Common Digital Filter Types

Why This Matters

Digital filters form the backbone of nearly every signal processing system you'll encounter, from audio engineering and telecommunications to biomedical instrumentation and radar. When you're tested on filter design, you're really being evaluated on your understanding of tradeoffs: stability versus computational efficiency, sharp transitions versus phase distortion, and implementation complexity versus performance. These tradeoffs determine whether a real-time system can meet its latency requirements or whether a filtered signal retains its essential characteristics.

The filters in this guide fall into distinct categories based on structure (how they compute outputs), frequency selectivity (which frequencies they pass or reject), and approximation method (how they achieve their frequency response). Don't just memorize filter names. Know what problem each filter solves, what tradeoffs it accepts, and when you'd choose one over another. If an exam question asks you to "design a filter for X application," your job is to match the application's constraints to the right filter architecture.

---

Filter Structures: FIR vs. IIR

The most fundamental distinction in digital filter design is whether the filter uses feedback. This structural choice affects everything from stability guarantees to computational requirements and phase behavior.

Finite Impulse Response (FIR) Filters

• No feedback architecture. The output depends only on current and past input samples: $y[n] = \sum_{k=0}^{M} b_k \, x[n-k]$
• Guaranteed BIBO stability since all poles sit at the origin (inside the unit circle by definition). FIR filters can also achieve linear phase when coefficients are symmetric or antisymmetric, meaning all frequency components experience equal group delay. This is critical for preserving waveform shape.
• Higher computational cost for sharp transitions. An FIR filter typically requires many more coefficients (taps) than an IIR design to achieve comparable roll-off steepness.

Infinite Impulse Response (IIR) Filters

• Feedback architecture. The output depends on past outputs as well as inputs: $y[n] = \sum_{k=0}^{M} b_k \, x[n-k] - \sum_{k=1}^{N} a_k \, y[n-k]$
• Computational efficiency is the main advantage. Sharp frequency responses are achievable with far fewer coefficients, making IIR filters preferred for resource-constrained real-time systems.
• Stability is not guaranteed. All poles of the transfer function $H(z)$ must lie strictly inside the unit circle in the z-plane. If any pole lands on or outside the unit circle, the filter will oscillate or diverge. Quantization effects in fixed-point implementations can shift poles and degrade stability, so you need to account for finite word-length effects during design.
• Nonlinear phase is inherent to most IIR designs. The feedback structure introduces frequency-dependent group delay, which distorts the shape of time-domain waveforms even though it doesn't alter their spectral content.

---

Frequency Selectivity: Shaping the Spectrum

These classifications describe what frequencies a filter passes or rejects, independent of whether the implementation is FIR or IIR. Understanding selectivity is essential for matching filters to application requirements.

Low-Pass Filters

• Passes frequencies below the cutoff $f_c$ while attenuating higher frequencies. This is the most common filter type in practice.
• Applications include anti-aliasing before ADC conversion (where you must remove content above $f_s/2$ to satisfy the Nyquist criterion), smoothing noisy sensor data, and extracting baseband signals in communication receivers.
• Transition bandwidth, the region between the passband edge and the stopband edge, determines how "sharp" the filter appears. Narrowing this transition requires increasing the filter order, which costs more computation.

High-Pass Filters

• Passes frequencies above the cutoff $f_c$ while attenuating lower frequencies. Spectrally, it's the complement of a low-pass filter.
• Removes DC offset and low-frequency drift. This is essential in AC-coupled measurements and baseline wander correction in ECG signals, where slow electrode drift would otherwise obscure the cardiac waveform.
• Can be derived from low-pass designs using spectral transformation techniques (e.g., the substitution $z^{-1} \rightarrow -z^{-1}$ for FIR, or analog-domain frequency transformations for IIR), which simplifies the design process considerably.

Band-Pass Filters

• Passes frequencies within a range defined by lower cutoff $f_L$ and upper cutoff $f_H$, attenuating everything outside.
• Center frequency $f_0 = \sqrt{f_L \cdot f_H}$ (geometric mean) and bandwidth $BW = f_H - f_L$ are the key design parameters. The quality factor $Q = f_0 / BW$ quantifies selectivity: higher Q means a narrower passband relative to center frequency.
• Essential for frequency-division multiplexing (FDM) where multiple signals share a medium by occupying different frequency bands. Each channel is isolated by a band-pass filter tuned to its allocated band.

