Passive and Active Filters — Part One

Coffee drips from a paper filter into a carafe beside a circuit-patterned mug and a tray of loose passive components.

Filters!

Few subjects in electronics are as rich or as fascinating, yet for many people filtering still boils down to a memorised formula for an RC section whose significance remains unclear. For others, it immediately conjures up an abstract avalanche of transfer functions, poles, zeros and Laplace transforms.

Behind that formalism, however, lies the art of shaping signals. This four-part series aims to demystify filtering and take you through the actual engineering reasoning, step by step—from a simple resistor dissipating heat to universal active filters and sophisticated audio equalisation.


What is a filter?

Everyone has used a filter in everyday life. Perhaps the most familiar example is the filter in your morning coffee maker: it mechanically retains the insoluble grounds while allowing the precious dark liquid to flow through.

Electronics constantly presents us with the same need to separate things. Extend an antenna into the air or connect an audio cable, and thousands of signals coexist: the carriers conveying a radio presenter’s voice, 50 Hz mains hum, interference from a nearby switching power supply, a neighbour’s gas lighter, a teenager’s modified moped and thermal noise. An electronic filter acts as a selective gatekeeper: it lets the wanted frequencies through and rejects unwanted ones.

Analogy between a coffee filter and an electronic filter
Fig. 1 — Analogy between a coffee filter and an electronic filter

But this selection never works like a magical on/off switch. In the real world of analogue components, there is no razor-sharp boundary at which a 999 Hz signal passes unchanged while a 1 001 Hz signal suddenly disappears. An electronic filter does not cut abruptly: it attenuates gradually. Between the region where the signal passes easily (the passband) and the region where it is heavily suppressed (the stopband), there is always a continuous transition region. Its slope falls more or less steeply depending on the circuit architecture, complexity and number of stages.

An electronic filter therefore shapes a signal by changing its amplitude (the size of the voltage, through its gain) and its phase shift (its timing relative to the input) as a function of frequency.

Fundamental rule: a linear filter changes amplitude and phase continuously
Fig. 2 — Fundamental rule: a linear filter changes amplitude and phase continuously

Filters are not confined to specialised or mysterious equipment. They are already at the heart of most ordinary electronic circuits, often even when nobody deliberately set out to build a “filter”!

Consider three everyday examples:

  • At the input of an audio amplifier: to prevent an unwanted DC voltage (coming from an audio source or sensor) from saturating the amplifier, we insert a coupling capacitor in series (which acts as a watertight barrier blocking DC current while letting sound-wave signals pass). Immediately after it, a bias resistor connects the circuit to ground to establish a stable quiescent potential. Without having planned it, the combination of this capacitor and resistor forms a natural high-pass filter that removes the signal’s DC component and lets the music through!
  • In a laboratory power supply: after transforming and rectifying (or high-frequency switching) the 230 V mains voltage, the resulting waveform is still chopped and undulating. Capacitors then act as energy reservoirs and form a low-pass filter that strongly reduces the ripple. The resulting voltage is never perfectly smooth: residual ripple remains, to which noise and transient disturbances may be added; regulation placed after the filter can attenuate these further.
  • In a timer or flashing-light circuit: the resistor slows the current and the capacitor takes a certain time to fill. This charging time acts as an electronic hourglass to create a waiting delay (for example, switching off a ceiling light after 30 seconds or setting the tempo of a flashing light). This filling rate is measured by the time constant, traditionally denoted by the lower-case Greek letter tau (\tau), equal to the direct product \tau = R \times C (expressed in seconds when R is in ohms and C in farads). It is precisely this time rhythm \tau that will later set the cutoff frequencies of some of our filters!

In other words, anyone who connects components is already working with filters, whether they realise it or not. Let us begin with their four main forms.


The four main magnitude responses—or the different ways to select frequencies—and the selective phase shifter

First, what exactly do we mean by a spectrum?

The word comes from the Latin spectrum, meaning an “appearance” or “apparition”. In the seventeenth century, Isaac Newton (Newton, 1672) used it in optics to describe the elongated coloured image produced when a prism separates a beam of white light. His text is dated 6 February 1671/72 under the English calendar then in use and was published on 19 February 1671/72 in the Philosophical Transactions—in 1672 by modern year numbering. The term was subsequently extended by analogy to sound, electrical signals and many other physical phenomena.

In electronics, a signal’s frequency spectrum represents its constituent frequencies and their relative contributions:

  • An ideal sinusoid of frequency f_0 contains only one frequency component: in the usual positive-frequency representation, its spectrum has a line at f_0.
  • A complex periodic waveform, such as a square wave, contains a fundamental frequency accompanied by harmonics.
  • A voice, a real musical sound or noise distributes its energy over a narrower or wider range of frequencies.
  • Any DC component, equal to the signal’s average value, is represented by a line at 0\ \text{Hz}.
Understanding the DC component at 0 Hz in a spectrum and how components respond to it
Fig. 3 — Understanding the DC component at 0 Hz in a spectrum and how components respond to it

As an initial picture, imagine an ideal keyboard whose keys each produce just one frequency. The spectrum would show which keys are being pressed and how strongly. The analogy has limits, however: a real piano note itself contains a fundamental and many harmonics.

Understanding the fundamental, harmonics and the origin of musical timbre
Fig. 4 — Understanding the fundamental, harmonics and the origin of musical timbre

The spectrum lets us view a signal from a different perspective. Instead of watching its voltage evolve over time on an oscilloscope, we discover the “recipe” of frequencies that make it up.

Depending on which region of the spectrum we want to preserve or attenuate, we distinguish four main magnitude responses:

  1. A low-pass filter passes DC and low frequencies with little attenuation, then progressively attenuates higher frequencies beyond its transition region. Typical uses include smoothing and reducing high-frequency interference.
  2. A high-pass filter strongly attenuates DC and low frequencies while passing higher frequencies with little attenuation. It is used to remove a DC component, couple two stages or reduce very slow drift.
  3. A band-pass filter favours a range between a lower frequency f_1 and an upper frequency f_2, attenuating frequencies on either side. In a radio receiver, a band-pass filter helps select the wanted channel.
  4. A band-stop filter, also called a band-reject filter, does the opposite: it attenuates a specified range while passing lower and higher frequencies. When the rejected band is very narrow, it is called a notch filter. For example, it can reduce interference at 50\ \text{Hz} or 60\ \text{Hz} while limiting its effect on the rest of the signal.

The transition between bands depends in particular on the filter order: the degree of the denominator of its transfer function after simplification. For a conventional low-pass filter without compensating zeros, each additional pole increases the asymptotic attenuation slope by 20 dB/decade.

A distinct fifth response: the all-pass filter, or phase-shifting network

There is a fifth, apparently paradoxical family: the all-pass filter, or phase-shifting network. Unlike the previous four, it rejects no frequencies and does not change the signal’s amplitude. Why call it a filter? Because its phase varies with frequency: it shifts some frequency components in time relative to others, without attenuating any of them.

In a multiway loudspeaker system, for example, the drivers and crossover filters introduce different phase shifts. An all-pass network can correct their timing alignment around the crossover frequency. All-pass networks also appear in phaser effects and some quadrature oscillators.

We will examine its equations and phase response later in the article, once the necessary tools are in place.

Different ways to implement the same filtering requirement

Given a filter specification, an engineer has three main technological approaches:

  • A passive filter consists entirely of components with no internal energy source: resistors, capacitors, inductors and transformers. It requires no auxiliary supply, can handle very high power—although magnetic-core inductors can saturate at high current—and needs no semiconductor devices. However, it cannot provide any power gain (a resonance can locally raise the voltage, but never creates energy), and its inductors become bulky at low frequencies.
  • An active filter combines resistors and capacitors with an amplifying device, such as a transistor or operational amplifier. It can implement many low-frequency filters without inductors, provide adjustable gain, and offer very high input impedance together with nearly zero output impedance.
  • A digital filter samples and digitises the analogue signal using an analogue-to-digital converter (ADC), then processes it through mathematical equations in a microcontroller, digital signal processor (DSP) or field-programmable gate array (FPGA), before converting it back through a digital-to-analogue converter (DAC).

This series is dedicated to the art of filtering: we begin here with physical principles and passive circuits (Part 1), before exploring active op-amp structures (Part 2), high-order polynomial approximations, switched-capacitor filters and digital filters (Part 3), and then high-fidelity audio applications (Part 4).

One input, one output and a transfer function

💡 Reading note — Mathematical formalism
This section establishes the rigorous theoretical and mathematical framework of filtering (the Laplace transform, the operational variable p, the transfer function, complex-impedance calculations, poles, zeros and the 3D surface). It is not strictly necessary to understand how practical circuits behave on the bench, but it forms the foundational bedrock of signal processing and belongs naturally here in the logical progression.

If you prefer to start with practical circuit behaviour, you can skip ahead without losing the thread: The first-order low-pass filter: the RC section.

Throughout this article, we work within the framework of linear time-invariant (LTI) systems. Why is this assumption so valuable? Because thanks to Fourier’s theorem (Fourier, 1822; see the appendix at the end of the article), any complex signal—whether a human voice, a cymbal strike or a square wave—can be decomposed into a sum of pure sinusoids. And the fundamental property of a linear system is this: a sinusoid that enters at a frequency f comes out at the same frequency f, simply attenuated or amplified and possibly shifted in time. The filter creates no new frequencies. Direct consequence: if we know how the filter treats each individual sinusoid, we know how it treats any signal, since any signal is simply a sum of sinusoids.

This is precisely what ceases to hold in a nonlinear system (a saturating amplifier, a diode or a transistor outside its normal operating region): there, a pure 440 Hz sinusoid can generate harmonics at 880 Hz, 1 320 Hz and so on that did not exist in the input signal—this is harmonic distortion. In a linear system, this phenomenon is by definition impossible.

