk by
min kW (ωk ) [H(ωk ) − D(ωk )]kp L/2
X
h
H(ωk ) = hne−jωk n
Where n=−L/2
Enforcing even symmetry in the impulse response, i.e.,
• ωk = suitable discrete set of frequencies hn = h−n, gives a zero phase FIR filter which we can
• H(ωk ) = frequency response of our FIR filter later right-shift L/2 samples to make a causal, linear
phase filter. In this case, the frequency response reduces
• D(ωk ) = desired (complex) frequency response
to a sum of cosines:
• W (ωk ) = (optional) weighting function XL/2
H(ωk ) = h0 + 2 hn cos(ωk n), k = 0, 1, 2, . . . , N − 1
Lets look at some specific cases: n=1
or in matrix form:
   
H(ω0) 1 2 cos(ω0 ) . . . 2 cos[ω0(L − 1)] h0
    
 H(ω1)   1 2 cos(ω1 ) . . . 2 cos[ω1(L − 1)]   h1 
 ...  =  ..   .. 
  .  . 
H(ωN −1) 1 2 cos(ωN −1 ) . . . 2 cos[ωN −1(L − 1)] hL/2
| {z } | {z }
A x

(Note that Remez exchange algorithms are also based on

this formulation internally.)
31 32
 Matrix Formulation: Optimal L2 Design, Cont’d Least Squares Optimization

In matrix notation, our filter design problem can be stated ∆

x̂ = arg min k Ax − b k2 = arg min k Ax − b k22
x x
min kAx − bk22
x Hence we can minimize
where, for zero-phase filters, k Ax − b k22 = (Ax − b)T (Ax − b)
 
1 2 cos(ω0) . . . 2 cos [ω0(L − 1)]
Expanding this, we have:
∆ 1 2 cos(ω1) . . . 2 cos [ω1(L − 1)] 
A= . . 

1 2 cos(ωN −1) . . . 2 cos [ωN −1(L − 1)] (Ax − b)T (Ax − b) = (bT − xT AT )(Ax − b)
This is quadratic in x, hence it has a global minimum
∆
x=h which we can find by taking the derivative, setting it to
zero, and solving for x. Doing this yields:
and b = [D(ωk )] is the desired frequency response at the
specified frequencies.
AT Ax − AT b = 0
The function firls in the Matlab Signal Processing Tool PSfrag replacements
Box implements frequency-domain weighted-least-squares These are the ωfamous normal equations whose solution is
FIR filter design (not available in Octave as of 2/2003). given by: π
0  
|H| x̂ = (AT A)−1AT b
2π
The matrix
−π ∆
A† = (AT A)−1AT
−2π
ω pseudo-inverse of the matrix A.
is known as the
33 0 34
|H|
π
ωc
PSfrag replacements ω0
ωu
ω
ωl
Geometrical π Interpretation of Least Squares
−π
0
|H| −ωc column-space of A
Typically, the number of frequency constraints is much −ω0 Ax̂
greater than2π the number of design variables (filter taps).
−π
In these cases, we have an overdetermined system of
−2π
equations (more equations than unknowns). Therefore,
ω
we cannot generally satisfy all the equations, and we are This diagram suggests that the error vector b − Ax is
0
left with minimizing some error criterion to find the orthogonal to the column space of the matrix A, hence it
|H|
“optimal compromise” solution. must be orthogonal to each column in A.
π
In the case ofωcleast-squares approximation, we are
minimizing theω0 Euclidean distance, which suggests the AT (b − Ax̂) = 0 ⇒ AT Ax̂ = AT b
following geometrical
ωu interpretation:
ωl This is how the orthogonality principle can be used to
derive the fact that the best least squares solution is
−π
−ωc given by
column-space of A
−ω0 x̂ = (AT A)−1AT b = A†b
Ax̂

Note that the pseudo-inverse A† projects the vector b

onto the column space of A.
Ax̂ = b + e In Matlab:
Minimizexkek2 = kb − Axk2
•x = A \ b
• x = pinv(A) * b

