Subject: Digital Signal Processing
Subject Code: BTEC-502-18
Module 3: Digital Filters Design
Question Bank Solution
CO3:
Design of different types of digital filters for various applications.
Dr. Ashanand
Assistant Professor
CCE, SET, CGC University, Mohali.
Q1. What is a direct form I realization of an IIR filter?
Answer: Direct Form I realization implements the difference equation directly, using separate delay
elements for the input (numerator) and output (denominator) sections. It requires 2(N+M) delay
elements for an N-th order IIR filter and is straightforward but uses more memory than Direct Form
II.
Q2. What is the difference between Direct Form I and Direct Form II?
Answer: Direct Form II (canonical form) uses a shared delay line for both the recursive and non-
recursive parts, requiring only max(N,M) delay elements instead of N+M. It is more hardware-
efficient. The transposed forms (IIR-I and IIR-II transposed) are obtained by transposing the signal
flow graph, offering different numerical properties.
Q3. How is a Cascade (series) form structure realized?
Answer: A cascade form factors the transfer function H(z) into second-order sections (SOS): H(z) =
H1(z)·H2(z)·...·Hk(z). Each section is typically a biquad (2nd order). This form is preferred for high-
order filters because it reduces coefficient sensitivity and limit cycle problems.
Q4. Explain Parallel form realization of a digital filter.
Answer: In parallel form, H(z) is expanded using partial fractions into a sum of first and second-
order sections: H(z) = H1(z) + H2(z) + ... + Hk(z). All sections receive the same input, and outputs are
summed. It allows parallel computation and can have better noise performance than cascade in some
cases.
Q5. What is a Lattice structure for FIR filters?
Answer: A lattice FIR filter uses a chain of two-multiplier lattice stages with reflection coefficients
k_m. The forward and backward prediction error signals propagate through the lattice. The reflection
coefficients directly relate to the parcor (partial correlation) coefficients and guarantee stability when
|k_m| < 1.
Q6. Describe the Lattice-Ladder structure for IIR filters.
Answer: The IIR lattice-ladder (lattice-ladder) structure combines a lattice section (all-pole/AR part)
with a ladder (MA) section. It is characterized by reflection coefficients {k_m} and ladder
coefficients {c_m}. This structure has modular form, is useful in adaptive filtering, and ensures
stability if |k_m| < 1 for all stages.
Q7. What are the advantages of lattice structures over direct form?
Answer: Lattice structures offer: (1) modular and systematic design, (2) guaranteed stability for IIR
filters when |k_m| < 1, (3) easier order update (adding one stage), (4) better numerical properties and
reduced sensitivity to coefficient quantization, and (5) direct connection to linear prediction theory.
Q8. What is the Impulse Invariance method for IIR filter design?
Answer: Impulse Invariance maps an analog filter to a digital filter by matching the impulse
response: h[n] = T·ha(nT), where T is the sampling period. The digital poles are placed at z = e^(saT).
It preserves the shape of the impulse response but suffers from aliasing since the frequency response
is a sum of shifted versions of Ha(jΩ).
Q9. What is the Bilinear Transformation method?
Answer: The Bilinear Transformation maps the entire s-plane to the z-plane using s = (2/T)·(z-
1)/(z+1). It eliminates aliasing and maps the jΩ axis to the unit circle exactly. The drawback is
frequency warping: the digital frequency ω is related to analog frequency Ω by Ω = (2/T)·tan(ω/2),
requiring pre-warping of critical frequencies.
Q10. What is frequency pre-warping in the Bilinear Transformation?
Answer: Before applying the bilinear transformation, critical frequencies (cutoff, passband edge)
must be pre-warped: Ω_c = (2/T)·tan(ω_c/2). The analog prototype is then designed at the pre-warped
frequency Ω_c. This ensures that the digital filter has the desired cutoff at exactly ω_c after the
bilinear transformation.
Q11. Compare Impulse Invariance and Bilinear Transformation methods.
Answer: Impulse Invariance: preserves impulse response shape, has aliasing (unsuitable for high-
pass or band-stop), mapping is not one-to-one. Bilinear Transformation: no aliasing, one-to-one
mapping of entire s-plane to z-plane, but introduces nonlinear frequency warping that requires pre-
warping of specifications. BZT is generally preferred except when time-domain matching is critical.
Q12. What is the Rectangular window and its characteristics?