The complete behaviour of the filter is summarised by its transfer function H(p) (the formula showing how the circuit transforms amplitude and phase), which relates the output voltage V_s (subscript s for French sortie, output) to the input voltage V_e (subscript e for French entrée, input) in the operational domain:

H(p)=\frac{V_s(p)}{V_e(p)} \qquad\Longleftrightarrow\qquad V_s(p) = H(p) \cdot V_e(p)
Canonical diagram of a linear two-port network and its transfer function H(p)
Fig. 5 — Canonical diagram of a linear two-port network and its transfer function H(p)

In other words, H(p) is the ratio of “what comes out” to “what goes in”. In English-language literature, the same quantities are denoted V_{\text{in}} (input) and V_{\text{out}} (output), or sometimes V_I and V_O—take care not to confuse the I of Input with current I, nor the O of Output with the digit zero.

In many French-language textbooks and engineering programmes—notably in France (BTS, BUT, engineering schools and control engineering)—the complex Laplace variable is traditionally denoted p. English-language literature generally uses s, a notation also increasingly encountered in simulation software (SPICE, MATLAB) and translated technical documents. Why use the Laplace transform? Because it turns differential calculus (the equations describing how a quantity varies from instant to instant under the action of capacitors and inductors) into simple algebra: with zero initial conditions, differentiating a signal with respect to time (measuring its instantaneous rate of change) simply amounts to multiplying it by p (\mathcal{L}\{\mathrm{d}x/\mathrm{d}t\} = pX(p)), and integrating it (accumulating its effect over time) amounts to dividing it by p (\mathcal{L}\{\int_0^t x(\tau)\,\mathrm{d}\tau\} = X(p)/p)!

To obtain the circuit’s response in sinusoidal steady state (steady state means that the signal has been oscillating long enough for any transient phase to have died out) at angular frequency \omega, we evaluate its transfer function on the imaginary axis by setting p = j\omega, where:

  • j represents the imaginary unit (j^2 = -1), (written i in mathematics but j in electronics, to avoid any disastrous confusion with instantaneous electric current i(t));
  • \omega (lower-case Greek letter omega) is the angular frequency, expressed in radians per second (\text{rad/s}), directly related to frequency f (in hertz, or cycles per second) by:
\omega=2\pi f

For a steady-state sinusoid of angular frequency \omega > 0, the ideal integration operator 1/p becomes 1/(j\omega) = -j/\omega: the amplitude of the result is divided by angular frequency \omega and its phase undergoes a constant phase shift of -90^\circ (a quarter-period lag, since 1/j = -j = e^{-j90^\circ}). Likewise, the ideal differentiation operator p becomes j\omega: the amplitude is multiplied by \omega and the phase undergoes a phase shift of +90^\circ (a quarter-period lead, since j = e^{+j90^\circ}). These \pm 90^\circ phase shifts are strict and constant for pure mathematical operators and ideal reactive components (C and L); in a complete filter combining resistors and capacitors (such as our RC section), the presence of the resistor makes the phase vary continuously and progressively from 0^\circ to -90^\circ as frequency increases.

A linear filter changes amplitude and phase while preserving frequency
Fig. 6 — A linear filter changes amplitude and phase while preserving frequency

This is precisely how we establish the expressions for complex impedances, traditionally denoted by the capital letter Z (impedance being the generalised resistance a component presents to alternating current, indexed by the component type: Z_C for a capacitor, Z_L for an inductor, and Z_R = R for a resistor):

Z_C(p) = \frac{1}{pC} \quad\Longrightarrow\quad Z_C(j\omega) = \frac{1}{jC\omega} = -\frac{j}{C\omega}
Z_L(p) = pL \quad\Longrightarrow\quad Z_L(j\omega) = jL\omega

(At zero frequency \omega = 0, however, the factor 1/(j\omega) tends to infinity. Integrating a DC component 0\ \text{Hz} does not produce a steady-state sinusoid but a ramp that grows indefinitely with time. This is the drift problem of the ideal integrator, which we will meet later.)

Using angular frequency \omega makes these impedance expressions much simpler by avoiding a factor of 2\pi in every calculation.

Do not confuse resistance with impedance:

  • A pure resistance R (in ohms \Omega) opposes the flow of current in the same way regardless of frequency: it irreversibly dissipates electrical energy as heat through the Joule effect (P = R I^2) and produces no phase shift between voltage and current.
  • Complex impedance Z = R + jX (in ohms \Omega) is the general concept covering all components. Its imaginary part X is called reactance (opposition to alternating current that stores and returns energy without dissipating it):
    • Capacitive reactance of a capacitor (X_C = -\dfrac{1}{C\omega} < 0): it stores energy in an electrostatic field and decreases in magnitude as frequency increases (90^\circ voltage lag);
    • Inductive reactance of an inductor (X_L = L\omega > 0): it stores energy in a magnetic field and increases proportionally with frequency (90^\circ voltage lead).

Unlike resistance, which consumes electrical energy, capacitive and inductive reactances periodically exchange energy with the circuit. Every resistance is a particular impedance with zero reactance X = 0, but not every impedance is a simple resistance.

We then obtain the harmonic frequency response (how the filter “responds”—in amplitude and phase shift—when driven by a sinusoid at each frequency):

H(j\omega)=|H(j\omega)|e^{j\varphi(\omega)}

This equation contains two essential pieces of information:

  • The magnitude |H(j\omega)|—the “size” of the complex number—gives the voltage gain: how much the filter attenuates or amplifies the sinusoid.
  • The argument \varphi(\omega), the Greek letter phi—the angle of the complex number—is the phase shift: the output’s lead (\varphi > 0) or lag (\varphi < 0) relative to the input.

To conveniently represent huge amplitude variations (from 1 to 1 000 000) and match the logarithmic sensitivity of the human ear (the Weber–Fechner law and auditory psychoacoustics: to double the perceived loudness, the acoustic power must be multiplied by roughly ten, or about +10\ \text{dB}—a fascinating subject we will explore in the loudness article), gain is traditionally expressed in decibels (\text{dB}):

G_{\mathrm{dB}}(\omega)=20\log_{10}|H(j\omega)|

Why the factor 20? Because electrical power is proportional to voltage squared (P = V^2 / R). For voltages compared across identical impedances, or using a constant reference impedance, a magnitude ratio of 1/\sqrt{2} \approx 0.707 corresponds to half the power, giving 10\log_{10}(0.5) = 20\log_{10}(1/\sqrt{2}) \approx -3.01\ \text{dB} in decibels.

This is the -3\ \text{dB} point, sometimes called the half-power point. That second name requires care: reducing voltage by a factor of 1/\sqrt{2} halves the power only when the voltages are compared across the same reference impedance or identical impedances. On a voltage-gain plot, the universally valid statement is simply that the magnitude is 1/\sqrt{2}\approx0.707 of its reference value. The corresponding frequency defines the cutoff frequency (f_c), used as a conventional passband boundary.

The −3 dB point on a low-pass filter's frequency response
Fig. 7 — The −3 dB point on a low-pass filter’s frequency response

(This response curve is an introductory example using a simple low-pass filter to illustrate the cutoff boundary and power ratio. The corresponding circuit, its transfer-function calculation and its phase response will be explained in detail later in this article.)

Filter order, poles and zeros: what the fraction tells us

When engineers calculate a filter’s transfer function H(p)—the output-voltage/input-voltage ratio introduced above—they obtain a mathematical fraction with a numerator on top and a denominator below:

H(p) = \frac{\text{Numerator}(p) \quad (\text{top})}{\text{Denominator}(p) \quad (\text{bottom})}

This fraction contains three fundamental ideas that determine the filter’s entire behaviour:

  1. Filter order: how many independent energy-storage variables are involved.
    The order is the degree of the polynomial in the denominator, after fully simplifying the fraction by cancelling any common factors between numerator and denominator:

    • If the highest power is p^1, or simply p, the filter is first order. Its dynamics involve one independent energy-storage variable, such as the voltage of a capacitor C in an RC section or the current through an inductor L in an RL circuit.
    • If the highest power is p^2, the filter is second order. Its dynamics involve two independent energy-storage variables, such as those of two capacitors in cascaded RC sections, or an inductor L and capacitor C in a resonant RLC circuit.
    • If the highest denominator power is p^3, p^4 \dots, the filter is third order, fourth order, and so on. Such filters are often built by combining first- and second-order sections.

    The number of independent reactive elements sets the maximum order the circuit can achieve. Its effective order may be lower, depending on their interconnection or cancellation between numerator and denominator.

    In a conventional all-pole low-pass filter, whose numerator is constant, each pole adds an asymptotic slope of -20\ \text{dB/decade}, or approximately -6\ \text{dB/octave}—exactly -6.02\ \text{dB/octave}. At sufficiently high frequency, an order-n low-pass filter therefore approaches a slope of -20n\ \text{dB/decade}:

    • First order: -20\ \text{dB/decade}, or approximately -6\ \text{dB/octave}.
    • Second order: -40\ \text{dB/decade}, or approximately -12\ \text{dB/octave}.
    • Third order: -60\ \text{dB/decade}, or approximately -18\ \text{dB/octave}.

    (This asymptotic slope is not reached immediately at the cutoff frequency f_c, but only sufficiently far above the characteristic frequencies of the various poles.)