35 36
 2π
−π
−2π
Chebyshev Optimal Linear Phase filter Design Hence we are minimizing a linear objective, subject to a
ω linear constraints. This is known as a linear
set of
0
programming problem (linprog in Matlab).
Consider the L∞-norm minimization problem: |H|
π
min k Ax − b k∞ ωc
x
ω0
We said earlier that the ∞ norm of a vector is the ωu
maximum element of that vector, hence minimizing the ωl
∞ norm is minimizing the maximum element of a vector: −π
−ωc
min max |aTk x − bk |
x k −ω0 Linear
Constraints
where aTk denotes the kth row of the matrix A.
We can then write this as: Linear
Objective
min t
s.t. |aTk x − bk | < t
 
∆ x
If we introduce a new variable x̃ = , then
t
t = f T x̃ = [ 0 . . . 0 1 ]x̃, and our optimization problem
becomes:
min f T x̃
x̃  
s.t. [ aTk 0 ] · x̃ − bk  < x̃f T x̃

37 38

More General Real FIR Filters Complex FIR Filter Design

So far we have looked at the design of linear phase FIR
filters. In this case, A, x and b are all real.
Sometimes we may want to design filters and specify both
the magnitude and the phase of the frequency response.
Examples:
• Minimum phase filters
• Inverse filters
• Fractional delay filters
• Interpolation polyphase subfilters
In linear-phase filter design, we assumed symmetry of our
filter coefficients [h(n) = h(−n)] ⇒
• The filter frequency response became a sum of
cosines (“zero phase”)
• The matrix A was real
• The desired magnitude response b was real
• The final zero-phase filter x̂ could be right-shifted
L/2 samples to get a corresponding causal
linear-phase FIR filter
Now we would like to specify a complex frequency
response. This means that:

• b is complex
• A is complex
• We still want x (our filter coefficients) to be real

If we try to use ’\’ or pinv in Matlab, we will generally

get a complex result for x̂

39 40
Summarizing our problem: Optimal FIR Filters — Arbitrary Magnitude and
Phase Specificiations
min k Ax − b k2
x
cremez (Matlab Signal Processing Toolbox) performs
where, A ∈ CN xM , b ∈ CN x1, and x ∈ RM x1 complex L∞ FIR filter design:
Hence we have,
• Documented online at The Mathworks
(search for cremez)
min k[R(A) + jI(A)] x − [R(b) + jI(b)]k22
x • Convergence theoretically guaranteed for arbitrary
which can be written as: magnitude and phase specifications versus frequency.
min k R(A)x − R(b) + j [I(A)x + I(b)] k22 • Reduces to Parks-McClellan algorithm (Remez second
x
or     2 algorithm) as a special case.
R(A) R(b)
min
x− • Written by Karam and McClellan. See “Design of
x I(A) I(b) 2
Optimal Digital FIR Filters with Arbitrary Magnitude
which is written in terms of only real variables. and Phase Responses,” by Lina J. Karam and James
H. McClellan, ISCAS-96. The paper may be
Hence, we can use the standard least squares solvers in
downloaded at
Matlab and end up with a real solution.
http://www.eas.asu.edu/~karam/papers/iscas96 km.html

Related paper

“Design of Fractional Delay Filters Using Convex

Optimization” (Mohonk-97, Music 421 handout):
http://ccrma.stanford.edu/~jos/eps/mohonk97.ps.gz
41 42

Second-Order Cone Problems (SOCP) Hilbert Transform Filter Design Example:

A Case Study in FIR Filter Design
• A linear function is minimized over the intersection of
an affine set and the product of second-order Design Example:
(quadratic) cones
• Ideal Single-Sideband Filter (Hilbert transformer)
• Nonlinear, convex problem including linear and
(convex) quadratic programs as special cases • Window Method using Kaiser Window

• Solved by efficient primal-dual interior-point methods • Optimal Chebyshev using Remez Exchange
• Number of iterations required to solve a problem
grows at most as the square root of the problem size
• Typical number of iterations ranges between 5 and
50, almost independent of the problem size

See “Applications of second-order cone programming,” by

M. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret,
Linear Algebra and its Applications, 284:193-228,
November 1998, Special Issue on Linear Algebra in
Control, Signals and Image Processing. Available online:
http://www.stanford.edu/~boyd/socp.html
For more on this topic, see also Prof. Boyd’s course on
convex optimization: EE 3642 and text book (Boyd and
Vandenberghe, Cambridge U. Press, 2004).
2
http://www.stanford.edu/class/ee364/

43 44
 Problem Statement Hilbert Transform

∆
y(t) = (hi ∗ x)(t) (Hilbert transform of x)
• Design a “negative-frequency filter” which converts a
real signal into its complex “analytic signal” ∆ 1
hi(t) = πt (Hilbert transform “kernel”)
counterpart by filtering out negative frequencies.