Answer: The rectangular window w[n] = 1 for 0 ≤ n ≤ M-1, zero otherwise. It has the narrowest main
lobe width (4π/M) but the highest peak sidelobe level of about -13 dB. This causes the Gibbs
phenomenon (9% overshoot) at discontinuities. It provides the sharpest transition but poorest
stopband attenuation (~21 dB).
Q13. Describe the Hamming window and its properties.
Answer: The Hamming window is w[n] = 0.54 - 0.46·cos(2πn/(M-1)). It has a peak sidelobe level of
-41 dB and provides about 53 dB stopband attenuation. The main lobe width is 8π/M. It eliminates
the largest sidelobe of the Hann window and offers a good trade-off between main lobe width and
sidelobe level for most practical applications.
Q14. Describe the Hanning (Hann) window and compare it to Hamming.
Answer: The Hanning window is w[n] = 0.5 - 0.5·cos(2πn/(M-1)). Its peak sidelobe is -31 dB and
provides ~44 dB stopband attenuation. Compared to Hamming: Hanning has slightly higher sidelobe
(-31 dB vs -41 dB) but its sidelobes roll off faster. Hamming is better for fixed attenuation
requirements; Hanning is preferred when sidelobe rolloff matters.
Q15. What is the Parks-McClellan (Equiripple) method for FIR design?
Answer: The Parks-McClellan algorithm uses the Chebyshev (minimax) approximation criterion and
the Remez exchange algorithm to design linear-phase FIR filters with equiripple error in both
passband and stopband. It minimizes the maximum weighted error, yielding the optimal filter for a
given order. It is implemented via the MATLAB function firpm (formerly remez).
Q16. What are the four types of linear-phase FIR filters?
Answer: Type I: Odd length, even symmetry — can be any filter type. Type II: Even length, even
symmetry — zero at z=-1 (ω=π), not suitable for high-pass/band-stop. Type III: Odd length, odd
(antisymmetric) — zeros at z=±1, used for Hilbert/differentiators. Type IV: Even length, odd
symmetry — zero at z=1 (ω=0), used for Hilbert/differentiators.
Q17. What is the Butterworth approximation for filter design?
Answer: Butterworth filters have maximally flat magnitude response in the passband: |H(jΩ)|² =
1/(1+(Ω/Ω_c)^(2N)). All zeros of (1+ε²|H|²) occur at the origin or infinity; there are no ripples. The
roll-off is -20N dB/decade. Poles lie on a circle in the s-plane. Filter order N = log[(10^(0.1αs)-
1)/(10^(0.1αp)-1)] / [2·log(Ωs/Ωp)].
Q18. What is the Chebyshev Type I approximation?
Answer: Chebyshev Type I filters have equiripple in the passband and monotone in the stopband.
The magnitude is |H(jΩ)|² = 1/(1+ε²TN²(Ω/Ωp)), where TN is the N-th order Chebyshev polynomial.
For a given order, it provides sharper transition than Butterworth. Poles lie on an ellipse in the s-
plane.
Q19. What is the Chebyshev Type II (Inverse Chebyshev) approximation?
Answer: Chebyshev Type II (inverse Chebyshev) has maximally flat (monotone) passband and
equiripple stopband. It has both poles and zeros in the s-plane. The zeros lie on the imaginary axis in
the stopband. For a given specification, it requires the same order as Type I but the passband is
smoother while the stopband has controlled equiripple.
Q20. What is the Elliptic (Cauer) filter approximation?
Answer: Elliptic filters have equiripple in both passband and stopband. For given passband ripple,
stopband attenuation, and transition ratio, elliptic filters achieve the minimum order (sharpest
transition band). They have both poles and zeros. The design uses Jacobi elliptic functions. The
penalty is the most nonlinear phase response among the classical approximations.
Q21. How do you design a digital high-pass filter using BZT?
Answer: Step 1: Specify digital high-pass cutoff ωc and pre-warp to Ωc = (2/T)·tan(ωc/2). Step 2:
Apply low-pass to high-pass frequency transformation: s → Ωc/s in the analog prototype. Step 3:
Design the analog low-pass prototype. Step 4: Apply the frequency transformation. Step 5: Apply
BZT: s = (2/T)·(z-1)/(z+1) to get H(z).
Q22. How is a digital bandpass filter designed from an analog LP prototype?