  2. Poles: where the denominator becomes zero.
    What happens if we find the values of p that make the denominator equal to zero? At the exact point where the denominator vanishes, the function is undefined. As p approaches that point, however, the denominator tends to zero and the magnitude |H(p)| tends to infinity. That point is a pole.
    Why a “3D” picture, and why the word “pole”?
    The Laplace variable p is a complex number with two “ground-level” coordinates: real part \sigma and imaginary part \omega. They form a 2D plane, like a sheet of paper on a table. Add a third dimension above that table—a vertical height representing the gain magnitude |H(p)|—and we obtain a genuine 3D landscape.
    Wherever the denominator vanishes, the surface does not stay flat: it is pulled sharply skyward to form an infinitely high peak. In English, the word pole denotes a ship’s mast, a post or a tent pole. These invisible peaks hold up the entire “canvas” of the circuit’s response, just like the masts under a circus big top!
    And when we test our filter with steady-state sinusoids in the laboratory, we simply walk along the imaginary axis p = j\omega drawn in the horizontal plane—that is, along the line where \sigma = 0: the curve we plot on our sheet is only a slice (a 2D cross-section) through that vast tent canvas. If a pole is planted right next to our path, the canvas is lifted very high and raises a resonance bump; if it is planted far to the left, the canvas is gentle and smooth. It is therefore these poles that decide whether the filter is calm, fast or oscillatory.
  3. Zeros: where the numerator becomes zero.
    Conversely, values of p that make the numerator equal to zero make the whole fraction zero (0 / \text{Denominator} = 0).
    On our 3D big top, the canvas comes right down to touch the ground (forming a funnel or natural hollow descending to zero altitude |H|=0, marked by a small white circle on the ground). When one of these zeros lies precisely on our measurement line (the imaginary axis p = \pm j\omega_0), the input signal at that frequency is completely blocked and nothing comes out of the filter (V_s = 0). These values are quite naturally called zeros. This is the fundamental principle used in band-stop filters to eliminate a specific interfering frequency (such as 50 Hz mains hum). (And if a zero lies at the exact same position as a pole, their effects cancel each other out: the fraction simplifies.)

The illustration below uses the normalised example:

H(p) = \frac{p}{(p+1)^2 + 4}

It has a zero at the origin z_1 = 0. On the frequency axis p = j\omega, this means H(j0) = 0, so DC is blocked. Its two conjugate poles p_{1,2} = -1 \pm j2 produce the two surface peaks.

3D magnitude surface of a transfer function, showing two poles, a zero at the origin and six constant-real-part slices
Fig. 8 — 3D magnitude surface of a transfer function, showing two poles, a zero at the origin and six constant-real-part slices

Reading the surface: six slices to understand the gain

To show how this 3D surface shapes the filter’s response, the diagram includes six parallel slices, each at a constant real part \sigma:

\sigma = 0,\;-0.5,\;-0.8,\;-1,\;-1.2,\;-1.5

These slices show how the profile changes as we move through the complex plane:

  • \sigma = 0, amber slice: the familiar physical case—the sinusoidal frequency response H(j\omega) measured in steady state along the imaginary axis.
  • \sigma = -0.5 and then \sigma = -0.8, violet and magenta slices: moving left brings us closer to the poles. The two peaks begin to rise distinctly above the plane.
  • \sigma = -1, coral-red slice: this slice passes through the real coordinate of both conjugate poles (p_{1,2} = -1 \pm j2). As p approaches either pole, magnitude |H(p)| tends to infinity. At the pole itself the function is undefined, so the plotted peaks are clipped.
  • \sigma = -1.2 and then \sigma = -1.5, green and blue slices: moving farther left, beyond the poles, the surface descends and the peaks gradually flatten.

For the normalised function in the illustration, its magnitude formula makes this geometry clear:

|H(p)| = \frac{|p|}{|p-p_1|\,|p-p_2|}

Each term |p - p_k| is simply the Euclidean distance between the evaluation point p and pole p_k in the complex plane. As that distance decreases, the denominator approaches zero and magnitude |H(p)| rises sharply. Conversely, the numerator term |p| measures the distance to the zero at the origin (p = 0), pulling the fabric down to ground level.

What about phase? Why the 3D surface does not show everything

This 3D landscape represents exclusively the magnitude |H(p)| (the signal amplitude). It does not show the phase \arg H(p) (the phase shift between output and input). Yet phase obeys a different geometric dynamics: passing to the left of a pole does not automatically produce a full turn of phase.

A simple denominator pole p_k contributes:

\arg\left(\frac{1}{p-p_k}\right) = -\arg(p-p_k)

The phase of the complete transfer function is the sum of the contributions from all its zeros and poles:

\arg H(p) = \arg K + \sum_i \arg(p-z_i) - \sum_k \arg(p-p_k)

Its behaviour therefore depends closely on the path taken around the singularity:

  • Crossing directly through the pole along a straight line: the function H(p) is undefined at the crossing (1/0), and its phase has a local discontinuity of 180^\circ (\pi\text{ rad}).
  • Going all the way around the pole on a small counterclockwise loop enclosing only that pole: by Cauchy’s argument principle, the vector (p - p_k) rotates through +360^\circ, so its denominator contribution produces a full phase change of -360^\circ (-2\pi\text{ rad}).
  • Passing to the left without crossing the pole: phase varies continuously by a finite amount determined by the distance and precise path.

Since the 3D diagram shows only height |H(p)|, another representation, such as a coloured phase map (domain colouring) or a surface of argument \arg H(p), would be needed to visualise this angular behaviour in the complex plane.


The first-order low-pass filter: the RC section

To understand a filter’s inner workings, start with the simplest and most widespread circuit: a resistor R followed by a capacitor C.

RC low-pass section followed by an op-amp voltage follower
Fig. 9 — RC low-pass section followed by an op-amp voltage follower

This circuit forms an elementary voltage divider (two components in series between which the input voltage is shared as a function of frequency, in proportion to their respective impedances). Whereas an ordinary divider made of two resistors divides voltage in a constant ratio regardless of frequency, here one of the two branches is a component whose impedance—complex opposition to alternating current, characterised by a magnitude and a phase—depends on angular frequency (\omega = 2\pi f):

Z_C=\frac{1}{jC\omega}

Consider its physical behaviour:

  • At DC (f = 0\ \text{Hz}, \omega = 0): the capacitor’s insulating dielectric blocks steady current. Its impedance is infinite (Z_C \to \infty), so it behaves as an open circuit. No current flows through resistor R, and its voltage drop is zero. The entire input voltage appears at the output (V_s = V_e); gain is 1, or 0\ \text{dB}.
  • At very high frequencies (f \to \infty): the capacitor charges and discharges very rapidly. Its impedance falls towards zero (Z_C \to 0), approaching a short circuit to ground. The output voltage tends to zero.

Applying the voltage-divider formula, covered in Module 1, Lesson 05 — Resistors in series and parallel (French), gives:

V_s=V_e\frac{Z_C}{R+Z_C}

Hence the sinusoidal transfer function:

H(j\omega)=\frac{V_s}{V_e}=\frac{\dfrac{1}{jC\omega}}{R+\dfrac{1}{jC\omega}}=\frac{1}{1+jRC\omega}

And in the Laplace domain:

H(p)=\frac{1}{1+pRC}

The general canonical form (canonical form = universal standardised mathematical expression that directly exposes the key physical quantities: DC gain, cutoff frequency and time constant) of a first-order low-pass filter is written:

H(p)=\frac{k\omega_c}{p+\omega_c} = \frac{k}{1 + \dfrac{p}{\omega_c}}

where:

  • k is the DC gain (the gain measured at DC, at zero frequency f = 0\ \text{Hz}—here k=1 for our passive section, meaning that a DC signal passes through the circuit without any attenuation, or 0\ \text{dB});
  • \omega_c is the cutoff angular frequency (expressed in \text{rad/s}), directly the reciprocal of the time constant \tau = RC (the capacitor’s characteristic charging time through the resistor: \tau = RC);
  • f_c is the cutoff frequency (expressed in \text{Hz}), which forms the conventional marker separating the passband from the stopband and corresponds to the abscissa of the -3\ \text{dB} point introduced earlier:
\omega_c = \frac{1}{RC} = \frac{1}{\tau} \qquad ; \qquad f_c = \frac{\omega_c}{2\pi} = \frac{1}{2\pi RC}
Gain and phase of a first-order RC low-pass filter
Fig. 10 — Gain and phase of a first-order RC low-pass filter

Let us examine exactly what happens at the cutoff frequency f = f_c, the -3\ \text{dB} point, where \omega = \omega_c = 1/RC:
At this angular frequency, the real and imaginary parts of the denominator are equal (1 + j1). The transfer-function magnitude is:

|H(j\omega_c)|=\frac{1}{|1+j1|}=\frac{1}{\sqrt{1^2+1^2}}=\frac{1}{\sqrt2}\approx0.707

This is the -3.01\ \text{dB} point. Across the same reference impedance, this voltage ratio also corresponds to half the power.
For phase, the denominator 1+j1 has an angle of +45^\circ, so the output phase is exactly -45^\circ in the ideal model: a lag of one-eighth of a period.

The first-order low-pass response therefore has three regions:

  • For f \ll f_c, in the passband, gain remains close to 1 (0\ \text{dB}) and phase is almost zero.
  • At cutoff f = f_c, the -3\ \text{dB} point, gain is -3.01\ \text{dB} relative to the passband plateau and phase is -45^\circ.
  • For f \gg f_c, in the stopband, gain approaches a straight-line slope of -20\ \text{dB/decade}, or -6\ \text{dB/octave}, while phase approaches -90^\circ: a quarter-period lag.

Circuit note: the op amp connected as a voltage follower does not perform the filtering. It is a buffer stage: its very high input impedance does not significantly disturb capacitor charging, while its nearly zero output impedance drives the next stage without degrading the signal.

From the −3 dB point to a filter tolerance mask

We have just seen that the -3\ \text{dB} point (and its associated cutoff frequency f_c) is a filter’s main reference marker. That is convenient in the laboratory, but in industrial practice this single number is not enough. Consider a low-pass filter before an ADC sampling at 48\ \text{kHz}. We want to preserve music up to 20\ \text{kHz} without significant alteration and attenuate everything above 24\ \text{kHz} enough to avoid corrupting conversion. The -3\ \text{dB} point alone tells us neither how much passband variation is acceptable nor how strongly higher frequencies must be suppressed.

