The Fourier Transform

Start with a concrete case. A train of rectangular pulses of height \(A\) and width \(\tau\), repeating with period \(T\). Its exponential coefficients are

Now hold \(\tau\) fixed and increase \(T\). Two things happen independently:

The envelope never moves. It is fixed by \(\tau\) alone, and its first null stays at \(\omega = 2\pi/\tau\) whatever the period. Only the sampling of that envelope gets finer, and the samples get shorter.

The difficulty and its cure. Every line tends to zero, so \(\lim c_n\) is useless. Multiply by \(T\) first:

— which is exactly the envelope, independent of \(T\). Writing \(\omega\) for \(n\omega_0\) as the spacing vanishes:

The envelope was the transform all along.

In general, substituting \(c_n = \frac1T\int f e^{-jn\omega_0t}dt\):

The \(1/T\) cancels and the limits open out. Reconstruction becomes an integral:

using \(1/T = \omega_0/2\pi \to d\omega/2\pi\).

What changed and what survived:

Every theorem survives; only the bookkeeping changes. That is why Problems 5 to 8 will look familiar from Set 33.

The units point deserves care. \(F(\omega)\) is a density:

so asking "how many volts at 1 kHz?" is meaningless for an aperiodic signal — a single frequency carries no energy in a continuum. Only integrals over a band are physical, which is Problem 11's subject.

The sufficient conditions (Dirichlet, plus absolute integrability):

together with finitely many maxima, minima and discontinuities in any finite interval. The integrability condition is the binding one — it guarantees the defining integral converges, since \(|e^{-j\omega t}| = 1\).

What passes and what fails:

The last row is the genuine failure. A growing exponential has no Fourier transform at all — which is exactly Set 31, Problem 16's point that the imaginary axis lies outside its region of convergence.

The rescue for the middle cases: allow impulses. A constant is not integrable, yet

Check by inverting:

by the sifting property. Admitting impulses extends the transform to every signal of finite power, not merely finite energy — which brings periodic signals back inside the theory (Problem 9).

The impulse-bearing transforms worth knowing:

The step is the interesting one. It is not simply \(1/j\omega\) — the extra \(\pi\delta(\omega)\) accounts for its non-zero mean value of \(\tfrac12\). Omitting it is a standard error.

Why the step needs both terms. Decompose:

The impulse carries the DC content and the \(1/j\omega\) carries the transition. Both are needed, and the decomposition makes the reason transparent.

Conventions vary, and mixing them is a frequent source of stray \(2\pi\) factors:

The hertz convention is tidiest — no factor anywhere — which is why signal processing prefers it. Always check which is in use before quoting a table.

aThe rectangular pulse, height \(A\), width \(\tau\), centred at the origin:

\(F(0)\) is always the area under \(f(t)\) — immediate from the definition with \(\omega = 0\), and the quickest of all checks.

bThe one-sided exponential \(e^{-at}u(t)\), \(a>0\):

Identical to the Laplace transform with \(s = j\omega\) — Problem 12 explains exactly when that equivalence holds. The magnitude \(1/\sqrt{a^2+\omega^2}\) is the familiar single-pole response of Set 28.

cThe two-sided exponential \(e^{-a|t|}\):

Purely real, because the signal is even — a general rule worth noting: even signals have real transforms, odd signals imaginary ones.

dThe Gaussian \(e^{-t^2/2\sigma^2}\):

A Gaussian transforms into a Gaussian — the only common function that is its own transform in shape. Its time width \(\sigma\) and frequency width \(1/\sigma\) are exactly reciprocal, which makes it the extremal case of Problem 15's uncertainty relation.

The working table:

Two universal checks worth applying to every transform computed:

Area in one domain is the zero-frequency value in the other. For the rect: \(F(0) = A\tau\) ✓, and \(\frac{1}{2\pi}\int A\tau\,\text{sinc}(\omega\tau/2)d\omega = A\) ✓.