Answer: Apply the LP-to-BP transformation: s → (s²+Ω0²)/(Bs), where Ω0² = Ωl·Ωu (geometric
mean) and B = Ωu - Ωl (bandwidth). After pre-warping the digital band edges, this doubles the filter
order. Then apply the BZT to obtain H(z). The resulting bandpass filter has center frequency at Ω0
and bandwidth B.
Q23. What is a band-stop (notch) filter and how is it designed?
Answer: A band-stop filter rejects a band of frequencies between Ωl and Ωu. Design uses the LP-to-
BS transformation: s → Bs/(s²+Ω0²). This maps the stopband of the LP prototype to the rejection
band of the BS filter. A special case is the notch filter, which eliminates a single frequency and can be
designed with a pair of complex zeros on the unit circle at the notch frequency.
Q24. What is the effect of coefficient quantization in FIR filters?
Answer: Quantizing FIR coefficients h[n] to B-bit fixed-point representation introduces errors Δh[n].
These shift the frequency response, moving zeros away from their ideal positions on the unit circle.
For linear-phase FIR, symmetry is preserved if quantization is symmetric. The passband/stopband
ripple increases and the transition band may widen. More bits (higher B) reduce these effects.
Q25. What is the effect of finite register length on FIR filter output?
Answer: In fixed-point FIR implementation, products h[k]·x[n-k] must be rounded or truncated. This
introduces roundoff noise at each multiplier output. For an N-tap FIR filter, there are N+1 roundoff
noise sources. The output noise power is approximately (N+1)·σ²_e, where σ²_e = 2^(-2B)/12 for
rounding. Unlike IIR, FIR filters have no feedback, so roundoff noise does not accumulate (no limit
cycles).
Q26. Why are FIR filters preferred over IIR for finite register length concerns?
Answer: FIR filters have no feedback (all-zero structure), so roundoff noise generated in
multiplications does not feed back and grow. There are no limit cycles or overflow oscillations
possible. IIR filters, due to their recursive nature, can exhibit limit cycles and potentially unstable
overflow behavior. FIR filters also guarantee linear phase, making them preferable in many
applications.
Q27. What is the dynamic range problem in fixed-point filter implementation?
Answer: In fixed-point arithmetic, signal levels must be scaled to prevent overflow while keeping the
signal well above the noise floor. Too large a signal causes overflow (clipping distortion); too small a
signal causes the signal-to-noise ratio (SNR) to degrade. Proper scaling between filter sections (in
cascade form) is critical. The ratio of maximum to minimum signal level defines the dynamic range
requirement.
Q28. What is parametric spectral estimation?
Answer: Parametric spectral estimation assumes the signal is generated by a model with finite
parameters (e.g., AR, MA, ARMA). The spectrum is estimated from the model parameters. Examples
include the Yule-Walker method (AR), Burg's method (AR lattice), and MUSIC/ESPRIT for
sinusoids in noise. These methods can have higher resolution than non-parametric methods,
especially for short data records.
Q29. What is non-parametric spectral estimation?
Answer: Non-parametric spectral estimation estimates the power spectral density (PSD) directly from
the data without assuming a signal model. The periodogram (|DFT|²/N) is the basic estimator.
Bartlett's method averages periodograms of non-overlapping segments. Welch's method uses
overlapping segments with windowing to reduce variance. These are simple but have limited
frequency resolution for short records.
Q30. What is the Bartlett method for PSD estimation?
Answer: The Bartlett method divides the N-point signal into K non-overlapping segments of length
M = N/K. The periodogram of each segment is computed and averaged: P(ω) =
(1/K)·Σ|DFT_k(ω)|²/M. Averaging reduces variance by a factor of K but also reduces frequency
resolution (resolution ≈ 1/M). It provides a consistent estimate of the PSD.
Q31. What is Welch's method of spectral estimation?
Answer: Welch's method extends Bartlett's by allowing overlapping segments (typically 50%) and
windowing each segment before DFT computation. The overlap increases the number of segments for
a given data length, further reducing variance. Windowing reduces spectral leakage. It offers better
variance-resolution trade-off than Bartlett and is widely used in practice (e.g., MATLAB's pwelch).
Q32. What is the MUSIC algorithm in parametric spectral estimation?