We express those requirements using a filter tolerance mask: a drawing that defines boundaries and forbidden regions the response must not enter. This applies to low-pass, high-pass, band-pass and band-stop filters. Whatever the family, the mask divides the spectrum into three kinds of region:

  • The passband (the range of useful frequencies to transmit with minimal alteration): bounded by its characteristic frequencies—for example between f_1 and f_2 for a band-pass filter, from 0 to a frequency f_1 for a low-pass filter, or from f_1 to infinity for a high-pass filter. Within this band, gain must remain between the nominal level (0\ \text{dB}) and a floor denoted -A_p (the maximum tolerated passband attenuation or ripple, for example A_p = 0.5\ \text{dB}).
  • The stopband (the range of unwanted frequencies to eliminate): located outside the useful band (for example beyond a frequency f_2 for a low-pass filter, below a frequency f_1 for a high-pass filter, or on either side of [f_1, f_2] for a band-pass filter). Gain must remain at or below a ceiling denoted -A_a (the minimum required attenuation, for example A_a = 60\ \text{dB}, a thousandfold division of the voltage).
  • The transition band (the intermediate interval separating the passband from the stopband): this is the region where the filter switches from “letting through” to “blocking”. In this magnitude-mask representation, no specific shape is imposed between its two boundaries: different responses may be suitable, provided they still satisfy tolerance A_p up to frequency f_p and have reached at least the minimum attenuation A_a from frequency f_a onwards.
Low-pass tolerance mask with several acceptable frequency responses
Fig. 11 — Low-pass tolerance mask with several acceptable frequency responses

For the conventional low-pass example above:

  • The passband-edge frequency f_1 is conventionally written f_p, for passband.
  • The start-of-stopband frequency f_2 is written f_a, for attenuation.
  • The transition band therefore extends from f_p to f_a.

A mask does not define a unique response. The three curves above differ considerably: one falls gradually, another more steeply, and the third has some ripple. All are acceptable because none enters the red hatched forbidden regions. The mask specifies what is acceptable without, by itself, determining the technology, order or mathematical approximation used to achieve it.

Filter order does constrain the eventual slope: far beyond the transition band, an all-pole low-pass filter of order n approaches -20n\ \text{dB/decade}. But it does not uniquely determine the shape between f_p and f_a. At the same order, Butterworth, Chebyshev, Bessel and elliptic approximations—different mathematical response families—fall differently. Conventional filter synthesis, calculating a filter from its required performance, therefore starts by choosing an approximation and then finding the minimum order that satisfies the mask. If the specification itself sets an order, that becomes an additional constraint.

In our audio example sampled at 48\ \text{kHz}, the requirements are a passband variation below A_p = 0.5\ \text{dB} up to f_1 = f_p = 20\ \text{kHz}, and attenuation of at least A_a = 60\ \text{dB} from f_2 = f_a = 24\ \text{kHz} onwards. The -3\ \text{dB} point remains a convenient reference, but these four key parameters (f_1, A_p, f_2, A_a) define the magnitude mask. A complete specification may add a maximum order, phase-shift requirements (changes in relative timing), group delay (delay of the signal envelope), input and output impedances, technology or cost.


The first-order high-pass filter: the CR section

What happens if we swap the resistor and capacitor?

CR high-pass section followed by an op-amp voltage follower
Fig. 12 — CR high-pass section followed by an op-amp voltage follower

The capacitor now sits in series with the signal path, and the output is measured across resistor R.

The physical behaviour is reversed:

  • At DC (f=0), the capacitor blocks steady current entirely, so the output voltage is zero (V_s = 0).
  • At high frequency, the capacitor’s impedance tends to zero (Z_C \to 0), so almost the whole input appears across R (V_s \approx V_e).

The sinusoidal transfer function follows immediately:

H(j\omega)=\frac{R}{R+\dfrac{1}{jC\omega}} =\frac{jRC\omega}{1+jRC\omega}

In the Laplace domain:

H(p)=\frac{pRC}{1+pRC}=\frac{p}{p+\omega_c}

Its behaviour mirrors the low-pass filter:

  • At cutoff f = f_c, the -3\ \text{dB} point, gain magnitude is 1/\sqrt{2} (-3.01\ \text{dB}), with a phase lead of +45^\circ.
  • For f \ll f_c, gain rises with a slope of +20\ \text{dB/decade}, and phase approaches +90^\circ as f \to 0^+.
  • For f \gg f_c, gain approaches 1 (0\ \text{dB}) and phase shift tends to zero (0^\circ).

Why two identical sections do not simply add together

Once we understand the basic first-order section and its -20\ \text{dB/decade} slope, an obvious idea suggests itself: if one section attenuates at -20\ \text{dB/decade}, surely placing two identical sections one after the other gives -40\ \text{dB/decade} and doubles the roll-off steepness.

Let us try it by directly cascading two identical R-C sections.

Two identical RC sections connected directly in cascade, without a buffer
Fig. 13 — Two identical RC sections connected directly in cascade, without a buffer

Measurement and calculation reveal a disappointment. The final slope does reach -40\ \text{dB/decade} at very high frequencies, but the response near cutoff is poor: the transition is rounded and flattened, and attenuation begins much too early in the passband.

Why? Our intuition silently assumed that each section would operate independently. But connecting the second section directly to the first makes resistor R_2 and capacitor C_2 draw a significant current from the first section. The second section loads the first and reduces its voltage. We have encountered the universal problem of loading and impedance interaction.

The Thévenin view: tracing the loading effect

💡 Reading note — Mathematical derivation of loading
The following paragraphs use a Thévenin equivalent, complex impedances and Millman’s theorem to show exactly where the unwanted interaction term R_1 C_2 comes from.

If you prefer to take away the physical result and move straight to the practical solutions, skip to The golden rule of passive cascading.

To see exactly how the load affects the circuit, use Thévenin’s theorem, covered in Module 2, Lesson 05 — Thévenin’s and Norton’s theorems (French) (forthcoming). Seen from its output terminals, the first stage behaves as an imperfect voltage source with an open-circuit voltage and internal impedance Z_{th}:

Thévenin equivalent and the effect of a load
Fig. 14 — Thévenin equivalent and the effect of a load

The Thévenin diagram deliberately remains general: impedances Z_1, Z_2 and Z_e can represent any loaded divider. Return to the two cascaded RC sections shown above. In the first section, Z_1 represents resistor R_1, while Z_2 is the impedance of capacitor C_1, namely 1/(pC_1). The load Z_e is the second section’s input impedance, R_2 + 1/(pC_2). We use the Laplace variable p because we are deriving the general transfer function H(p). To find the sinusoidal response at angular frequency \omega, set p=j\omega; the capacitor impedance then becomes Z_C=1/(jC\omega).

To find the first stage’s internal Thévenin impedance, replace the ideal voltage source with a short circuit. The complete transformation is shown below without skipping any step:

Three-step calculation of an RC section's Thévenin impedance: original circuit, voltage source shorted, then resistor and capacitor redrawn in parallel
Fig. 15 — Three-step calculation of an RC section’s Thévenin impedance: original circuit, voltage source shorted, then resistor and capacitor redrawn in parallel

In the middle diagram, the left terminal of R_1 is grounded through the short replacing the source. It therefore belongs to the same node as the lower terminal of C_1. The final diagram changes no connections: it simply makes the two parallel branches between the first stage’s output and ground explicit. The internal impedance is:

Z_{th}=R_1\parallel\frac{1}{pC_1} =\frac{R_1}{1+pR_1C_1}

Connecting the second section (R_2, C_2) places its input impedance in series with Z_{th}. Together they form an additional, unintended voltage divider that substantially changes the overall dynamics.

To calculate the intermediate node voltage without writing a large system of equations, we use Millman’s theorem, covered in Module 2, Lesson 05 — Thévenin’s and Norton’s theorems (French) (forthcoming).

Published by Jacob Millman in 1940 (Millman, 1940), this theorem expresses Kirchhoff’s current law directly in terms of node potentials. When n branches (k = 1, 2 \dots n) meet at node M, each connecting a potential V_k through an impedance Z_k, or admittance Y_k = 1/Z_k, the node potential V_M is a weighted average of the neighbouring potentials:

V_M = \frac{\displaystyle\sum_{k=1}^n \frac{V_k}{Z_k}}{\displaystyle\sum_{k=1}^n \frac{1}{Z_k}} = \frac{\displaystyle\sum_{k=1}^n Y_k \cdot V_k}{\displaystyle\sum_{k=1}^n Y_k}

Each term has a direct physical meaning:

  • V_M is the unknown potential at node M, in volts \text{V}.
  • \sum_{k=1}^n, capital Greek sigma, is the summation symbol: add the contributions of every connected branch, from branch k = 1 to branch k = n.
  • n is the total number of branches meeting at the node—here n = 3 for our intermediate node.
  • V_k is the potential at the far end of branch k: input voltage V_e, ground 0\ \text{V} or output potential V_s.
  • Z_k is the complex impedance of branch k, in ohms \Omega, opposing current flow—for example R_1, R_2 or Z_{C1} = 1/(pC_1).
  • Y_k = \dfrac{1}{Z_k} is the admittance of branch k (admittance = the ease with which a component passes alternating current, the exact reciprocal of impedance Y = 1/Z, expressed in siemens \text{S}, denoted G_k = 1/R_k for a plain resistor): the lower an impedance, the larger its admittance;
  • The numerator (top, \sum Y_k V_k = \sum \dfrac{V_k}{Z_k}) represents the sum of the currents that each branch would inject towards the node (I = V/Z);
  • The denominator (bottom, \sum Y_k = \sum \dfrac{1}{Z_k}) represents the sum of all the admittances connected to the node (its total conductance).

💡 An intuitive picture of the weighted average:
Millman’s formula resembles a weighted average of exam marks:

  • The external voltages V_k are the marks.
  • The admittances Y_k = 1/Z_k are their weights.
    A branch with lower resistance—higher admittance and greater weight—pulls node potential V_M more strongly towards its own voltage. An open branch (Z = \infty) has zero weight (Y = 0) and no influence. A short circuit (Z = 0) has infinite weight and imposes its voltage.

Apply the formula to the intermediate node at potential V_1 between the two sections:

  • The input branch connects potential V_e through resistor R_1, giving term V_e / R_1.
  • The middle branch connects ground (0\ \text{V}) through capacitor C_1 of impedance Z_{C1} = 1/(pC_1), giving term 0 / Z_{C1} = 0.
  • The output branch connects potential V_s through resistor R_2, giving term V_s / R_2.