• Equivalently, design a “single sideband” (SSB) filter  −j, ω > 0

∆
which outputs a single-sideband signal in response to Hi(ω) = j, ω < 0 (Hilbert frequency response)
a complex-conjugate pair of sidebands (a Hermitian 
 1, ω = 0
spectrum). ∆
xa(t) = x(t) + jy(t) (Analytic signal from x)
• We could also call it a “positive-frequency-pass” filter. Z ∞
• The imaginary part of such a filter is called a = π1 X(ω)ejωtdω (Note: Lower limit usually −∞)
Hilbert transform filter. 0
Proof:
• The ideal Hilbert transform is an allpass filter that
delays positive-frequency components by 90 degrees, Xa(ω) = X(ω) + jY (ω) (By linearity of Fourier transform)
and advances negative-frequency components by 90
degrees. = X+ + X− (Apply frequency response)
+j[−jX+ + jX−]

= 2X+(ω)

45 46

PSfrag replacements
ω
π
0
|H|
Kaiser Window 2π 0 Kaiser Window Transform, FFT size = 2048, Window length = 257

−π
−20
Kaiser window in causal, linear-phase form: −2π
ω −40

w = kaiser(M,beta); % M=length=257, beta=8

|H| −60

Zero-pad by a factor of 8 and convert to zero-phase form

Gain (dB)

π −80
(N = FFT size = 2048): ωc
ω0 −100

wzp = [w((M+1)/2:M),zeros(1,N-M),w(1:(M-1)/2)]; ωu −120

Kaiser window transform: ωl

−π −140

W = fft(wzp); −ωc
−ω0
−160
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2
Normalized Frequency (cycles/sample)
0.3 0.4 0.5

47 48
eplacements
ω
π
Kaiser window transform, close-up on main lobe Oversimplified Window Method
0
|H| Sample the ideal Hilbert-transform kernel
Close Up on Kaiser LF Response, FFT size = 2048, Window length = 257
2π 0
h(t) = 1/πt
−π
−2π −20 to get
ω ∆ 1
ĥi(n) = (Sampled Hilbert transform kernel)
−40 πnT
|H| ∆
ĥw (n) = w(n)ĥi(n) (Windowed ideal impulse response)
Gain (dB)

π −60

ωc Ĥw (ω) = (W ∗ Ĥ)(n) (Smoothed ideal frequency response)

ω0 −80

ωu Design Parameters:
ωl −100 fs = 22050; % sampling rate (Hz)
−π T = 1/fs; % sampling period (sec)
−ωc M = 257; % FIR filter length = window length
−ω0
−120
−0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 0.02
Normalized Frequency (cycles/sample)
0.03 0.04 0.05
N = 8*(M-1); % for interpolated spectral displays
beta = 8; % beta for Kaiser window

Filter Design:
n = [-N/2+0.5:N/2-0.5]; % Time axis (avoid t=0)
hi = T ./ (pi*n*T); % Sampled Hilbert kernel
hr = sin(pi*n) ./ (pi*n); % 1/2 sample delay filter
h = (hr + j*hi)/2; % Sampled ideal final filter
plot(f,fftshift(max(-100,20*log10(abs(fft(h)))))); grid;
...

