Two-Port Networks

What \(H(s)\) actually says. It is defined as a ratio measured under one specific condition:

The output must be open-circuited. Every transfer function quoted in Sets 28 to 34 carried this hidden condition, and it was harmless only because nothing was ever connected.

The failure, concretely. Set 32, Challenge C1 cascaded two \(RC\) sections:

42% of the bandwidth lost, and a repeated pole split into two distinct ones. The prediction is not slightly optimistic; it describes a different circuit.

Why it fails. Connecting stage 2 draws current from stage 1's output, so \(I_2 \ne 0\) and the defining condition is violated:

The first stage no longer behaves as characterised. Nothing is wrong with \(H_1\) — it is simply being used outside the conditions of its own definition.

What is missing. To predict the loaded behaviour one must know:

This is not a coincidence of counting. A two-port has four terminal variables, two of them independent, so a linear relation between them needs exactly four coefficients.

When the shortcut is safe. The product rule works if the interface carries no power:

This is why active filters cascade cleanly and passive ladders do not — Set 30, Problem 16 designed the ladder as a whole for exactly this reason.

The programme for this set:

Problem 10 returns to the failed example and gets \(s^2+3s+1\) in two lines of matrix multiplication.

The defining equations, taking both currents as independent:

All four have units of impedance, hence the name. Both currents flow into the network — the standard convention.

How each is measured. Set one current to zero by open-circuiting that port:

The \(z\) parameters are the open-circuit parameters — every measurement leaves one port open, which is easy at low frequency and awkward at high (a genuine open circuit is hard to realise, Problem 18).

Apply to a T-network — series \(Z_a\), series \(Z_b\), shunt \(Z_c\) between them:

The shunt arm appears in every entry — it is the element common to both loops, which is exactly what a transfer impedance measures.

Read the structure:

The last row is worth noting: a reciprocal two-port has only three independent parameters, and the T-network's three elements match them exactly. That is why any reciprocal two-port has a T-equivalent (Problem 14).

A numerical case used throughout: \(Z_a = 10\), \(Z_b = 20\), \(Z_c = 30\ \Omega\):

Reciprocal (\(z_{12} = z_{21}\)) but not symmetric (\(40 \ne 50\)) — as expected, since \(Z_a \ne Z_b\).

When \(z\) parameters do not exist. An ideal transformer has \(V_1/V_2 = n\) regardless of current, so no finite \(z_{11}\) exists:

No single parameter set describes every two-port. Problem 19 tabulates which sets fail for which networks — a series impedance has no \(y\), a shunt admittance has no \(z\), and the ideal transformer has neither.

The defining equations, now with voltages independent:

The dual of the \(z\) set, and measured by short-circuiting a port rather than opening it — which is why they are the short-circuit parameters.

The measurements:

Short circuits are easier to realise than open circuits at high frequency, which is one reason \(y\) parameters are preferred for transistor characterisation above audio.

The Pi-network — shunt \(Y_1\), series \(Y_2\), shunt \(Y_3\):

The negative sign is not an error — it is the same minus that appears off-diagonal in nodal analysis (Set 5), and for the same reason. The Pi-network is to \(y\) what the T-network is to \(z\).

The relationship to \(z\) is matrix inversion, not term-by-term reciprocal:

This is the single commonest error in the topic. The parameters are entries of inverse matrices, so \(y_{11}\) involves \(z_{22}\) — the two are measured under different terminations and cannot be reciprocals of each other.

Demonstrate on the running T-network, \(\mathbf{z} = \begin{bmatrix}40&30\\30&50\end{bmatrix}\), \(\Delta z = 1100\):

The two differ by nearly a factor of two. Physically, \(1/z_{11}\) is the input admittance with the output open; \(y_{11}\) is with it shorted — different circuits, different answers.

When \(y\) fails. The inversion requires \(\Delta z \ne 0\):

A lone series impedance has \(\mathbf{z} = \begin{bmatrix}Z&Z\\Z&Z\end{bmatrix}\), which is singular — shorting the output gives infinite current, so no finite \(y_{11}\) exists.

The hybrid set takes \(I_1\) and \(V_2\) as independent:

The name "hybrid" is literal — the four parameters have four different units:

Why this suits a transistor. Each parameter matches a quantity that is both physically meaningful and easily measured:

The last is decisive. A transistor's \(z\) parameters need the input open, which for a BJT means no base current and therefore a device that is switched off. The measurement destroys the operating point — which is why \(z\) parameters are useless for a transistor and \(h\) parameters became the industry standard.