The theorem. If \(f(t) \leftrightarrow F(\omega)\), then

Taking the transform's shape as a time function returns the original shape, reflected, with a factor \(2\pi\).

The proof is a relabelling. Start from the inverse transform:

Now exchange the names of the variables \(t\) and \(\omega\):

The forward and inverse transforms differ only in a sign and a constant, which is why the theorem exists at all.

Immediate consequences, each obtained free:

The second row is the one that matters most. A sinc-shaped pulse has a rectangular spectrum — it is perfectly band-limited, which is the basis of the ideal filter of Problem 16 and the reconstruction formula of Problem 17.

Work one through in full. The Lorentzian \(f(t) = \dfrac{1}{1+t^2}\):

By duality, taking \(F(t) = \dfrac{2}{1+t^2}\):

A contour integral avoided entirely. Direct evaluation of \(\int\frac{e^{-j\omega t}}{1+t^2}dt\) requires residues; duality gives it in two lines.

The symmetry properties that duality implies:

The third is conjugate symmetry, the same property that made \(c_{-n} = c_n^*\) in Set 33. It means the negative-frequency half carries no independent information for a real signal — which halves the storage in any practical spectrum computation.

The theorem:

Proof by substituting \(\tau = at\) (taking \(a>0\)):

The \(|a|\) handles time reversal, where \(a<0\) flips the integration limits.

Reading it. Two separate effects:

Speeding a signal up spreads its spectrum in exact proportion. Playing a recording at double speed doubles every frequency present — the scaling theorem, audible.

The invariant product. For a rectangular pulse the first spectral null is at \(\omega = 2\pi/\tau\):

Constant, exactly. The scaling theorem guarantees it: shrinking \(\tau\) by any factor expands every spectral feature by the same factor, so any product of a time width and a frequency width is scale-invariant.

The general statement:

The rectangular case is a caution: its RMS bandwidth is infinite, because \(\int\omega^2|F|^2d\omega\) diverges for a \(1/\omega\) tail. Practical bandwidth measures use the first null or the \(-3\) dB point instead, and Problem 15 treats the rigorous version.

Where this shows up:

The third is a familiar engineering headache: a designer who speeds up an edge for timing margin has widened the emission spectrum by the same factor, and the compliance test will say so.

aTime shift. Substituting \(\tau = t-t_0\):

Magnitude untouched; phase rotated linearly in \(\omega\) — the continuous version of Set 33, Problem 18, and once again the definition of a pure delay (constant group delay, Set 28, Problem 17).

bFrequency shift, the dual:

Multiplying by a complex exponential translates the spectrum. Duality guaranteed this the moment the time-shift theorem was proved.

The modulation theorem follows by writing the cosine as two exponentials:

The spectrum splits into two half-height copies, centred at \(\pm\omega_0\). This single result is the basis of every radio transmission ever made.

A worked case. A 4 kHz-bandwidth audio signal modulating a 1 MHz carrier:

The bandwidth doubles — upper and lower sidebands both appear, and for a real signal they are mirror images carrying identical information. That redundancy is what single-sideband transmission removes.

Demodulation is modulation again. Multiplying the received signal by the same carrier:

Two multiplications and a filter recover the original. The catch is that the second carrier must match the first in frequency and phase — a phase error of \(\theta\) scales the output by \(\cos\theta\), and 90° recovers nothing at all.

The general principle:

Modulation is the special case where \(F_2\) is a pair of impulses, and convolving with an impulse simply relocates — which is why the spectrum translates rather than smearing. Problem 18 shows what happens when \(F_2\) is not impulsive.

Differentiation. Integrating by parts, with the boundary term vanishing since \(f \to 0\) at infinity:

No initial-condition terms — unlike Laplace (Set 31, Problem 5). The two-sided transform has no starting instant to attach them to, which is precisely why Fourier cannot handle initial conditions and Laplace can.