Millman’s formula directly gives:

V_1 = \frac{\dfrac{V_e}{R_1} + \dfrac{0}{Z_{C1}} + \dfrac{V_s}{R_2}}{\dfrac{1}{R_1} + \dfrac{1}{Z_{C1}} + \dfrac{1}{R_2}} = \frac{\dfrac{V_e}{R_1} + \dfrac{V_s}{R_2}}{\dfrac{1}{R_1} + pC_1 + \dfrac{1}{R_2}}

Combining this with the second-stage voltage divider (V_s = V_1 \cdot \dfrac{Z_{C2}}{R_2 + Z_{C2}} = \dfrac{V_1}{1 + pR_2C_2}) gives the overall transfer function of the interconnected passive sections:

H(p)=\frac{1}{1+(R_1C_1+R_2C_2+R_1C_2)p+R_1R_2C_1C_2p^2}

Look closely at the first-degree coefficient in the denominator. Alongside the individual time constants R_1C_1 and R_2C_2, an unwanted cross-term appears: R_1C_2. This is the mathematical signature of interaction between the two stages.

With identical components (R_1 = R_2 = R and C_1 = C_2 = C), the equation becomes:

H(p)=\frac{1}{1 + 3RCp + (RC)^2 p^2}

The middle term is 3RC rather than the 2RC expected for independent stages!
The damping ratio consequently becomes \xi = 1.5, giving a very small quality factor Q = 1/3 \approx 0.333. Loading makes the filter strongly overdamped: its knee rounds off prematurely, producing appreciable attenuation well before angular frequency \omega_0=1/RC.

The golden rule of passive cascading

Why, then, do two identical sections not simply add together?

Because a passive circuit is not isolated: the second section draws current from the first, changes its voltage division and introduces an unwanted interaction term (R_1 C_2).

🎯 The fundamental rule of cascading: before multiplying the transfer functions of directly connected stages, check their interaction. The product H_{\text{total}}(p)=H_1(p)\,H_2(p) is exact when each transfer function already includes its actual load, or when the following stage does not appreciably load the preceding one. For our directly connected RC sections, multiplying the unloaded transfer functions would be wrong: loading reduces the quality factor from Q=0.500 to Q=0.333 and rounds the knee prematurely.


Reducing loading: passive and active solutions

Faced with this inevitable degradation of the direct passive cascade, the engineer has two main avenues: the passive trick of impedance scaling, or the active solution using a buffer stage.

1. The passive solution: impedance scaling

To limit interaction without an active component, the standard technique is to scale the impedances. Give the second stage a much larger resistance and a proportionally smaller capacitance—for example, by a factor of ten:

R_2=10R_1 \qquad ; \qquad C_2=\frac{C_1}{10}

For example, set the reference natural frequency to f_0 \approx 1.59\ \text{kHz} (\tau = 0.1\ \text{ms}). The actual cutoffs at the -3\ \text{dB} point will be 596\ \text{Hz} for 1:1 scaling, 956\ \text{Hz} for 1:10 and 1.024\ \text{kHz} with an op-amp buffer:

  • First stage: R_1 = 1\ \text{k}\Omega and C_1 = 100\ \text{nF} (\tau_1 = 10^3 \times 100\cdot 10^{-9} = 0.1\ \text{ms}).
  • Second stage: R_2 = 10\ \text{k}\Omega (\times 10) and C_2 = 10\ \text{nF} (\div 10), giving \tau_2 = 10^4 \times 10\cdot 10^{-9} = 0.1\ \text{ms}.
Impedance scaling in two cascaded passive RC sections: component values and damping comparison
Fig. 16 — Impedance scaling in two cascaded passive RC sections: component values and damping comparison

The two time constants remain exactly equal (\tau_2 = \tau_1 = \tau). But look at what happens to the unwanted cross-term in the general expression derived using Millman’s theorem:

R_1 C_2 = (1\ \text{k}\Omega) \times (10\ \text{nF}) = 0.01\ \text{ms} = \mathbf{0.1 \cdot \tau}

The transfer-function denominator becomes:

D(p) = 1 + (1 + 1 + 0.1)\tau p + \tau^2 p^2 = 1 + \mathbf{2.1 \tau p} + \tau^2 p^2

The middle term is no longer 3.0\tau, but only 2.1\tau!
The damping ratio falls to \xi \approx 1.05 and the quality factor rises to Q = 1/2.1 \approx 0.476. The filter recovers 95.2\% of the ideally buffered quality factor (Q = 0.500), corresponding to 93.3\% of its bandwidth at the -3\ \text{dB} point, without adding an op amp or a supply rail.

This elegant passive technique nevertheless has practical physical limits:

  • Excessively large resistances (R_2) increase Johnson thermal noise (e_n = \sqrt{4kTR\Delta f}) and make the circuit vulnerable to leakage currents and surrounding electromagnetic interference.
  • Excessively small capacitors (C_2) make the response sensitive to stray capacitances on the circuit board and between neighbouring components.

Why impedance scaling succeeds where identical sections fail

It comes down to the ratio between the first stage’s output impedance and the second stage’s input impedance:

Impedance comparison: identical and impedance-scaled cascades
Fig. 17 — Impedance comparison: identical and impedance-scaled cascades
  1. In the identical cascade (1:1 ratio): the second stage presents an input impedance of 1\ \text{k}\Omega, identical to the first stage’s output impedance (1\ \text{k}\Omega). It is like connecting too heavy a load to a small source: the second stage “pumps” current from the first, pulls down the intermediate voltage V_1 and prematurely flattens the filter’s knee (Q = 0.333).
  2. In the impedance-scaled cascade (1:10 ratio): by multiplying the resistance by 10 (10\ \text{k}\Omega) and dividing the capacitance by 10 (10\ \text{nF}), the time constant remains identical, but the second stage becomes 10 times more resistant to current flow. The interaction cross-term R_1 C_2 is reduced to only 0.1\tau (10% of its initial value): the loading effect on the first stage is drastically reduced, the intermediate voltage V_1 is remarkably well preserved and the parasitic attenuation is almost completely eliminated (Q = 0.476).

The improvement appears directly in the Bode plots. Where the identical cascade had its -3\ \text{dB} point at only 0.374 f_0, scaling moves it to 0.601 f_0, widening the useful bandwidth by +60.5\%. It also reduces the attenuation at f_0 to -6.44\ \text{dB} instead of -9.54\ \text{dB}:

Bode-response comparison: identical passive cascade versus 1:10 impedance scaling
Fig. 18 — Bode-response comparison: identical passive cascade versus 1:10 impedance scaling

⚖️ The physical trade-off:
Impedance scaling is a remarkable passive trick requiring no active component, but it has an unavoidable trade-off: the filter’s overall output impedance is now 10\ \text{k}\Omega instead of 1\ \text{k}\Omega. The filter has itself become more fragile with respect to whatever next circuit is connected to its output V_{\text{out}}.
Avoiding this trade-off is precisely why we turn to an active solution.

2. The active solution: an op-amp voltage-follower buffer (ideal decoupling)

In modern electronic designs, the universal solution is to insert an operational amplifier between the sections, connected as a voltage follower with gain \times 1:

H(p)=H_1(p) \cdot H_2(p) =\frac{1}{(1+pR_1C_1)(1+pR_2C_2)}

In the ideal model, the op-amp follower presents an infinite input impedance (it draws no current from the first section, eliminating the loading effect) and a zero output impedance (in the ideal model, it drives the second section like a perfect voltage source). The interaction cross-term then vanishes strictly and the overall transfer function becomes the exact product of the two independent sections: H(p) = H_1(p) \cdot H_2(p).

Two RC sections isolated by an op-amp follower, compared with a directly connected passive cascade
Fig. 19 — Two RC sections isolated by an op-amp follower, compared with a directly connected passive cascade

Two identical sections isolated by a follower give ideal critical damping: \xi = 1.0 and Q = 0.500. Their independence is restored, but their attenuations still add. At the individual sections’ natural frequency f_0, each has gain -3.01\ \text{dB}, so the combination has gain -6.02\ \text{dB}. The complete filter’s cutoff, its -3\ \text{dB} point, is therefore lower, at approximately 0.644f_0, before the slope approaches -40\ \text{dB/decade} asymptotically.

Three Bode responses: identical passive cascade, impedance-scaled cascade and op-amp-buffered cascade
Fig. 20 — Three Bode responses: identical passive cascade, impedance-scaled cascade and op-amp-buffered cascade

Comparing the three second-order arrangements (f_0 = 1\ 000\ \text{Hz})

  1. Identical passive cascade, ratio 1:1:
    • Quality factor: Q = 0.333, overdamped with \xi = 1.5.
    • -3\ \text{dB} point, actual cutoff: f_c = 0.374\ f_0 = \mathbf{374\ \text{Hz}}, reducing the useful bandwidth by 62.6\%.
    • Gain at natural frequency f_0: -9.54\ \text{dB}, leaving V_{\text{out}} = 0.333\ \text{V} for an input of 1\ \text{V}.
    • Assessment: ❌ Strongest loading; the response falls prematurely.
  2. Impedance-scaled passive cascade, ratio 1:10:
    • Quality factor: Q = 0.476, nearly critically damped with \xi = 1.05.
    • -3\ \text{dB} point, actual cutoff: f_c = 0.601\ f_0 = \mathbf{601\ \text{Hz}}, recovering +60.5\% of useful bandwidth relative to the identical cascade.
    • Gain at natural frequency f_0: -6.44\ \text{dB}, leaving V_{\text{out}} = 0.476\ \text{V}.
    • Assessment: ⚖️ An excellent passive compromise with no active component, but with higher output impedance (10\ \text{k}\Omega).
  3. Active cascade with a voltage follower, ideal isolation:
    • Quality factor: Q = 0.500, ideal critical damping with \xi = 1.0.
    • -3\ \text{dB} point, actual cutoff: f_c = 0.644\ f_0 = \mathbf{644\ \text{Hz}}, a maximum useful-bandwidth increase of +72\% relative to the identical passive cascade.
    • Gain at natural frequency f_0: -6.02\ \text{dB}, leaving V_{\text{out}} = 0.500\ \text{V}.
    • Assessment: ✅ Ideal decoupling: optimal bandwidth, zero parasitic loss and zero output impedance in the ideal model.