Answer: MUSIC (MUltiple SIgnal Classification) is a subspace-based method that estimates the
frequencies of sinusoids in noise. It decomposes the data correlation matrix into signal and noise
subspaces. The MUSIC pseudospectrum is 1/(a^H(ω)·E_n·E_n^H·a(ω)), where E_n is the noise
eigenvector matrix. Peaks of the pseudospectrum indicate signal frequencies. It provides super-
resolution for closely-spaced sinusoids.
Q33. What is the Matched Z-Transform method?
Answer: The Matched Z-Transform (MZT) places digital zeros at locations corresponding to analog
zeros via z = e^(sT): for each analog zero at s = sa, the digital zero is at z = e^(saT). Poles are
similarly mapped. Unlike impulse invariance, the zeros are matched explicitly. It avoids the aliasing
problem somewhat and is used when the analog filter has finite zeros (e.g., elliptic, inverse
Chebyshev filters).
Q34. What are the analog frequency transformations used in filter design?
Answer: Standard analog LP-to-X transformations: (1) LP→LP: s → s/Ω_c; (2) LP→HP: s →
Ω_c/s; (3) LP→BP: s → (s²+Ω0²)/(B·s); (4) LP→BS: s → B·s/(s²+Ω0²). These map a normalized LP
prototype to the desired filter type. After transformation, the BZT is applied to obtain the digital filter.
Q35. What are digital frequency transformations?
Answer: Digital frequency transformations map a digital LP prototype H(z) to other filter types
directly in the z-domain using allpass substitutions. For LP→HP: z⁻¹ → -(z⁻¹+α)/(1+αz⁻¹). For
LP→BP: z⁻¹ → -(z⁻²-a1z⁻¹+a0)/(a0z⁻²-a1z⁻¹+1). These preserve the unit circle and maintain stability
(allpass transformations map inside of unit circle to itself).
Q36. What is the role of pre-warping in digital filter design via BZT?
Answer: Pre-warping corrects the nonlinear frequency compression introduced by the bilinear
transformation. The BZT maps ω ∈ [0,π] to Ω ∈ [0,∞) via Ω = (2/T)tan(ω/2), compressing high
frequencies. Pre-warping ensures critical frequencies (passband edge, stopband edge) are mapped
exactly: Ω_prewarped = (2/T)tan(ω_desired/2). The analog prototype is then designed at
Ω_prewarped.
Q37. What is fixed-point representation in digital signal processing?
Answer: Fixed-point numbers represent real values as integers scaled by a fixed power of two. In a
B-bit 2's complement representation, the range is [-1, 1-2^(-(B-1))] for fractional format, with
resolution 2^(-(B-1)). It is computationally simpler and faster than floating point but has limited
dynamic range, making overflow and underflow management critical in filter implementations.
Q38. What is floating-point representation and its advantage over fixed-point?
Answer: Floating-point represents numbers as m × 2^e, where m is the mantissa (fraction) and e is
the exponent. IEEE 754 single precision: 1 sign bit, 8 exponent bits, 23 mantissa bits. Advantage:
much larger dynamic range (2^(-126) to 2^127 for single precision), relative error is nearly constant
across the range (approximately 2^(-23)), reducing the need for manual scaling. The trade-off is
higher hardware complexity and power consumption.
Q39. What is round-off noise in digital filters?
Answer: Round-off noise arises from quantizing multiplier outputs to a finite number of bits. In the
linear noise model, each quantizer is replaced by an additive white noise source with variance σ²_e =
2^(-2B)/12 (for rounding, B bits). In IIR filters, this noise is filtered by the transfer function from the
noise source to the output, and can be amplified by filter poles near the unit circle (high-Q filters are
more sensitive).
Q40. What is a limit cycle in a recursive digital filter?
Answer: A limit cycle is a low-level periodic oscillation that persists in a recursive (IIR) digital filter
even when the input is zero or constant. It arises from the nonlinearity of quantization in the feedback
loop. Granular limit cycles are small oscillations near zero; overflow limit cycles are large oscillations
due to overflow. They can be suppressed by using two's complement arithmetic (which prevents
overflow oscillations) and rounding (vs truncation) for granular cycles.
Q41. What are overflow limit cycles and how are they prevented?