Integration, which needs care:

The impulse term is essential and often dropped. It accounts for the DC level the integration produces: if \(f\) has non-zero area, its integral tends to a non-zero constant, and a constant needs an impulse in its spectrum (Problem 2). When \(F(0) = 0\) the term vanishes and the simple form is correct.

Use differentiation to find a transform. Take the triangular pulse of height \(A\) and half-width \(\tau\). Differentiating once gives two rectangles:

Differentiating again gives three impulses:

Impulses are trivial to transform — that is the whole trick.

Transform the impulses using the shift theorem:

Now divide by \((j\omega)^2 = -\omega^2\):

Using \(1-\cos x = 2\sin^2(x/2)\):

The triangle's transform is sinc-squared — obtained without evaluating a single integral.

Check it:

The decay rate matches Set 33, Problem 10's rule exactly — a corner gives \(1/\omega^2\), a jump gives \(1/\omega\), in the continuum just as in the series.

Why the triangle is the rect convolved with itself. Since \(\text{sinc}^2 = \text{sinc}\times\text{sinc}\), the convolution theorem (Problem 8) says

which is easily verified directly — sliding one rectangle across another gives a linearly rising then falling overlap. Two independent routes to the same result, and a useful consistency check on both theorems.

The pair of theorems:

Duals of one another, as Problem 4 requires. The \(1/2\pi\) appears in the second because of the convention chosen; in the hertz convention both are clean.

The first theorem is filtering. A linear circuit convolves its input with \(h(t)\):

This is Set 31, Problem 12's result, now applying to any finite-energy signal rather than only causal ones — and it is why "filtering" and "multiplying the spectrum" are the same sentence.

The second theorem is windowing, and it is the one that causes trouble. Observing a signal for a finite time multiplies it by a window:

The true spectrum is smeared by the window's transform. Since a finite window is a rect and a rect's transform is a sinc with \(-13.3\) dB sidelobes, every measured spectrum is blurred whether or not anyone intended it. Problem 18 develops this.

A worked example of the first theorem. Two rectangular pulses of width \(\tau\):

which is the triangle of Problem 7 ✓. Convolving two identical pulses gives a triangle of twice the width — and the spectral consequence is that the sidelobes drop from \(-13.3\) dB to \(-26.6\) dB, since squaring doubles the decibels.

Convolution with an impulse is the case that recurs everywhere:

The third row is the key to the whole sampling theorem: sampling in time replicates the spectrum in frequency, and aliasing is those replicas overlapping.

The practical payoff: fast convolution. Convolving two \(N\)-point sequences costs \(N^2\) operations directly, but

costs \(O(N\log N)\) with the FFT of Problem 19. For \(N = 10^6\) that is a speedup of about 52 000, which is why long filters are always implemented this way.

Start from the series. Any periodic \(f(t)\) is \(\sum c_ne^{jn\omega_0t}\), and each exponential transforms to an impulse (Problem 2):

A periodic signal's transform is a train of impulses at the harmonics, weighted by the Fourier coefficients. The line spectrum of Set 33 is literally a set of impulses — which is why it was drawn as vertical lines.

Why impulses rather than finite values. A periodic signal has infinite energy:

so \(|F|^2\) cannot be an ordinary function. The infinite energy is concentrated at discrete frequencies, and an impulse is the only object that carries finite weight at a single point. Problem 11 returns to the distinction between energy and power signals.

The impulse train, which is the single most useful case:

Its coefficients:

An impulse train transforms into an impulse train. Spacing \(T\) in time becomes spacing \(2\pi/T\) in frequency — reciprocal, exactly as Problem 5 demands.

The reciprocity is worth tabulating:

Denser in time means sparser in frequency. The two limiting rows recover the \(\delta \leftrightarrow 1\) pair from either direction.