💡 What the graph reveals—and the trap to avoid:

  1. The blue 1:10 curve and the green buffered curve really are very close.
    This is no illusion: by simply scaling the impedances in a ratio of 10, the parasitic interaction term weighs only 0.1\tau in the denominator (D(p) = 1 + 2.1\tau p + \tau^2 p^2 instead of 1 + 2.0\tau p + \tau^2 p^2). Passive scaling thus recovers 95.2\% of the ideal quality factor (and 93.3\% of the bandwidth) of a buffered stage without consuming a single milliampere of supply current!
  2. The large gap is the identical cascade, shown in red.
    With a 1:1 ratio, the unwanted interaction contributes 1.0\tau (D(p) = 1 + 3.0\tau p + \tau^2 p^2). Loading reduces useful bandwidth by 62.6\% (f_c = 0.374 f_0) and reduces the signal to -9.54\ \text{dB} at the natural frequency.
  3. Why use an op amp?
    Because the curve shows the filter only when unloaded! The impedance-scaled cascade presents a high output impedance (10\ \text{k}\Omega): as soon as a downstream stage (amplifier, ADC) is connected to it, it will in turn suffer a destructive new loading effect. An op-amp follower, by contrast, has zero output impedance in the ideal model (Z_{\text{out}} = 0\ \Omega): in that model, it guarantees that the filter retains its exact response whatever resistive load is connected downstream. (In a real op amp, maximum output current, bandwidth, voltage swing and capacitive-load stability impose limits that must be checked for the chosen component.)

A different view: what if the frequency axis were linear?

Electronics relies so heavily on logarithmic Bode plots that we sometimes forget the shape of a transfer function on a linear frequency axis, as available on a network analyser or swept-frequency measurement setup.

To see the bandwidth increase without logarithmic compression, plot the same three responses on a linear frequency axis from 0 to 2\cdot f_0:

Responses of the three cascades on a linear frequency axis from 0 to 2 f0
Fig. 21 — Responses of the three cascades on a linear frequency axis from 0 to 2 f0

This representation highlights two things:

  1. The 72% bandwidth increase becomes immediately visible: without logarithmic compression, the difference between the identical cascade’s cutoff (0.374 f_0) and the buffered cutoff (0.644 f_0) stands out clearly.
  2. We understand Hendrik Bode’s stroke of genius: with a linear frequency axis, the -40\ \text{dB/decade} asymptote loses its straightness and plunges along a curve. The logarithmic scale is indispensable for drawing straight-line asymptotes over several decades (10\ \text{Hz} \to 100\ \text{kHz}), whereas the linear view gives a concrete reminder of the signal’s spectral distribution.

Second order: distinguishing f_0, f_c and f_r, and understanding \xi, \omega and Q

First order was simple: a single time constant \tau = RC set the cutoff f_c = 1/(2\pi RC), attenuation approached -20\ \text{dB/decade} and phase moved from 0^\circ to -90^\circ. There was no overshoot or oscillation to surprise us.

As soon as we cross the threshold of second order, the physical picture changes radically. A second-order filter has two independent energy stores: for example, two capacitors or an inductor–capacitor pair. Two capacitors both store electrical energy. In a passive RC network they produce second-order dynamics, but not the oscillatory energy exchange needed for resonance. An LC pair, by contrast, exchanges energy between the capacitor’s electric field and the inductor’s magnetic field. Depending on damping, resonance can then occur.

A single cutoff frequency no longer describes the response adequately. We need two independent physical parameters:

  1. The natural angular frequency \omega_0 (omega zero, \omega_0 = 2\pi f_0), expressed in \text{rad/s}, which sets the frequency scale and where the circuit concentrates its dynamic behaviour.
  2. Damping, which determines the response shape. Engineers describe it using two related quantities:
    • The quality factor Q (the dimensionless number indicating how lightly damped and selective a resonance is): favoured by radio and acoustics engineers, it measures the circuit’s damping and the proximity of the poles to the imaginary axis. The higher Q is, the less damped the system and the livelier its dynamics (resulting, for a canonical low-pass filter (conforming to the standardised reference expression that highlights the circuit’s key parameters), in a sharper resonance).
    • The damping ratio \xi (the dimensionless number measuring how strongly oscillations are braked): this lower-case Greek letter xi is pronounced “xi” or “ksi”. This notation is widely used by control and mechanical engineers.

What does \xi mean, and why the factor of two?

These two quantities describe the same physical behaviour from opposite directions:

Q = \frac{1}{2\xi} \qquad \Longleftrightarrow \qquad 2\xi = \frac{1}{Q}

Why is the factor 2 always attached to \xi?
When physicists write the general differential equation of an oscillating system:

\frac{\mathrm{d}^2 s(t)}{\mathrm{d}t^2} + 2\xi\omega_0 \frac{\mathrm{d}s(t)}{\mathrm{d}t} + \omega_0^2 s(t) = \omega_0^2 e(t)

The factor of two is introduced to simplify the solution. It makes the reduced discriminant of the characteristic equation \Delta' = \omega_0^2(\xi^2 - 1), avoiding cumbersome denominator fractions when calculating the roots—the circuit’s poles.

Pole positions and damping regime are fundamental properties determined by the denominator. The consequences described below for the step response in time and the frequency-response magnitude curve assume a canonical second-order low-pass filter with unity DC gain (canonical form = standardised mathematical reference model); the numerator and its possible zeros can modify these response shapes:

  • Overdamped, non-oscillatory regime (\xi > 1, or Q < 0.5): strong damping produces two distinct real poles. The response is slow and the magnitude falls gradually without peaking.
  • Critical damping (\xi = 1, or Q = 0.5): the two real poles coincide (p_1 = p_2 = -\omega_0). For a fixed natural frequency, this gives the fastest step response without overshoot or oscillation in this canonical family.
  • Underdamped without magnitude peaking (1/\sqrt{2} \le \xi < 1, or 0.5 < Q \le 0.707): the poles are complex conjugates. Some step overshoot appears in time, but the frequency-response magnitude remains strictly decreasing, without a peak. The boundary \xi = 1/\sqrt{2} \approx 0.707 (Q = 0.707) is the Butterworth case, with a maximally flat passband.
  • Underdamped regime with resonance (\xi < 1/\sqrt{2} \approx 0.707, or Q > 0.707): the poles move closer to the imaginary axis. An overvoltage bump (resonance peak) springs up on the magnitude curve before cutoff.
    (The theoretical limiting case is an undamped oscillator with \xi = 0 and Q \to \infty, whose poles lie directly on the imaginary axis p = \pm j\omega_0.)

In its normalised canonical form (canonical form = universal standardised expression where the coefficients are directly Q, \xi and \omega_0 without intermediate calculation), the transfer-function denominator D(p) is written interchangeably with Q or with \xi:

D(p)=1+\frac{p}{Q\omega_0}+\frac{p^2}{\omega_0^2} =1+\frac{2\xi}{\omega_0}p+\frac{p^2}{\omega_0^2}

Three frequencies that must not be confused

Beginners often treat the natural frequency, cutoff frequency and resonance-peak frequency as the same thing. They are three distinct quantities:

  1. f_0, the natural frequency (f_0 = \omega_0 / 2\pi), is set by component values (LC or RC). For the canonical unity-DC-gain low-pass filter, at f = f_0 phase is exactly -90^\circ and gain magnitude equals Q directly: |H(j\omega_0)| = Q, or G_{\text{dB}}(f_0) = 20\log_{10} Q.
  2. f_c, the cutoff frequency at the -3\ \text{dB} point, is where voltage-gain magnitude equals 1/\sqrt{2} \approx 0.707 of its reference value. This 1/\sqrt{2} magnitude corresponds to half power only when the compared impedances are identical.
  3. f_r, the resonant frequency, is where the magnitude response reaches its maximum when a peak exists.

For a canonical unity-DC-gain low-pass filter, f_0 = f_c only in the specific Butterworth case (Q = 1/\sqrt{2}), which has no magnitude peak.

Once Q > 1/\sqrt{2} in that same canonical low-pass filter, a peak appears at resonant frequency f_r, strictly below f_0:

f_r=f_0\sqrt{1-\frac{1}{2Q^2}} < f_0

As Q increases, resonance becomes sharper and f_r approaches f_0 from below, while the -3\ \text{dB} point moves well above f_0.

Placing the poles in the complex plane

The denominator’s roots are the transfer function’s poles. A stable first-order low-pass filter has one negative real pole. An underdamped second-order filter has a complex-conjugate pair of the form p = -\alpha \pm j\omega_d \qquad (\alpha > 0).

Their positions give an immediate picture of the filter’s dynamics:

Poles in the complex plane
Fig. 22 — Poles in the complex plane

A causal rational filter is bounded-input, bounded-output stable—every bounded input produces a bounded output—only when all its poles lie strictly in the left half-plane (\text{Re}(p) < 0). As poles approach the imaginary axis at a fixed natural frequency, damping \xi decreases, quality factor Q increases and resonance sharpens. Crossing into the right half-plane makes the circuit unstable: a real pole produces exponential divergence, while a complex pair produces growing oscillations.

Two passive RC sections

For the second-order passive RC low-pass networks considered so far, containing only resistors and capacitors without amplification or positive feedback, the poles remain on the negative real axis (\xi \ge 1). They cannot produce an underdamped resonance peak.

The two cases already studied are:

  • Two identical directly connected sections, with substantial loading:

    \xi=1.5 \qquad ; \qquad Q=\frac13 \approx 0.333
  • Two identical sections isolated by an ideal voltage-follower buffer:

    \xi=1.0 \qquad ; \qquad Q=\frac12 = 0.500

Can we improve Q without an amplifier by redistributing the resistors and capacitors?
The following diagram compares several component-value combinations:

Examples of second-order passive RC low-pass networks
Fig. 23 — Examples of second-order passive RC low-pass networks

Redistributing resistance and capacitance changes Q in these networks, but does not necessarily increase it. Depending on the combination, Q_{\max} = \sqrt{b}/2, where b is the time-constant ratio. For b = 0.5, Q_{\max} \approx 0.354, a small improvement over Q = 1/3; for b = 1/3, Q_{\max} \approx 0.289, a deterioration. The ideal buffered value Q = 0.5 remains out of reach. To exceed it passively and obtain resonance, we need another kind of reactive energy storage: an inductor.


