BMEN90033 · Week 7
Back
BMEN90033 · WEEK 7 · BIOINSTRUMENTATION

High-order filters.

The first-order RC filter is a pedagogical starting point, not an engineering endpoint. Bioinstrumentation front-ends for electrocardiography, pulse oximetry, and electroencephalography routinely require approximations to the ideal low-pass response that are sharper in the transition band, flatter in the passband, or deeper in the stopband than a single pole can provide. In what follows we present a comparative study of five canonical analog low-pass designs, relating the distribution of poles and zeros in the $s$-plane to the resulting magnitude response and to the steady-state behaviour of each filter under a common input.

1st-order Butterworth Chebyshev I & II Elliptic
scroll
01why go beyond first order

The limitations of the first-order response.

The elementary single-pole low-pass filter, realised as a series resistor followed by a shunt capacitor, has the transfer function

$$H_1(s) = \frac{\omega_c}{s + \omega_c}.$$

It possesses a single pole at $s = -\omega_c$ and no finite zeros. Asymptotically, the magnitude response attenuates at $20$ dB per decade above the cutoff — a single order of magnitude reduction for each decade of frequency. This roll-off rate is often inadequate in practice. When an out-of-band interferer such as the $50$ Hz mains fundamental lies within a factor of two of the intended signal band, a first-order response provides negligible rejection.

Increasing the slope of the transition band requires additional poles: an $N^{\text{th}}$-order filter attenuates at $N \cdot 20$ dB per decade in the asymptotic stopband. However, the placement of $N$ poles in the left half-plane is not unique, and the resulting degrees of freedom admit a continuum of possible responses. The classical filter families — Butterworth, Chebyshev type I, Chebyshev type II, and elliptic — correspond to the optimal solutions of four distinct approximation problems, each minimising a different measure of deviation from the ideal brick-wall response.

Four competing design objectives. Any higher-order approximation must balance passband flatness, stopband attenuation, transition-band steepness, and phase linearity. Improvement in any one criterion is purchased at the expense of at least one of the others, and the selection of a filter family is therefore equivalent to a declaration of which approximation error is tolerable for the application at hand.
first-order · bode magnitude
02the butterworth filter

The maximally flat approximation.

The Butterworth filter is defined by the requirement that the magnitude response be maximally flat at the origin: the first $2N-1$ derivatives of $|H(j\omega)|^2$ with respect to $\omega$ vanish at $\omega = 0$. The unique rational approximation satisfying this constraint places all $N$ poles on a semicircle of radius $\omega_c$ in the left half-plane, uniformly spaced by $\pi/N$ radians.

$$|H(j\omega)|^2 = \frac{1}{1 + (\omega/\omega_c)^{2N}}$$

The transfer function contains no finite zeros and all poles are confined to the unit circle when $\omega_c = 1$. The magnitude response is strictly monotonically decreasing on $[0, \infty)$, passes through the half-power point ($-3$ dB) at $\omega = \omega_c$ for every order $N$, and exhibits an asymptotic stopband slope of $N \cdot 20$ dB per decade.

Pole placement

For normalised cutoff $\omega_c = 1$, the poles are located at

$$p_k = \exp\!\left(j\,\frac{\pi(2k + N - 1)}{2N}\right), \quad k = 1,\dots,N.$$

For $N = 4$ the poles occupy angular positions of $112.5^\circ, 157.5^\circ, 202.5^\circ$, and $247.5^\circ$, arranged symmetrically about the negative real axis and bounded away from the imaginary axis. This symmetry is a direct consequence of the flatness constraint.

Applicability. The Butterworth response is the preferred choice when passband fidelity is paramount and no ripple can be tolerated. Typical applications include anti-aliasing prior to analog-to-digital conversion, pre-integration smoothing, and any signal chain in which subsequent processing is sensitive to amplitude distortion within the passband. The associated cost is a comparatively shallow transition band relative to Chebyshev and elliptic designs of equivalent order.
s-plane · butterworth
bode · butterworth vs 1st-order
|H(s)| surface · butterworth — jω trace highlighted (drag to rotate · dbl-click resets)
03chebyshev type i

Equiripple approximation in the passband.

If a bounded, known variation of the passband magnitude can be accepted, a substantially steeper transition band may be obtained for the same filter order. The Chebyshev type I design embodies this trade: it is the rational function of order $N$ that minimises the maximum passband error subject to a monotonically decreasing stopband response. The resulting poles lie on an ellipse in the left half-plane rather than a circle.

$$|H(j\omega)|^2 = \frac{1}{1 + \varepsilon^2\, T_N^2(\omega/\omega_c)}$$

Here $T_N$ denotes the Chebyshev polynomial of the first kind of degree $N$, and $\varepsilon = \sqrt{10^{R_p/10} - 1}$ is chosen to yield a prescribed passband ripple $R_p$ in decibels. Because $T_N$ oscillates between $\pm 1$ on the interval $[-1, 1]$, the passband magnitude equi-ripples between unity and $1/\sqrt{1 + \varepsilon^2}$. Outside the passband $T_N(\omega/\omega_c)$ grows as $2^{N-1}(\omega/\omega_c)^N$, which accounts for the steeper roll-off relative to the Butterworth response.

Pole placement

Let $v_0 = N^{-1}\sinh^{-1}(1/\varepsilon)$. The poles of the type I design are given by

$$p_k = -\sinh(v_0)\sin\theta_k + j\cosh(v_0)\cos\theta_k, \quad \theta_k = \frac{(2k-1)\pi}{2N}.$$

The locus is an ellipse with semi-minor axis $\sinh v_0$ along the real direction and semi-major axis $\cosh v_0$ along the imaginary direction. Increasing the admissible ripple decreases $v_0$, flattens the ellipse, and translates the poles toward the imaginary axis, thereby sharpening the transition band at the cost of increased passband variation.