Building a periodic signal by convolution. Any periodic signal is one period convolved with an impulse train:

By the convolution theorem:

Comparing with the boxed result gives \(c_n = F_1(n\omega_0)/T\) — the Fourier coefficients are samples of the single-pulse transform. Problem 1's observation that the series samples an envelope, now proved.

The unification is complete:

Fourier series is not a separate theory but a special case — periodicity in time forces discreteness in frequency, and that duality between periodicity and discreteness recurs in Problems 17 and 19.

The theorem — the transform's counterpart of Parseval:

Energy is the same whether counted in time or in frequency. It follows from the convolution theorem applied to \(f\) and \(f^*\), or directly by substituting the inversion integral.

Verify on the two-sided exponential \(f(t) = e^{-a|t|}\) with \(a = 2\). In the time domain:

In the frequency domain, with \(F = 2a/(a^2+\omega^2)\):

Numerical evaluation confirms both to six figures.

Using it to evaluate integrals. The theorem runs in either direction, which makes it a computational tool. Take the rectangular pulse of unit height and width \(\tau\):

Rearranging with \(x = \omega\tau/2\):

A standard integral, obtained in three lines without contour integration — the same trick that produced number series in Set 33.

Energy signals and power signals must be distinguished:

A signal is one or the other, never both (except the trivial zero signal). Which it is determines whether Rayleigh or Parseval applies, and whether the spectrum is a curve or a set of lines.

The practical use: energy in a band. Because the integrand is a density, energy over any frequency range follows by integration:

(the factor 2 collecting the negative-frequency half for a real signal). This answers questions like "what fraction of a pulse's energy lies below 1 MHz?" — which is precisely what an emissions limit asks.

The definition. Rayleigh's theorem identifies the integrand as an energy density:

Note it depends on \(|F|\) only — phase information is discarded entirely. Two signals with the same magnitude spectrum and different phases have identical energy distributions and can look nothing alike (Set 33, Problem 8).

Apply to a rectangular pulse of height \(A\) and width \(\tau\):

The cumulative energy up to the \(k\)-th null at \(\omega = 2\pi k/\tau\):

The main lobe carries 90% of the energy, which is why \(B \approx 1/\tau\) is the standard working bandwidth for a pulse. Squeezing out the last few per cent costs ten times the bandwidth.

Why the tail is so persistent. The sidelobes fall as \(1/\omega^2\) in energy (since \(|F| \sim 1/\omega\)), so

— the residual energy falls only as \(1/\Omega\). Halving the leftover energy requires doubling the bandwidth, and this is a direct consequence of the pulse's discontinuous edges (Set 33, Problem 10).

Rounding the edges changes everything. Comparing shapes of the same duration:

This is why transmitted pulses are shaped. A rectangular pulse's spectral tails interfere with neighbouring channels indefinitely; a smoothed pulse of the same duration confines its energy.

Power spectral density is the counterpart for signals of infinite energy:

The middle row — computing \(|F_T|^2/T\) for a truncated record and letting \(T\) grow — is what a spectrum analyser actually does, and Problem 18 shows what the truncation costs.

A practical calculation. A 1 µs, 5 V rectangular pulse:

That last 1% is what an EMC test measures. It is negligible as energy and entirely sufficient to fail a radiated-emissions limit.

The two definitions side by side:

The condition for equality — both must hold:

The second is Set 31, Problem 16's condition again — and for a causal signal it means all poles strictly in the left half-plane.

Test three cases:

The middle row is the instructive one. The step's ROC has the imaginary axis as its boundary, not its interior, and that borderline case is exactly where the extra impulse appears. Substituting \(s = j\omega\) naively loses it.

Which to use:

A rough division: Laplace for circuits, Fourier for signals. Laplace answers "what does this network do?"; Fourier answers "what is in this waveform?"

The deeper relationship. The Laplace transform is the Fourier transform of an exponentially weighted signal:

The weighting \(e^{-\sigma t}\) is what tames a growing signal so that a transform exists at all — Set 31, Problem 2's convergence factor, seen from the Fourier side. Fourier is the special case \(\sigma = 0\), which works only when no taming is needed.