Returning to the second-order all-pass filter

The numerator and denominator have the same coefficients except for the first-degree p term, whose sign reverses: - in the numerator and + in the denominator. Substituting p with j\omega for sinusoidal steady state makes the polynomials complex conjugates, with equal magnitudes and opposite phases. Their ratio therefore has magnitude 1: gain is constant, while the overall phase is twice the numerator’s phase. Conditions Q>0 and \omega_0>0 put the poles in the left half-plane and ensure stability here; coefficient symmetry alone would not guarantee a stable physical circuit.

With this convention, unwrapped phase progresses from 0^\circ to -360^\circ as frequency increases, passing through -180^\circ at f_0. The final -360^\circ is equivalent to 0^\circ, but records one complete phase rotation. Quality factor Q controls transition steepness: a high Q concentrates the change near f_0, while a low Q spreads it over a wider range.

Constant gain and varying phase of a second-order all-pass filter
Fig. 24 — Constant gain and varying phase of a second-order all-pass filter

An all-pass filter is not necessarily a pure delay. Delaying every frequency by the same time requires phase to vary linearly with frequency, which this response only approximates over a limited range.


From low-pass to high-pass

Before introducing the inductor, let us note a remarkable property of mathematical symmetry: to transform any reference low-pass function of angular frequency \omega_0 into its high-pass equivalent with the same cutoff frequency, we simply apply in the Laplace domain the spectral duality transformation:

p\longrightarrow\frac{\omega_0^2}{p}

The substitution inverts the frequency scale: DC (p=0) maps to infinity, and high frequencies (p \to \infty) map towards the origin.

For first order:

\frac{1}{1+p/\omega_0} \longrightarrow \frac{1}{1 + \dfrac{\omega_0^2}{p\omega_0}} = \frac{1}{1 + \dfrac{\omega_0}{p}} = \frac{p/\omega_0}{1+p/\omega_0}

In an RC circuit diagram, this duality is implemented by exchanging resistors and capacitors, with values rescaled to retain the desired frequency:

C' = \frac{1}{R\omega_0} \qquad ; \qquad R' = \frac{1}{C\omega_0}
Transforming a low-pass filter into a high-pass filter
Fig. 25 — Transforming a low-pass filter into a high-pass filter

The second-order passive RLC low-pass filter

We have seen that a passive resistor–capacitor circuit remains overdamped (Q \le 0.5). Capacitors store and release electric-field energy, but a network containing only resistors and capacitors lacks a complementary magnetic reservoir to support natural oscillatory energy exchange.

To obtain natural resonance without an amplifier, we need that complementary reservoir: an inductor L, storing energy in a magnetic field (E = \frac{1}{2} L I^2).

By combining an inductor and a capacitor, we create a genuine energy ping-pong: the capacitor discharges into the inductor, which accumulates the energy in its magnetic field, and the inductor then returns that energy by recharging the capacitor with the opposite polarity.

Second-order passive RLC low-pass filter
Fig. 26 — Second-order passive RLC low-pass filter

In the ideal theoretical model—zero source impedance, unloaded output and total loop losses represented by series resistor R—the transfer function measured across the capacitor is:

H(p)=\frac{1}{1+RCp+LCp^2}

Identifying terms in the canonical form 1 + \dfrac{p}{Q\omega_0} + \dfrac{p^2}{\omega_0^2} gives:

\omega_0=\frac{1}{\sqrt{LC}} \qquad ; \qquad f_0=\frac{1}{2\pi\sqrt{LC}}
Q=\frac{1}{R}\sqrt{\frac{L}{C}}

This expression for Q is particularly informative:

  • Resistor R represents the circuit’s series resistive losses, including source resistance, the inductor winding resistance and equivalent dielectric losses.
  • As loss resistance R tends to zero, Q tends to infinity: energy oscillates without damping.
  • Adjusting R lets us choose the quality factor in this model: Q = 0.707 for Butterworth, Q = 0.577 for Bessel or Q > 1 for a selective resonant filter. In practice, source resistance, output loading and component tolerances affect the actual Q.
Effect of quality factor on a second-order low-pass response
Fig. 27 — Effect of quality factor on a second-order low-pass response

The second-order passive RLC high-pass filter

Just as with first order, we can rearrange the RLC circuit to create a second-order high-pass filter. Measure the output across inductor L instead of across the capacitor.

The physical behaviour is reversed:

  • At DC and very low frequency (f \to 0), the inductor has almost zero impedance (Z_L = jL\omega \to 0), effectively shorting the output to ground. The rising stopband slope is +40\ \text{dB/decade}, or +12\ \text{dB/octave}.
  • At very high frequency (f \to \infty), the inductor’s impedance is large and the capacitor approaches a short circuit. Almost the entire input appears across L, and gain approaches 1 (0\ \text{dB}).
Effect of quality factor on a second-order high-pass response
Fig. 28 — Effect of quality factor on a second-order high-pass response
  • For f \ll f_0, gain approaches a straight rising slope of +40\ \text{dB/decade}.
  • At sufficiently high quality factor Q (Q > 1/\sqrt{2}), a peak appears above f_0, at the high-pass resonant frequency:

    f_r = \frac{f_0}{\sqrt{1 - \dfrac{1}{2Q^2}}} > f_0
  • For f \gg f_0, gain approaches its unity plateau.

In practice, winding and source resistances limit the maximum attainable Q and round the transition.


Passive RC band-pass filters

What if we want to isolate an intermediate frequency band—for example, speech between 300 Hz and 3 400 Hz in a telephone system?

The obvious approach combines a high-pass filter, removing low frequencies, with a low-pass filter, removing high frequencies. Implementing this using only passive resistors and capacitors soon reveals the severe limitations of RC networks:

Three second-order passive RC band-pass structures
Fig. 29 — Three second-order passive RC band-pass structures

The general canonical form (canonical form = standardised mathematical reference expression that directly exposes the centre-frequency gain A, the natural angular frequency \omega_0 and the quality factor Q) of a second-order band-pass filter is written:

H(p)=A\frac{\dfrac{1}{Q}\dfrac{p}{\omega_0}} {1+\dfrac{1}{Q}\dfrac{p}{\omega_0}+\dfrac{p^2}{\omega_0^2}}

Here A = |H(j\omega_0)| is the maximum gain at the centre of the band.

For these structures, the geometric centre frequency and cutoff frequencies at the -3\ \text{dB} points relative to the peak obey the exact relations:

\omega_0=\sqrt{\omega_1\omega_2}
\omega_1=\omega_0\left(\sqrt{1+\frac{1}{4Q^2}}-\frac{1}{2Q}\right)
\omega_2=\omega_0\left(\sqrt{1+\frac{1}{4Q^2}}+\frac{1}{2Q}\right)

The bandwidth between the two -3\ \text{dB} points (\Delta f) is inversely proportional to quality factor Q:

\Delta f=f_2-f_1=\frac{f_0}{Q}
Second-order band-pass response for several values of Q
Fig. 30 — Second-order band-pass response for several values of Q

Engineering assessment: higher Q means a narrower, more selective band. Purely passive RC networks have only real poles, however, so their quality factor is bounded by Q \le 0.5 (\xi \ge 1). With equal components, direct cascading gives Q = 1/3 \approx 0.333. Strong impedance scaling—for example, R_1 = 10R_2 and C_1 = \dfrac{C_2}{10} in the second structure, giving Q = A = \dfrac{1}{2.1} = \dfrac{10}{21} \approx 0.476—lets some arrangements approach Q = 0.5, but never exceed it. Two sections isolated by an ideal buffer also satisfy Q \le 0.5. Centre-frequency gain A additionally incurs insertion loss depending on topology and scaling: about -9.54\ \text{dB} for equal sections with A = 1/3, reduced to about -6.44\ \text{dB} with A = 10/21 \approx 0.476. (The plotted curves are normalised to a peak of 0\ \text{dB}. For an actual passive circuit, shift the whole curve vertically by 20\log_{10}A.) A passive RC band-pass filter is therefore suitable only for very broad, unselective filtering.


The passive RLC band-pass filter

For a narrow, selective passive band-pass filter with unity gain (A = 1), we need an inductor again.

A series RLC circuit with its output measured across resistor R is a canonical resonant band-pass filter:

Passive series-RLC band-pass filter
Fig. 31 — Passive series-RLC band-pass filter

Its transfer function is:

H(p)=\frac{RCp}{1+RCp+LCp^2}

with:

f_0=\frac{1}{2\pi\sqrt{LC}} \qquad ; \qquad Q=\frac{1}{R}\sqrt{\frac{L}{C}}

Consider the physical behaviour at resonant frequency f = f_0:
At angular frequency \omega_0 = 1/\sqrt{LC}, the inductor impedance (+jL\omega_0) and capacitor impedance (-j/C\omega_0) have equal magnitudes and opposite signs:

Z_{LC} = jL\omega_0 + \frac{1}{jC\omega_0} = j\left(L\omega_0 - \frac{1}{C\omega_0}\right) = 0

In this model without additional parasitic losses, the reactances cancel. The LC branch behaves as a short circuit at resonance, and output ideally equals input (V_s = V_e): unity gain, or 0\ \text{dB}, with zero phase shift. Winding resistance and source impedance introduce some attenuation in practice.

On either side of f_0, impedance Z_{LC} shoots up and chokes off the signal, with asymptotic slopes of +20\ \text{dB/decade} in the low frequencies (bass) and -20\ \text{dB/decade} in the high frequencies (treble).


The twin-T notch filter

The complementary response to band-pass is band-stop, or band-reject.