49 50

eplacements
ω
π
0
FFT of Truncated(2048) Ideal Impulse Response Hwp = [Hw(N/2+2:N), Hw(1:N/2+1)]; % Neg. freqs on left
|H| Hwpn = abs(Hwp); Hwpn = Hwpn/max(Hwpn);
2π plot(f,20*log10(Hwpn)); grid;
PSfrag replacements
...
−π 20
FFT of Truncated Analytically Specified "Ideal" Impulse Response

−2π ω
ω 0 FIR
π Frequency Response by the Window Method
0
|H| -20 |H|
Length 257 Analytic−Derived FIR Frequency Response, FFT size = 2048

π 2π 0
Gain (dB)

ωc
-40 −π
ω0 −2π −20

-60
ωu ω
−40

ωl -80
|H| −60
−π
Gain (dB)

−ωc -100
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
π
−ω0 Normalized Frequency (cycles/sample) ω c −80

ω0
h = [h(N/2+1:N),h(1:N/2)]; % zero-phase form (almost) ωu −100
hw = wzp .* h; % Apply window to ideal impulse response ωl
−120
Hw = fft(hw); % Frequency response we really get −π
−ωc −140
% Compute total stopband attenuation:
−ω0 −0.5 −0.4 −0.3 −0.2Normalized −0.1 0 0.1 0.2 0.3 0.4 0.5
Frequency (cycles/sample)
ierr = norm(Hw(N/2+2:N))/norm(Hw)
= 0.0378 = 3.8 percent

For spectral displays, rotate buffer so negative frequencies

are on the left and DC is at k = N/2:
51 52
eplacements
ω
π Zoom in on low-frequency transition band Ideal Bandlimited Impulse Response
0
|H|
Analytic−Derived Low−Frequency Transition Band, f1/fs=0.023926 (marked by "+")
Since we are working with sampled data, the ideal
2π 20 single-side-band filter impulse response is
−π Z π
−2π 0 ∆ 1
h(n) = IDTFTn(u) = u(ω)ejωndω
ω −20 2π −π
where u(ω) is the unit step function in the frequency
|H| −40
domain: (
Gain (dB)

π −60 ∆ 1, ω ≥ 0
ωc u(ω) =
0, ω < 0
ω0 −80

ωu We can evaluate the IDTFT directly as follows:

−100 π
Z π
ωl 1 1 jωn ejπn − 1
h(n) = ejωndω = e  =
−π −120
2π 0 2πjn 0 2πjn
−ωc (
−ω0
−140
−0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 0.02
Normalized Frequency (cycles/sample)
0.03 0.04 0.05
(−1)n − 1 0, n even, n 6= 0
= = 1
2πjn j πn , n odd
• Poorly positioned transition region from positive to
For n = 0, going back to the original integral gives
negative frequencies. Z π
1 1
• Need to first bandpass-filter the ideal impulse h(0) = ejω·0dω = .
2π 0 2
response.
Thus, the real part of h(n) is δ(n)/2, and the imaginary
• High-frequency transition identical, by symmetry
part is 1/(πn) for odd n and zero otherwise (as a result
of bandlimiting). There is no aliasing. However, we also
53 54

PSfrag replacements
ω
need a transition band. We can work this out analytically, π Desired Frequency Response
or we can simply “draw” it in the frequency domain, take 0
a (large) inverse DFT, and window that. Since the latter |H|
Desired Frequency Response, FFT size = 2048
approach is more general, we’ll do that. 2π 0
−π
Window Method applied to Bandlimited Ideal −2π
Impulse Response ω
−50

Further Design Parameters: |H|

Gain (dB)

π
f1 = 530; % lower passband limit (Hz) ωc
N = 2^(nextpow2(8*fs/f1)) % FFT size from xition width ω0 −100
if N<8*M, N = 8*(M-1); end; % spectral interpolation ωu
ωl
Bandlimited “Ideal” Hilbert transformer frequency
response (FR), including transition bands ∆ matched to −π
FFT size N such that N  fs/∆: −ωc
−ω0
−150
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2
Normalized Frequency (cycles/sample)
0.3 0.4 0.5

H = [ ([0:k1-2]/(k1-1)).^8,ones(1,k2-k1+1), ...
([k1-2:-1:0]/(k1-1)).^8,zeros(1,N/2-1)];
Hp = [H(N/2+2:N), H(1:N/2+1)];
f = [-0.5 + 1/N : 1/N : 0.5]; normalized freq axis
plot(f,Hp); grid; ...