The equivalent circuit reads directly off the equations:

A Thévenin input and a Norton output — which is exactly how a transistor behaves, and why the model is so durable.

Typical values for a small-signal BJT in common emitter:

Note the enormous spread — from \(10^{-4}\) to \(10^3\). The two small parameters, \(h_{re}\) and \(h_{oe}\), are routinely neglected, and Problem 16 checks whether that is justified.

The reciprocity test is different for this set:

Emphatically not reciprocal — forward gain 50, reverse coupling \(2.5\times10^{-4}\), a ratio of 200 000. That asymmetry is amplification, and it is only possible because the transistor is active. Problem 7 shows every passive bilateral network must be reciprocal.

The defining equations, expressing the input in terms of the output:

The minus signs are deliberate. With \(-I_2\) as the current flowing out into a load, the output variables of one stage become the input variables of the next — which is what makes cascading a matrix product (Problem 10).

The four measurements:

Note \(A = V_1/V_2\), not \(V_2/V_1\) — the transmission parameters run backwards, describing what input is needed for a given output. The open-circuit voltage gain is therefore \(1/A\).

The two elementary building blocks:

Both have determinant 1 and are as simple as a matrix can be without being the identity. Every ladder network is a product of these two, which is the whole reason the transmission set exists.

A worked case — one \(RC\) section (series \(R\), then shunt \(C\)), which Problem 10 will need:

With \(R = 1\), \(C = 1\):

Check the gain: \(1/A = 1/(1+s)\) ✓ — the familiar unloaded \(RC\) response. Note the order of multiplication matches the physical order of the elements, input first.

Other useful blocks:

All have unit determinant — a signature of reciprocity (Problem 7). The transformer row also shows why \(ABCD\) succeeds where \(z\) and \(y\) both fail.

The \(ABCD\) set is the odd one out in a useful way:

Derive rather than memorise. Start from the \(z\) equations and solve for \(V_1, I_1\) in terms of \(V_2, -I_2\). From the second equation:

Comparing with \(I_1 = CV_2-DI_2\) gives \(C = 1/z_{21}\) and \(D = z_{22}/z_{21}\) immediately.

Substitute into the first equation:

Everything is divided by \(z_{21}\) — so the conversion fails for a network with no forward transmission, which is exactly right.

A structural check. Compute the determinant:

So \(AD-BC = 1\) precisely when \(z_{12} = z_{21}\) — the two reciprocity tests are the same statement, which Problem 7 pursues.

Apply to the T-network \(\mathbf{z} = \begin{bmatrix}40&30\\30&50\end{bmatrix}\), \(\Delta z = 1100\):

Verify against direct circuit analysis. Terminating in \(Z_L\), the transmission parameters predict

Direct analysis of the T — \(Z_a\) in series with \(Z_c\) parallel \((Z_b+Z_L)\):

The last row is a good check: with the output open, \(Z_{\text{in}}\) must equal \(z_{11}\) by definition, and it does.

The conversion map. All six sets are interconvertible where they exist:

Rather than memorise the table, re-derive. Each conversion is two linear equations rearranged, and doing so takes less time than recalling which index goes where.

The physical statement. Drive port 1 with a current and measure the open-circuit voltage at port 2; then swap. A reciprocal network gives the same answer both ways:

This is the two-port form of the reciprocity theorem met in Set 12 — source and detector may be exchanged without altering the reading.

The test in each set:

The \(h\) and \(ABCD\) forms look unrelated but are not — Problem 6 showed \(AD-BC = z_{12}/z_{21}\), so the determinant condition is the equality condition, rewritten.

What guarantees reciprocity. A network is reciprocal if it contains only:

The rule: passive and bilateral implies reciprocal. Mutual inductance is worth noting — despite its asymmetric appearance, \(M_{12} = M_{21}\) always (Set 17), so coupled coils are reciprocal.

Why non-reciprocity is valuable. Amplification requires it:

A reciprocal device cannot amplify. If forward and reverse transmission were equal, any output signal would feed back to the input in equal measure — which is why every gain block ever built is non-reciprocal, and why the isolation ratio above is a figure of merit.

The gyrator is the cleanest non-reciprocal passive element:

Antisymmetric rather than symmetric. Its useful property is impedance inversion: a gyrator terminated in a capacitor looks inductive,

which is how integrated circuits synthesise inductance without a coil — Set 30's active filters rely on it.