Applicability. The Chebyshev type I response is appropriate where a specified, bounded passband ripple (typically fractions of a decibel to a few decibels) is acceptable but a gradual transition is not. It is commonly encountered in anti-aliasing stages whose sampling rate lies close to the signal band and in applications requiring efficient use of filter order.
s-plane · chebyshev i
bode · chebyshev i vs butterworth
|H(s)| surface · chebyshev i — jω trace highlighted (drag to rotate · dbl-click resets)
04chebyshev type ii

The introduction of finite zeros.

The inverse approximation problem exchanges the roles of the two frequency bands: the passband is required to be monotonically smooth, while equi-ripple behaviour is admitted in the stopband. The resulting filter is the Chebyshev type II, also termed the inverse Chebyshev design. It is the first of the families considered here whose transfer function contains finite transmission zeros — located on the imaginary axis — producing deep, localised notches in the stopband.

$$|H(j\omega)|^2 = \frac{1}{1 + \left[\varepsilon^2\, T_N^2(\omega_s/\omega)\right]^{-1}}$$

Whereas the Chebyshev polynomial in the type I response oscillates over the passband, its argument in the type II formulation is inverted so that the oscillation is mapped onto the stopband. At each zero of $T_N(\omega_s/\omega)$ the magnitude vanishes exactly, producing the transmission zeros that are observed as sharp nulls in the magnitude response.

Zero placement

$$z_k = \pm \frac{j\,\omega_s}{\cos\theta_k}, \quad \theta_k = \frac{(2k-1)\pi}{2N}, \quad k = 1,\dots,\lfloor N/2\rfloor.$$

For $N = 4$ with stopband edge $\omega_s = 1.5\,\omega_c$, the zeros are located at approximately $\pm j\,1.62\,\omega_c$ and $\pm j\,3.92\,\omega_c$, situated within the stopband so as to maximise local attenuation.

Applicability. The Chebyshev type II response is indicated when passband fidelity is required and specific narrow-band interferers must be attenuated. Typical examples include the rejection of mains harmonics at 50 or 60 Hz, suppression of switching-converter noise, and the removal of coherent carrier components. Placement of the transmission zeros permits targeted attenuation at frequencies of known interference.
s-plane · chebyshev ii
bode · chebyshev ii vs butterworth
|H(s)| surface · chebyshev ii — jω trace highlighted (drag to rotate · dbl-click resets)
05the elliptic filter

Equiripple in both bands.

The elliptic, or Cauer, filter is the rational approximation obtained when equi-ripple behaviour is permitted in both the passband and the stopband simultaneously. For a given filter order, it achieves the steepest transition band of any rational transfer function, and is optimal in the minimax sense over both bands. Its poles lie on an ellipse, as in the Chebyshev type I design, while its zeros are distributed along the imaginary axis, as in the Chebyshev type II design. Both sets of singularities are determined by the Jacobian elliptic functions, from which the family derives its name.

$$|H(j\omega)|^2 = \frac{1}{1 + \varepsilon^2\, R_N^2(\omega/\omega_c,\, L)}$$

Here $R_N(\cdot, L)$ denotes the elliptic rational function of degree $N$, parameterised by a selectivity factor $L$ that fixes the ratio of stopband to passband edge frequencies. For the design illustrated here ($N = 4$, $R_p = 1$ dB, $R_s = 40$ dB), the transition from passband to stopband is completed within approximately $1.1\,\omega_c$; a Butterworth filter of the same order requires a further $6$ dB of attenuation at this frequency to reach comparable stopband levels.

Singularity structure

Applicability. The elliptic family is the standard choice where a near-ideal brick-wall magnitude response is required: anti-aliasing stages ahead of delta-sigma converters, channel-selection filters in narrow-band radio-frequency receivers, and output reconstruction in Class-D audio amplifiers. The principal disadvantage is pronounced phase nonlinearity near the band edges, which can produce substantial group-delay distortion. Applications in which waveform morphology is diagnostic—for instance, clinical electrocardiography and oscilloscope front-ends—generally require a design with a more linear phase response.
s-plane · elliptic
bode · elliptic vs butterworth
|H(s)| surface · elliptic — jω trace highlighted (drag to rotate · dbl-click resets)
06side by side

Comparative response to a common input.

The input applied to all five filters is the sum of a low-frequency component situated within the passband and a high-frequency component situated beyond the cutoff. An ideal low-pass filter preserves the former and eliminates the latter; the deviation of each design from this ideal is evident in the five output panels below. All filters share a common normalised cutoff, and each output is computed from the steady-state sinusoidal response $y(t) = |H(j\omega)|\sin(\omega t + \arg H(j\omega))$ applied to each input component independently.

0.30 ωc
2.00 ωc
0.60
4
low-frequency signal (pass) high-frequency interferer (reject) filter output
input signal · x(t) = sin(ω₁t) + A sin(ω₂t)
1st-order
butterworth
chebyshev i
chebyshev ii
elliptic
all five · bode magnitude overlay
Interpretation of the traces. The faint blue envelope in each panel denotes the target low-frequency component. The elliptic and Chebyshev type II outputs exhibit the closest agreement with this reference, reflecting their superior stopband attenuation at the interferer frequency. The Chebyshev type I response attenuates the interferer comparably but introduces visible amplitude variation within the passband. The Butterworth output is free of passband distortion but retains a greater fraction of the interferer. The first-order response provides negligible rejection. The Bode magnitude overlay at the bottom of the figure relates these time-domain observations to the underlying frequency-domain characteristics.