55 56
 PSfrag replacements
ω
Desired Impulse Response π Real Part of Desired Impulse Response
0
|H|
Ideal SSB Filter Impulse Response, Real Part, FFT size = 2048
h = ifft(H); % zero-phase form 2π 0.6

−π
% measure of round-off error:
ierr = norm(imag(h(1:2:N)))/norm(h) % zero?
−2π 0.5

ω 0.4

ierr =
|H| 0.3

Amplitude
4.1958e-15 π
ωc 0.2

% rough estimate of time-aliasing error: ω0

aerr = norm(h(N/2-N/32:N/2+N/32))/norm(h)
ωu 0.1

aerr =
ωl
0
−π
4.8300e-04 −ωc
−ω0
−0.1
−150 −100 −50 0
Time (samples)
50 100 150

% for plots, prefer negative times on the left:

hp = [h(N/2+2:N), h(1:N/2+1)];

57 58

eplacements
PSfrag replacements
ω
π ω
Imaginary Part of Desired Impulse Response π Final FIR Filter Frequency Response
0
|H| 0
Ideal SSB Filter Impulse Response, Imaginary Part, FFT size = 2048 |H|
2π 0.4 Length 257 FIR filter Frequency Response, FFT size = 2048

−π 2π 0
−2π
0.3 −π −20
−2π
ω 0.2
ω −40

|H| 0.1 −60

|H|
Amplitude

π 0 −80
Gain (dB)

ωc π
ω0 −0.1 ωc −100

ωu ω0 −120
−0.2

ωl ωu
−140
ωl
−π −0.3

−ωc −π −160

−ω0
−0.4
−150 −100 −50 0 50 100 150 −ωc
Time (samples)
−ω0
−180
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2
Normalized Frequency (cycles/sample)
0.3 0.4 0.5

Apply window to ideal impulse response:

hw = wzp .* h; % Final FIR coefficients
Hw = fft(hw); % Frequency response we’ll see

Total stopband energy:

ierr = norm(Hw(N/2+2:N))/norm(Hw)
disp(sprintf([’Stop-band energy is %g percent of’,...
total spectral energy’], 100*ierr));

59 60
eplacements
Zoom-in
ω on Transition Region of Final Frequency Remez Multiple Exchange Method
π Response
0
|H| • Remez exchange algorithm for optimal Chebyshev
2π 20
Low−Frequency Transition Band, f1/fs=0.023926 (marked by "+") (equiripple) FIR filter design.
−π • Original reference: J. H. McClellan, T. W. Parks, and
−2π 0
L. R. Rabiner, “A Computer Program for Designing
ω −20 Optimum FIR Linear Phase Digital Filters,” IEEE
Trans. Audio Electro., vol AU-21, Dec. 1973,
|H| −40
pp. 506-526.
Gain (dB)

π −60
• Text reference: L. R. Rabiner and B. Gold, Theory
ωc
and Application of Digital Signal Processing,
ω0 −80
Prentice-Hall, 1975.
ωu −100

ωl • Pertinent recent reference: A. Reilly, G. Frazer, and

−π −120 B. Boashash, “Analytic Signal Generation—Tips and
−ωc Traps,” IEEE Tr. Signal Processing, vol. 42, no. 11,
−ω0
−140
−0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 0.02
Normalized Frequency (cycles/sample)
0.03 0.04 0.05
Nov. 1994.
Design an optimal lowpass filter of the desired width
• Transition band better positioned (takes quite a while):
• ≈ 100 dB stob-band attenuation hrm = remez(M-1, [0,(f2-fs/4)/fn,0.5,1], ...
[1,1,0,0], [1,10]);

The weighting [1,P] says make the passband ripple P

times that of stopband. For steady-state audio spectra,
61 62

PSfrag replacements
passband ripple can be 0.1 dB or more. However, ω Optimal Chebyshev FIR(257) Frequency
consider an FM signal in which a sinusoid is sweeping π Response
back and forth in the passband. In that case, the 0
passband ripple generates AM sidebands, so a spec more |H|
Length 257 Optimal Chebyshev FIR Single−Sideband−Filter Freq. Response
like that in the stopband may be called for. Here, we 2π 50

allow the passband ripple to be 10 times the stopband −π

ripple, as a compromise. −2π
0
Modulate it up to where it’s a single-sideband filter. I.e., ω
right-shift by π/2 in the frequency domain. I.e., rotate
the frequency response counterclockwise along the unit |H| −50
Gain (dB)

circle by 90 degrees: π
ωc −100
hr = hrm .* j .^ [0:M-1];
ω0
ωu
ωl −150

−π
−ωc
−ω0
−200
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2
Normalized Frequency (cycles/sample)
0.3 0.4 0.5

