Questions: Filter Order, Rolloff Rate, and Transition Band

Question 1

A client specifies a filter that must pass all signals below 1 kHz and reject all signals above 1.001 kHz — a transition band 1000 times narrower than a filter with a 1 kHz to 2 kHz transition. What does this demand in terms of filter order?

• A. Roughly the same order, since the passband edge frequency is identical in both cases
• B. A dramatically higher order — because the minimum order scales with log(Ωs/Ωp) in the denominator, halving the transition band nearly doubles the required order, so the extremely narrow band requires far more poles
• C. Only a different filter family such as Chebyshev rather than Butterworth, with no change in order
• D. Fewer poles, since a narrow transition band is easier to achieve in IIR designs

Answer: (B)

The minimum filter order formula has log(Ωs/Ωp) in the denominator. Moving the stopband edge closer to the passband edge (narrowing the transition band) shrinks this log term, forcing the required order up. A transition band ratio of 2 (Ωs/Ωp = 2) might need 4th order; a ratio of 1.001 would require orders of magnitude more poles for the same attenuation. This is why narrow transition bands are extremely expensive in terms of filter complexity.

Question 2

An engineer implements a high-order IIR filter using fixed-point arithmetic and finds that the output contains spurious oscillations not present in the floating-point prototype. What is the most likely cause and remedy?

• A. The filter order is too high for any implementation; the order must be reduced
• B. Coefficient quantization in fixed-point shifts poles and zeros from their designed locations — the standard remedy is to implement the filter as a cascade of second-order sections (biquads), which are far less sensitive to quantization errors
• C. Fixed-point arithmetic cannot represent negative filter coefficients, requiring a different filter family
• D. The sampling rate must be increased to accommodate higher-order filters

Answer: (B)

High-order IIR filters implemented directly in fixed-point arithmetic are prone to instability from coefficient quantization — small rounding errors in coefficients can move poles outside the unit circle. Cascading second-order sections (biquads) solves this because each biquad has only 5 coefficients and its poles are numerically well-separated, making them much less sensitive to quantization. This is why cascaded biquads are the standard implementation form for high-order IIR filters in fixed-point systems.

Question 3

An 8th-order Butterworth filter rolls off at approximately 160 dB/decade in the stopband.

• T. True
• F. False

Answer: True

The asymptotic rolloff rate for an Nth-order Butterworth filter is 20N dB/decade. For N = 8, this gives 20 × 8 = 160 dB/decade. Equivalently, this is 8 × 6 = 48 dB/octave. The 20N dB/decade rule is specific to Butterworth designs; other filter families achieve different (often steeper) rolloffs at the same order by accepting ripple.

Question 4

A higher-order filter generally produces better performance than a lower-order filter for any given application.

• T. True
• F. False

Answer: False

Higher order provides a sharper transition band, which is sometimes necessary — but it comes with real costs. High-order IIR filters are more sensitive to coefficient quantization and can become unstable in fixed-point arithmetic. They introduce more phase distortion (group delay variation), which can degrade performance in communications and audio applications. In analog hardware, each additional pole requires an additional reactive component (inductor or capacitor), adding cost and size. Higher order is only better if the sharper transition is actually required and the additional complexity can be managed.

Question 5

Why does halving the transition band width require roughly double the filter order, rather than a proportionally small increase in order?

Model answer: The minimum filter order is proportional to 1/log(Ωs/Ωp), where Ωs is the stopband edge and Ωp is the passband edge. When the transition band is halved, Ωs moves closer to Ωp, shrinking the ratio Ωs/Ωp. Because of the logarithm, the change in the denominator is roughly proportional to the change in the frequency ratio — halving the transition band approximately doubles 1/log(Ωs/Ωp), and therefore approximately doubles the required order. This logarithmic sensitivity means that extremely narrow transition bands require very high orders.

This is a fundamental constraint, not a limitation of any specific filter family — all linear filter families face it. The Chebyshev and elliptic families can achieve lower order for a given transition band compared to Butterworth (by tolerating ripple), but they cannot escape the underlying logarithmic relationship between order and transition band width.