Band-Stop (Notch) Filters

• Attenuates frequencies within a specific range while passing everything else. This is the inverse of a band-pass filter.
• Eliminates interference at known frequencies, such as 50/60 Hz power line hum or specific harmonic distortion components. In biomedical signal processing, notch filters at the mains frequency are nearly universal.
• Narrow notch filters require high Q-factor designs. These are often implemented as cascaded second-order IIR sections (biquads) for both efficiency and numerical stability, since a single high-order section is more sensitive to coefficient quantization.

All-Pass Filters

• Unity magnitude response at all frequencies: $|H(e^{j\omega})| = 1$ for all $\omega$. The filter passes everything but modifies the phase relationship between frequency components.
• Phase equalization is the primary application. All-pass sections can compensate for phase distortion introduced by other system components or transmission channels, flattening the overall group delay.
• Group delay manipulation enables signal alignment in multi-path systems and is also used in audio effects like phasers, where frequency-dependent phase shifts create the characteristic sweeping sound.

---

Approximation Methods: Achieving the Response

These filter types describe how a desired frequency response is mathematically approximated. Each method accepts different tradeoffs between passband flatness, stopband attenuation, transition sharpness, and phase linearity. All three below are classical analog prototype methods, typically converted to IIR digital filters via the bilinear transform or impulse invariance.

Butterworth Filters

• Maximally flat passband with no ripple. The squared magnitude response is: $|H(j\omega)|^2 = \frac{1}{1 + \left(\omega/\omega_c\right)^{2N}}$ where $N$ is the filter order.
• Roll-off rate is $20N$ dB/decade (equivalently $6N$ dB/octave). This is the most gradual roll-off among the three classical approximations for a given order, so you'll need higher $N$ to achieve sharp transitions.
• Best choice when passband flatness is the priority and you can tolerate a wider transition band. Common in audio applications where amplitude ripple would be audible.

Chebyshev Filters

• Steeper roll-off than Butterworth for the same order, achieved by allowing controlled ripple in either the passband or the stopband.
• Type I permits equiripple in the passband with a monotonically decreasing stopband. Type II (also called inverse Chebyshev) has a monotonically flat passband with equiripple in the stopband.
• Ripple amplitude (specified in dB) is a design parameter you choose. More ripple tolerance yields a sharper transition for the same filter order. For example, allowing 0.5 dB of passband ripple in a Type I design can save you one or two orders compared to Butterworth for the same transition width.

Elliptic (Cauer) Filters

• Sharpest possible transition for a given filter order among all classical approximation methods. This is a provable optimality result: no causal, stable filter of the same order can have a narrower transition band.
• Equiripple in both passband and stopband. You accept ripple everywhere in exchange for minimizing the transition bandwidth.
• Preferred for strict frequency separation requirements where computational resources limit filter order, such as channelization in software-defined radio or anti-aliasing filters with tight spectral masks.
• The tradeoff is that elliptic filters have the most nonlinear phase response of the three, which matters if waveform fidelity is important.

---

Quick Reference Table

---

Self-Check Questions

1. You need to filter a biomedical signal where preserving the exact timing relationship between waveform features is critical. Would you choose FIR or IIR, and why?

2. Compare Butterworth and Chebyshev Type I filters: what tradeoff does each make, and in what application would you prefer one over the other?

3. A software-defined radio must separate two adjacent channels with minimal guard band. Which approximation method provides the sharpest transition for a given filter order?

4. What structural property distinguishes IIR filters from FIR filters, and what design constraint does this property impose?

5. You're designing a system to remove 60 Hz power line interference from an audio recording while preserving the rest of the spectrum. What filter selectivity type would you use, and what implementation consideration affects how narrow you can make the rejection band?

6. An IIR filter designed with double-precision coefficients is being ported to a fixed-point DSP. What stability concern does this raise, and which filter structure (direct form, cascade biquads, etc.) is more robust to this problem?