63 64
eplacements
PSfrag replacements
ω
ω
π Zoom-In on Transition Band Passband Ripple
π
0
|H| 0
Low−Frequency Transition Band, f1/fs=0.023926 (marked by "+")
|H|
2π 50 x 10
−4 Passband Ripple

2π 4

−π
−π
−2π 3
0 −2π
ω
ω 2

|H| −50 1
|H|
Gain (dB)

Gain (dB)
π 0
ωc
−100 ωc
ω0 −1
ω0
ωu
ωu
ωl −150
−2

ωl
−π
−π −3

−ωc −200
−0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 0.02 0.03 0.04 0.05 −ωc
−ω0 Normalized Frequency (cycles/sample)
−4
0.18 0.2 0.22 0.24 0.26 0.28 0.3 0.32
−ω0 Normalized Frequency (cycles/sample)

65 66

eplacements
ω
π
Initial Impulse Response Summary of Trade-Offs
0
|H| −4 First 50 impulse−response coefficients Window-method filter:
x 10
8
2π
−π 6 • Stopband ”droops” which ”wastes” filter taps if the
−2π worst-case stopband attenuation is the only stopband
ω 4
spec. However, such ”roll-off” is natural in the
2
frequency domain and corresponds to a ”smoother
|H| pulse” in the time domain.
Amplitude

π 0
• Passband is degraded by early roll-off. Edge is off.
ωc
ω0
−2
• Filter length can be thousands of taps as needed for
ωu −4
full audio band.
ωl
−π −6 Optimal Remez-Exchange filter:
−ωc
−ω0
−8
0 5 10 15 20 25 30 35 40 45 50
• Stopband is ideal, equiripple.
Index

• Transition region close to half that of the window

method.
• Pass-band is ideal, equiripple.
• Time-domain impulses from passband ripple are small.
• Design time orders of magnitude larger than that for
window method.

67 68
 • Fails to converge for filters much longer than 256 Matlab hilbert() function
taps. (Need to increase working precision to get
longer filters.) The Matlab Hilbert function returns an analytic signal
given its real part.
PSfrag replacements
Nh
ω = M-2; % This looks best
delta = [1,zeros(1,Nh)]; % zero-phase impulse
π
hh = hilbert(delta); % zeros negative-freq FFT bins
0 = fft(hh); % FIR frequency response
Hh
|H|
Frequency Response of Length 256 hilbert() using Matlab
2π 0
−π
−2π −50

ω
−100

|H| −150

Gain (dB)
π
ωc −200

ω0
ωu −250

ωl
−300
−π
−ωc
−ω0
−350
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2
Normalized Frequency (cycles/sample)
0.3 0.4 0.5

Looks pretty good, but let’s zero-pad the impulse

response to interpolate the frequency response:
69 70

PSfrag replacements
ω
eplacements π
0
ω |H|
π = [hh(Lh/2+1:Lh), zeros(1,N-Lh), hh(1:Lh/2)];
hhzp Low−Frequency Transition Band, f1/fs=0.023926 (marked by "+")
2π 50
0 = fft(hhzp); % Frequency response
Hhzp
−π
|H| 0
Interpolated Frequency Response of Length 256 hilbert() using Matlab −2π
2π 0
ω −50
−π
−2π −50
|H| −100

ω
Gain (dB)

−100 π −150

ωc
|H| −150 ω0 −200
Gain (dB)

π ωu
ωc −200
−250

ωl
ω0
−π −300

ωu −250
−ωc
ωl −350
−0.05 −0.04 −0.03 −0.02 −0.01 0 0.01 0.02 0.03 0.04 0.05
−300 −ω0 Normalized Frequency (cycles/sample)
−π
−ωc −350
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 Problems:
−ω0 Normalized Frequency (cycles/sample)

• Transition bandwidth only 2 bins wide.

• Massive time aliasing.
• Only justifiable for periodic signals with period exactly
equal to filter length Nh. In this case only, time
aliasing is equivalent to overlap-add from adjacent
periods.

71 72
More generally, any real signal given to hilbert() must be
interpreted as precisely one period of a periodic signal.

73