A practical use of the test:

The determinant test is the best single check on any hand computation in this topic — it catches most algebra errors immediately.

The two properties are independent and are constantly confused:

Symmetry is the stronger condition: it implies reciprocity for passive networks, but reciprocity does not imply symmetry.

The tests:

Reciprocity constrains the off-diagonal; symmetry constrains the diagonal. That is the cleanest way to remember which is which.

Test the running T-network, \(Z_a = 10\), \(Z_b = 20\), \(Z_c = 30\):

Physically obvious once seen: looking in from port 1 the first series arm is 10 Ω; from port 2 it is 20 Ω. The network is not the same from both sides.

Make it symmetric by setting \(Z_a = Z_b = 10\), keeping \(Z_c = 30\):

Both conditions now hold. A symmetric reciprocal two-port has only two independent parameters — down from four — which is why such networks are so tractable.

What symmetry buys:

This is why filter sections and attenuator pads are almost always built symmetric — Set 30's ladder prototypes, and the T and Pi pads of Problem 15.

A counter-example worth holding on to. An ideal transformer:

Reciprocal but not symmetric — exactly as expected, since a step-up transformer viewed from the other side is a step-down. It proves the two properties are genuinely independent.

aSeries connection — inputs in series, outputs in series. The currents are shared and the voltages add:

The \(z\) set is chosen because it takes current as the independent variable — and current is what the two networks share.

bParallel connection — the dual. Voltages shared, currents add:

Example: two reciprocal networks in parallel,

Both reciprocal and symmetric, and so is the sum — these properties survive addition, which is a useful check.

cCascade connection — output of the first becomes input of the second:

The second condition is the crucial one. The current flowing out of network 1 flows into network 2 — and the \(ABCD\) convention was defined with \(-I_2\) precisely so that this becomes an identity rather than a sign nuisance.

The derivation. For network 1:

and for network 2, \(\begin{bmatrix}V_1''\\I_1''\end{bmatrix} = \mathbf{T}''\begin{bmatrix}V_2''\\-I_2''\end{bmatrix}\). Substituting:

An ordinary matrix product, in the physical order of the stages. Loading is handled automatically because the shared variables are matched exactly, not assumed away.

Two consequences of using matrices:

The second is elegant: since each reciprocal block has determinant 1, any cascade of them does too. Reciprocity is preserved under all three interconnections.

The summary table:

Each parameter set exists because some interconnection makes it add. That is the organising principle of the whole topic — and it explains why there are exactly these sets and not others.

A caution. The addition rules assume the interconnection does not disturb the individual networks' port conditions. That assumption can fail, and Problem 17 gives the test — the cascade rule, however, is always valid.

The one section, from Problem 5:

Unloaded gain \(1/A = 1/(1+s)\) — correct for a single section standing alone.

Cascade the two by multiplying:

Entry by entry:

The transfer function with the output open is \(1/A\):

Exactly Set 32, Challenge C1's answer, obtained in three lines of arithmetic rather than a full nodal analysis — and with no opportunity to make the loading mistake, because the matrix product cannot express it.

Where the extra \(s\) comes from. Track the term:

The loading term is \(B'C''\), and it is precisely what the naive product omits. \(B' = R = 1\) is the first stage's series resistance; \(C'' = sC = s\) is the second's shunt admittance. Their product is the current stolen.

Compare the two predictions:

The pole product is 1 in both cases; loading does not move the poles' product, it spreads them apart — one moving toward the origin (slower) and one away (faster), with the slow one setting the bandwidth.

The golden-ratio curiosity. The poles are \(-\varphi^2\) and \(-1/\varphi^2\) where \(\varphi = (1+\sqrt5)/2\):

A consequence of the continued-fraction structure of the ladder — pretty, and a memorable check that the numbers are right.

The buffered alternative. Insert a unity-gain buffer between the stages:

The buffer forces \(I_2' = 0\), restoring the condition under which \(H_1\) was defined. This is exactly why Set 30's Sallen–Key stages cascade cleanly and why passive ladder filters must be designed as a whole.

From the \(ABCD\) set, the quickest route. The load imposes \(V_2 = -I_2Z_L\):

A bilinear (Möbius) transformation of \(Z_L\) — the same mathematical form that governs the Smith chart, and the reason circles map to circles there.