Answer: Overflow limit cycles occur when arithmetic overflow in the feedback loop causes large-
amplitude oscillations. They can be prevented by: (1) saturating arithmetic (clip to max/min instead of
wrapping around), (2) proper scaling to prevent overflow, and (3) using two's complement overflow
detection. Saturation arithmetic eliminates overflow oscillations since it breaks the periodic orbit that
wrap-around arithmetic creates.
Q42. What is the effect of coefficient quantization on IIR filter poles?
Answer: Quantizing IIR coefficients moves poles from their ideal positions, altering the frequency
response and potentially destabilizing the filter if poles move outside the unit circle. High-order direct
form structures are most sensitive because small changes in one coefficient can dramatically shift
multiple poles (root sensitivity). Cascade and parallel second-order section implementations are much
less sensitive, as each pole is controlled by local coefficients.
Q43. What is the signal-to-quantization-noise ratio (SQNR)?
Answer: SQNR = 10·log10(signal power / quantization noise power). For a B-bit uniform quantizer
with a full-scale sinusoidal input, SQNR ≈ 6.02B + 1.76 dB. Each additional bit adds approximately 6
dB of SQNR. This formula guides the choice of word length: for 16-bit audio, SQNR ≈ 98 dB; for 24-
bit, SQNR ≈ 146 dB.
Q44. What is the significance of pole-zero sensitivity in filter structures?
Answer: Pole-zero sensitivity measures how much poles/zeros shift when coefficients are perturbed
(quantized). Direct form structures have high sensitivity for high-order or narrowband filters (poles
close to unit circle). Cascade form with second-order sections has low sensitivity since each biquad
has independent coefficients. The sensitivity matrix S = ∂z_k/∂a_i quantifies how coefficient errors
map to root errors.
Q45. Explain the concept of deadband in digital filters.
Answer: A deadband is a region of input values near zero for which the filter output does not change,
due to quantization in the feedback loop. In fixed-point IIR filters with rounding, the state variables
can become trapped in a constant non-zero state (a trivial limit cycle) if the recursive term is small
enough that quantization rounds it to zero net change. This causes a DC offset in the output rather
than a true zero.
Q46. What is the two's complement number system and why is it used in DSP?
Answer: Two's complement represents negative numbers so that N + (-N) = 2^B (ignoring overflow).
Negative number: invert all bits and add 1. Range for B bits: -2^(B-1) to 2^(B-1)-1. It is preferred in
DSP because: (1) addition and subtraction use the same hardware, (2) overflow in addition of
oppositely-signed numbers is self-correcting, and (3) two's complement overflow (wrap-around) in
accumulation can actually avoid overflow limit cycles in certain structures.
Q47. What is the noise gain of a digital filter structure?
Answer: Noise gain is the ratio of output noise power to input noise source power. For a filter with
transfer function H(z) carrying noise from an internal node, the output noise power is σ²_e · ||H(z)||² =
σ²_e · Σ|h[n]|². Direct form II has higher noise gain than Direct Form I for certain filters because the
noise at the state nodes sees the full H(z) transfer function. Transposed forms can have different noise
gains.
Q48. What is scaling in fixed-point filter implementation?
Answer: Scaling adjusts signal levels at internal nodes to prevent overflow while maximizing SNR.
L∞ scaling ensures the maximum signal amplitude at each node never exceeds 1.0; this uses the L1
norm of impulse responses. L2 scaling keeps the RMS value at each node equal (equal power), using
the L2 norm. Proper scaling between cascade sections is critical: scaling factors must be absorbed
into the filter coefficients without changing the overall transfer function.
Q49. What is the effect of the number of bits (word length) on filter performance?
Answer: Increasing word length B improves performance in all aspects: (1) coefficient quantization
error decreases as 2^(-B), reducing pole/zero displacement; (2) round-off noise power decreases as
2^(-2B)/12; (3) dynamic range increases by 6 dB per bit; (4) limit cycle amplitudes decrease. Modern
DSP processors use 16-24 bit fixed-point or 32-bit floating point, with the choice depending on cost,
power, and required precision.
Q50. What is the role of the accumulator register length in fixed-point DSP?
Answer: In fixed-point DSP processors, the accumulator is typically longer (e.g., 40 bits for a 16-bit
processor) to prevent intermediate overflow during multiply-accumulate operations. All partial
products are accumulated in the extended precision accumulator before the final result is
rounded/truncated to the output word length. This guard bit technique allows summation of many
products without overflow, improving accuracy significantly over rounding each product
individually.