How can we eliminate a specific interfering frequency (such as 50 Hz hum induced by the AC mains) without using a bulky inductor? The engineers’ historical answer is a masterpiece of passive ingenuity: the twin-T network.

Balanced symmetrical passive twin-T notch filter
Fig. 32 — Balanced symmetrical passive twin-T notch filter

The twin-T uses destructive interference between two parallel signal paths:

  • The upper path is low-pass, with two series resistors R and a shunt capacitor 2C from their midpoint to ground.
  • The lower path is high-pass, with two series capacitors C and a shunt resistor R/2 from their midpoint to ground.

At the tuned frequency:

f_0=\frac{1}{2\pi RC}

The two paths contribute equal-amplitude signals in exact phase opposition (180^\circ). At the output summing node they cancel (V_{s1} + V_{s2} = 0).

Under ideal conditions (perfect components, zero source resistance, infinite load impedance), the theoretical rejection at f_0 is infinite (an abyssal notch at -\infty\ \text{dB}).

Practical construction constraint: actual notch depth depends critically on component matching. Depending on their combination, 2% resistor or capacitor errors can limit rejection to a few tens of decibels. Bruce Carter analyses this sensitivity in detail (Carter, 2006). The passive twin-T also has a modest quality factor (Q = 0.25, giving a broad notch \Delta f = 4f_0) and collapses as soon as a resistive load is connected to it. Part Two shows how an active op-amp arrangement can substantially increase Q and reduce sensitivity to the external load, without eliminating sensitivity to component matching.


When should you choose a passive filter?

After this overview, an engineering question remains: despite the apparent dominance of integrated circuits and op amps, when is a purely passive filter still appropriate?

Passive filters are especially useful in four situations:

  1. No auxiliary supply and minimal complexity: a passive section needs no separate power source. This suits self-contained equipment, passive loudspeaker crossovers and input protection.
  2. High power and large signals: correctly rated high-voltage capacitors and heavy-wire inductors can handle hundreds of volts and tens of amperes—for example in EMI/RFI mains filters and 500 W loudspeaker crossovers—where an ordinary op amp would be destroyed.
  3. Very high frequencies (RF and microwave): at radio frequencies, although very-wide-bandwidth op amps now exist up into the gigahertz range, passive LC, ceramic, quartz or transmission-line filters often remain preferred for their high Q factor, absence of active noise, power handling and excellent linearity within their nominal operating range (even though inductors can saturate at high current and some capacitor dielectrics exhibit nonlinearities under high voltage).
  4. Ultra-low electronic noise: a passive network generates no active semiconductor noise (1/f noise, shot noise), only the unavoidable thermal noise of its resistors.

At lower frequencies, roughly 0.1 Hz to 100 kHz, and with small signals, active filters are often preferable. They implement many responses without bulky, costly inductors, can provide input impedances of megohms to gigohms, and simplify independent adjustment of frequency and damping.


From passive filters to practical integrators

To bridge the gap to the active circuits of Part Two, consider one of analogue electronics’ fundamental topologies: the operational-amplifier integrator.

Replace the feedback resistor in an inverting amplifier with a capacitor C. Input current i = v_e / R then charges that capacitor directly:

H(p)=-\frac{1}{RCp}

In the time domain (for zero initial conditions), the output voltage is the exact antiderivative (integral) of the input signal (the inverse operation of the derivative: instead of measuring the rate of change, we accumulate the signal over time):

v_s(t)=-\frac{1}{RC}\int_0^t v_e(t')\,\mathrm{d}t'
Ideal and practical op-amp integrators
Fig. 33 — Ideal and practical op-amp integrators

The ideal integrator’s trap—and practical solutions

On paper, the ideal integrator performs its task perfectly. In practice, it has a serious weakness: at DC (p=0), the capacitor is an open circuit, removing DC negative feedback.

Any residual DC component at the input—whether the op amp’s offset voltage V_{os} (the tiny imbalance inherent in its input transistors), gate or base input bias currents, or a DC component in the signal—injects a steady current i_0 into the capacitor. Because the op amp relentlessly integrates this current (v_s(t) = -\frac{1}{C}\int i_0\,\mathrm{d}t = -\frac{i_0}{C}t), the output voltage runs away along a linear drift ramp and eventually saturates against one of the supply rails (+15 V or −15 V). The pure ideal integrator is therefore unstable and unusable as-is at DC without a reset loop.

To tame this circuit, the engineer has three practical variants shown in the figure above, all sharing the same asymptotic high-frequency slope of -20\ \text{dB/decade} while differing in gain, polarity and phase shift:

  1. Practical inverting integrator (R_f \parallel C): placing a leakage resistor R_f in parallel with the capacitor restores DC negative feedback:

    H(p)=-\frac{R_f}{R}\frac{1}{1+pR_fC}

    DC gain is limited to A_{\text{max}} = -R_f / R, eliminating the unbounded drift ramp. Well above the corner frequency f_c = \dfrac{1}{2\pi R_f C}, the circuit approaches an ideal inverting integrator, with slope -20\ \text{dB/decade} and phase +90^\circ.

  2. Buffered non-inverting RC low-pass filter: the passive RC section feeds the non-inverting input of an op-amp follower (H(p) = \dfrac{+1}{1+pRC}). DC gain is unity (0\ \text{dB}), the signal is not inverted, phase varies from 0^\circ to -90^\circ, and output impedance is nearly zero.
  3. Amplified non-inverting RC low-pass filter: adding a gain-setting divider (R_1, R_2) in the feedback loop (H(p) = \dfrac{1 + R_2/R_1}{1+pRC}) combines first-order low-pass filtering with adjustable, non-inverting voltage gain in one stage.

These practical integrators are the building blocks used in the next instalment to synthesise the highest-performance universal active filters.


Conclusions and what comes next

Our journey began with a resistor and a capacitor.

Following their physical behaviour led us to cutoff frequency, logarithmic gain and phase shift. Simply trying to combine two sections then exposed loading and the need to control impedance interaction.

That brought us to second order: why two independent parameters—\omega_0 and either \xi or Q—are needed, how poles govern stability in the complex plane, and why an inductor is needed for resonance in these passive lumped-component circuits.

Above all, we encountered the limitations of passive networks: inductor size and losses, destructive interaction between stages, and the impossibility of freely adjusting selectivity without collapsing the signal level.

To cross that threshold and design custom-tailored filters that are steep, stable and adjustable with the turn of a potentiometer, we must bring amplification and active feedback onto the stage.

That is the subject of Part Two: the stabilised differentiator, Sallen–Key and Rauch multiple-feedback (MFB) topologies, the spectacular universal state-variable filter using three or four op amps, and gyrators, which emulate inductance over a specified frequency range subject to noise, dynamic-range and bandwidth limits.

Closing illustration: a large circuit board with inductors, capacitors and a transistor, beside a smaller board with resistors, capacitors and an operational amplifier.

Appendix: Fourier analysis and spectral decomposition

The decompositions associated with French mathematician and physicist Joseph Fourier (Fourier, 1822) were developed in his foundational Théorie analytique de la chaleur, published in 1822. They became a cornerstone of signal processing and linear circuit analysis.

1. Periodic signals: Fourier series

A periodic signal x(t) of period T, fundamental angular frequency \omega_0 = \dfrac{2\pi}{T} and fundamental frequency f_0 = \dfrac{1}{T} admits a Fourier-series representation under the usual sufficient Dirichlet conditions: in particular, integrability over a period and finitely many discontinuities and extrema. At a point of continuity, the series converges to the signal value; at a discontinuity, it converges to the mean of the limits on either side:

x(t) = a_0 + \sum_{n=1}^{+\infty} \left( a_n \cos(n\omega_0 t) + b_n \sin(n\omega_0 t) \right)

Equivalently, in amplitude-and-phase form:

x(t) = A_0 + \sum_{n=1}^{+\infty} A_n \cos(n\omega_0 t + \theta_n)

where:

  • A_0 = a_0 = \dfrac{1}{T}\int_0^T x(t)\,\mathrm{d}t is the mean value, or DC component at 0\ \text{Hz};
  • n = 1 gives the term at frequency f_0, the fundamental;
  • n \ge 2 gives integer-multiple frequencies n f_0, the harmonics of order n;
  • A_n = \sqrt{a_n^2 + b_n^2} is each harmonic’s amplitude, and \theta_n = \operatorname{atan2}(-b_n, a_n) is its phase at the time origin. Function \operatorname{atan2} preserves the correct quadrant according to the signs of a_n and b_n.

2. Nonperiodic signals: the Fourier transform

To move intuitively from a periodic waveform to a single transient—an isolated pulse, a spoken syllable or a disturbance—imagine period T increasing without limit (T \to \infty). The spacing between discrete frequencies, \Delta f = 1/T, tends to zero, and the spectrum becomes continuous. Under suitable existence conditions, this limit leads to the Fourier transform:

X(f) = \int_{-\infty}^{+\infty} x(t)\,e^{-j2\pi f t}\,\mathrm{d}t \qquad \Longleftrightarrow \qquad x(t) = \int_{-\infty}^{+\infty} X(f)\,e^{j2\pi f t}\,\mathrm{d}f

The spectrum is now continuous: X(f) describes the complex amplitude and phase of the frequency components. For a finite-energy signal, |X(f)|^2 is its energy spectral density.

3. Why Fourier analysis matters for linear filters

Every linear time-invariant (LTI) circuit obeys superposition. Knowing its sinusoidal frequency response H(j\omega) lets us obtain output y(t) for a periodic input x(t) by applying gain |H(j n\omega_0)| and phase shift \arg H(j n\omega_0) to each harmonic at its own frequency:

y(t) = A_0 \cdot H(0) + \sum_{n=1}^{+\infty} A_n \cdot |H(j n\omega_0)|\,\cos\big(n\omega_0 t + \theta_n + \arg H(j n\omega_0)\big)

Within its linear operating region and with zero initial conditions, a stable analogue LTI filter can thus be fully characterised by its sinusoidal frequency response H(j\omega).


Sources and background


Part Two: active filters, Sallen–Key, Rauch and state variables →