From the \(z\) set, eliminating \(I_2 = -V_2/Z_L\):

This form is the more revealing of the two. The input impedance is the open-circuit value \(z_{11}\) reduced by a term that depends on the load — and the reduction is proportional to \(z_{12}z_{21}\), the product of forward and reverse transfer.

The limiting cases, which are the best checks available:

Both must agree, and they do — since \(A/C = z_{11}\) from Problem 6's conversion. Any discrepancy signals an algebra error.

The running T-network, \(A = 4/3\), \(B = 110/3\), \(C = 1/30\), \(D = 5/3\):

Every value confirmed by elementary series–parallel reduction. Note \(Z_{\text{in}}\) varies only from 22 to 40 Ω as the load swings from short to open — the shunt arm \(Z_c\) partially isolates the input from the load.

Output impedance is the mirror image, found by looking in at port 2 with a source impedance \(Z_S\) at port 1:

\(A\) and \(D\) exchange places — which is exactly what reversing a two-port does, and why a symmetric network (\(A = D\)) has the same formula in both directions.

Why the bilinear form matters. As \(Z_L\) traces a curve, \(Z_{\text{in}}\) traces its image:

The last row is the basis of impedance measurement through a fixture: measure \(Z_{\text{in}}\), know the fixture's \(ABCD\), and invert to recover the device's true impedance — de-embedding, in instrument terminology.

Three gains must be distinguished, and confusing them is a standard error:

The loaded gain. From \(V_1 = AV_2-BI_2\) with \(I_2 = -V_2/Z_L\):

As \(Z_L \to \infty\) this tends to \(1/A\) — the open-circuit gain of Problem 5, recovered as it must be.

The overall gain adds the input divider between \(Z_S\) and \(Z_{\text{in}}\):

Substituting \(Z_{\text{in}}\) from Problem 11 and simplifying:

All four parameters appear, together with both terminations. This single expression contains everything: it reduces to \(1/A\) when \(Z_S = 0\) and \(Z_L = \infty\), and degrades from there.

Quantify on the running T-network (\(A = 4/3\), \(B = 110/3\), \(C = 1/30\), \(D = 5/3\)):

The terminations cost more than the network itself. The bare network attenuates by 2.5 dB; realistic source and load together turn that into 12.2 dB. Quoting an unloaded gain for a circuit that will be terminated is therefore not a small optimism.

Current gain and power gain follow the same pattern:

For a passive network \(G_P \le 1\) always; the value of \(Z_L\) maximising it is the conjugate-matched load, connecting back to Set 12's maximum power transfer theorem — the one method Set 32 found could not move into the \(s\)-domain freely.

The design lesson:

These three goals conflict, and choosing between them is the substance of interface design. Voltage amplifiers want mismatch; power stages want matching; transmission systems want \(Z_0\) termination.

The defining question: is there a load \(Z_0\) such that the input impedance equals the same \(Z_0\)? Setting \(Z_{\text{in}} = Z_L = Z_0\):

For a symmetric network \(A = D\), so the linear term vanishes:

Simple only because of the symmetry. For an unsymmetric network the quadratic has two distinct roots — the image impedances \(Z_{I1}\) and \(Z_{I2}\), one for each port.

Verify the claim. Substituting \(Z_L = \sqrt{B/C}\) with \(D = A\):

Multiplying numerator and denominator appropriately and using \(A^2-BC = 1\), this reduces identically to \(\sqrt{B/C}\) ✓ — confirmed symbolically.

A worked case: symmetric T with \(Z_a = Z_b = 10\), \(Z_c = 30\ \Omega\):

Also worth checking: \(Z_0 = \sqrt{Z_{\text{oc}}Z_{\text{sc}}}\), the geometric mean of the open- and short-circuit input impedances — a two-measurement route to \(Z_0\) for any network.

Why it matters for cascades. Terminate a chain of identical symmetric sections in \(Z_0\):

Every section sees the same impedance, so every section behaves identically. This is what makes ladder filters and transmission lines analysable: \(Z_0\) is the impedance that propagates unchanged, and it is why loading — the problem this whole set exists to solve — disappears when a chain is properly terminated.

The propagation constant completes the description:

Losses add when matched, exactly as decibels are meant to. When not matched they do not, which is precisely the naive-cascade error of Problem 10.

The counting argument. A reciprocal two-port has three independent parameters (\(z_{11}\), \(z_{22}\), and \(z_{12} = z_{21}\)). Both T and Pi have three elements:

The match is exact, so a unique equivalent exists — no more elements are needed, and no fewer will do.

The T equivalent. Comparing with Problem 2's result \(\mathbf{z} = \begin{bmatrix}Z_a+Z_c&Z_c\\Z_c&Z_b+Z_c\end{bmatrix}\):

For the running network \(\mathbf{z} = \begin{bmatrix}40&30\\30&50\end{bmatrix}\):

Recovering the original elements exactly — as it must, since that network was a T.

The Pi equivalent, from the \(y\) parameters:

Using \(\mathbf{y}\) from Problem 3:

A different circuit with identical terminal behaviour. No measurement made at the two ports can distinguish this Pi from the original T.

Verify the equivalence at the open-circuit condition:

Matching \(z_{11} = 40\ \Omega\) from the T. The equivalence is exact, not approximate.

This is the Y–Δ transformation of Set 3, in two-port language:

The generalisation is free — the transformation was always about network topology rather than about resistance, so replacing \(R\) by \(Z(s)\) costs nothing.

The limits of the equivalence — three real restrictions:

The third deserves care. A negative element is mathematically valid and physically unrealisable as a passive component — it appears legitimately in equivalent circuits for coupled coils (Set 17's T-model of a transformer, where one arm is \(L_1-M\) and can be negative for tight coupling).

The definition compares the load voltage with and without the network inserted:

It is a property of the network and its terminations, not of the network alone — which is exactly why it differs from the transfer function.

A simple case: a 50 Ω series resistor between a 50 Ω source and 50 Ω load.

Note this is not the same as the voltage ratio across the network — insertion loss accounts for the fact that inserting the network also changes what the source delivers.

The general expression in terms of \(ABCD\):

The numerator is the denominator of Problem 12's overall gain; the denominator is what a direct connection would give. For a matched, lossless network IL = 0 dB.

Design a matched T-pad. Requirements: \(Z_0 = 50\ \Omega\) at both ports, and 6 dB of attenuation. For a symmetric T:

where \(N\) is the voltage ratio. For 6 dB, \(N = 10^{6/20} = 1.9953\):

Check: terminated in 50 Ω, \(Z_{\text{in}} = 16.61+66.93\|(16.61+50) = 16.61+33.39 = 50.0\ \Omega\) ✓ — matched, as designed.

Why attenuators are built this way. A simple series resistor also attenuates, but:

The pad attenuates without disturbing the system impedance — so it can be inserted anywhere in a chain without affecting anything else. That is worth the extra two resistors, and it is why every laboratory attenuator is a pad rather than a resistor.

A common application: improving a poor match. A pad between a mismatched source and load reduces the reflection seen at each end:

A 6 dB pad improves the return loss by 12 dB, because a reflected wave traverses the pad twice. The cost is 6 dB of signal — a trade routinely accepted at the input of sensitive instruments.

The device and the load:

Derive the input impedance. The load imposes \(V_2 = -I_2R_L\). From the second \(h\) equation:

Substituting into the first:

Evaluate. The key intermediate quantity:

The gains follow:

The minus sign on \(A_v\) is the familiar phase inversion of a common-emitter stage.

Now test the simplification. Setting \(h_{re} = h_{oe} = 0\):

Errors of 10 to 25% — not negligible. The usual claim that both parameters can be dropped is too casual, and it is worth identifying which one is responsible.

Separate the two effects:

\(h_{oe}\) is the parameter that matters here, not \(h_{re}\). The reason is the load: \(h_{oe}R_L = 0.25\) is not small, because \(1/h_{oe} = 40\) kΩ is only four times \(R_L\). The transistor's own output resistance is shunting the load appreciably.

The condition for the simplification to hold:

The rule of thumb "neglect \(h_{oe}\)" is safe only for low-impedance loads. This stage's 10 kΩ load is precisely where it breaks down — and high-gain stages naturally use high load resistances, so the approximation fails exactly when the stage is most interesting.

What the two-port view adds. The output impedance depends on the source impedance, which a one-number model cannot express:

With \(R_S = 600\ \Omega\): \(Z_{\text{out}} = 56.7\) kΩ, against \(1/h_{oe} = 40\) kΩ if \(h_{re}\) were ignored.

Here \(h_{re}\) does matter — a 42% effect on output impedance, because it provides the internal feedback path that couples the source into the output. Whether a parameter is negligible depends on which quantity is being computed.

The hidden assumption. Deriving \(\mathbf{z} = \mathbf{z}'+\mathbf{z}''\) required that each network still be a two-port after connection — that the current entering the upper terminal of a port equals the current leaving its lower terminal:

This is the port condition, and it is what the word "port" means. Connecting two networks can destroy it.

How it fails. If the interconnection creates a path allowing current to enter one port and leave through another network's ground return, the two terminals of a "port" no longer carry equal and opposite currents:

The commonest cause is a shared ground connection that provides an unintended circulating path — which is easy to create accidentally when two three-terminal networks are joined.

The Brune test decides the matter. For a series connection, short the corresponding port terminals of the two networks and measure the voltage across the resulting break:

A zero reading means the connection carries no current it should not, so the two networks remain independent two-ports and their matrices may be added.

Why the cascade rule never fails. Compare the three connections:

In a cascade the output port of one network is the input port of the next — the same two terminals, so whatever current leaves one enters the other by construction. The port condition is satisfied identically, not merely usually.

This is a second reason the transmission parameters are preferred. Beyond composing by multiplication, they carry no validity caveat:

The practical remedy when a series or parallel connection fails the test: insert an ideal 1:1 transformer in one branch.

The transformer breaks the offending ground loop without altering the intended signal path — which is exactly why isolation transformers appear in bridge and instrumentation circuits.

The measurement problem. Every parameter set so far requires an open or a short circuit:

At microwave frequencies neither is realisable. A short has inductance, an open has capacitance, and at 10 GHz a 1 nH lead is 63 Ω of reactance — comparable with the impedances being measured. The terminations become part of the network.

A second difficulty: an active device terminated in a short or open frequently oscillates, destroying the measurement and sometimes the device.

The resolution: measure with matched terminations instead. Define incident and reflected wave variables:

\(a\) is the wave going in, \(b\) the wave coming back, both normalised so that \(|a|^2\) and \(|b|^2\) are powers.

Each parameter is measured with the other port terminated in \(Z_0\), so \(a_2 = 0\) — no reflection returns:

A matched termination is easy to build at any frequency — it is just a well-made 50 Ω resistor — and it does not provoke oscillation. That single change is what makes microwave measurement possible.

The quantities engineers actually quote are \(S\) parameters in disguise:

A datasheet reading "return loss 20 dB" is stating \(|S_{11}| = 0.1\) — 1% of the incident power reflected, and a VSWR of 1.22.

The properties carry over in recognisable form:

The same structure as Problems 7 and 8 — off-diagonal equality for reciprocity, diagonal equality for symmetry. The lossless condition is simply conservation of power: whatever is not reflected is transmitted.

Cascading uses the \(T\)-matrix (a rearrangement of \(S\)), which multiplies just as \(ABCD\) does — the same principle as Problem 9, in wave variables. And \(S\) converts to \(z\), \(y\) or \(ABCD\) whenever those exist, so nothing is lost.

First criterion: what is being done to the network.

Second criterion: existence. Not every set exists for every network:

The ideal transformer row is the striking one. Three of the four sets fail, and only \(ABCD\) survives — because a transformer is naturally a mapping from one port to the other, which is exactly what the transmission set encodes.

Why the failures occur. Each is a singular matrix:

A failure signals a constraint the chosen variables cannot express — not a defect in the network, and the remedy is always to change parameter set.

Third criterion: what can actually be measured.

The decision procedure in order:

Step 4 is a reasonable default. The transmission parameters exist for every network in the table above, compose without a validity test, and convert freely to the others when needed.

The architecture. Every method rested on two conservation laws stated in Set 1:

Nothing added afterwards was independent of these. Mesh and nodal analysis are systematic applications; superposition, Thévenin and Norton are consequences of linearity; phasors and transforms change the algebra without touching the physics.

The three parts, and what each bought:

The recurring pattern. Each part made the same move — find a transformation that turns a hard problem into an easy one:

Six versions of one strategy. The circuits never changed; the descriptions did.

The debts, all settled:

The last was the most stubborn, surviving four sets and three transforms — because it was never a transform problem at all.

What the whole book was about, stated once:

The habits worth keeping, which recurred in every set:

The fourth is the one that matters most. Almost every error in these thirty-five sets — the naive cascade, the power-factor formula, the substitution \(s = j\omega\) for an unstable system, neglecting \(h_{oe}\) — was a formula used outside the conditions of its own derivation.