Why both survive. Neither subsumes the other:

Each relaxes one restriction and imposes another. The two-sided Laplace transform relaxes both and is the true parent of each — but it requires stating an ROC explicitly every time, which is why the specialised versions are preferred in practice.

The three steps are the same as always:

For an \(RC\) low-pass with \(\tau_c = RC\):

The input, a pulse of height \(A\) from 0 to \(T_0\). Writing it as two steps:

The impulse term vanishes because \((1-e^{-j\omega T_0}) = 0\) at \(\omega = 0\), leaving

which is the finite-energy pulse spectrum, as it must be.

Multiply and invert:

Partial-fractioning the \(\omega\)-dependent part and applying the delay theorem term by term:

Identical to Set 32, Problem 12's result — as it must be, since the circuit is stable and starts from rest.

Compare the two routes honestly:

For a circuit problem Laplace is simply better. The Fourier route here is an exercise, not a recommendation — and its awkwardness with the step's impulse illustrates why.

Where Fourier does win. The transform's advantage is in the spectral question, not the time-domain answer:

The energy question is a good illustration. With \(\tau_c = T_0\):

One integral, no inversion required — whereas the Laplace route would need \(y(t)\) first and then \(\int y^2dt\).

The division of labour that has now emerged across three sets:

Both use the same \(H\), evaluated on the same axis. They are two questions about one object, not two competing methods.

The debt. Set 28, Problem 19 asserted \(t_rf_c = 0.35\) and could not derive it, because \(H(j\omega)\) contains no information about time. It is derivable now.

Take a first-order low-pass. Its step response is

Find the 10% and 90% crossings:

Subtract:

Convert to hertz with \(\omega_c = 2\pi f_c\):

Exact, not empirical. The familiar 0.35 is \(\ln9/2\pi\) rounded to two figures — a closed form that has been quoted as a rule of thumb for decades.

How much does the shape matter? Computing the same product for other responses:

The constant varies by only a few per cent across very different filters, which is exactly what makes the rule useful. It is not a coincidence but a consequence of the scaling theorem: any product of a time width and a bandwidth is scale-invariant (Problem 5), so only the shape can affect it — and shape has surprisingly little leverage.

Why the product is bounded below. A fast edge needs high-frequency content:

and a filter that removes it must slow the edge. No filter can make an edge faster than its own bandwidth allows — which is Problem 15's uncertainty principle in its engineering form.

Cascading. When several stages each slow the edge, the rise times combine in quadrature:

The instrument dominates here. A 1 ns edge measured with a 2 ns scope reads 2.69 ns — an error of 169%. The quadrature rule is exact for Gaussian responses and a good approximation otherwise.

The practical rules that follow:

Define the widths as normalised second moments, which makes the statement precise:

The theorem:

with equality only for a Gaussian. In hertz this reads \(\Delta t\,\Delta f \ge 1/4\pi = 0.0796\).

Verify numerically for Gaussians of three widths:

Exactly one half in every case, and the widths trade off precisely inversely — confirming both the bound and its attainment. Note \(\Delta t = \sigma/\sqrt2\) for this definition.

A caution about the rectangular pulse. Its time width is obviously finite, yet

because the \(1/\omega\) tails make the integral diverge. A rectangular pulse has infinite RMS bandwidth — the product is trivially infinite and the inequality is satisfied uselessly. This is why practical bandwidth measures use nulls or \(-3\) dB points instead (Problem 11).

The engineering consequences are all versions of one statement:

Every one is the same inequality wearing different clothes. The Set 29 entry is worth noting: a resonance's ringing time \(2/\text{BW}\) is precisely this bound, which is why high-\(Q\) circuits ring for so long.

The spectrum-analyser limit, made concrete:

To resolve two tones 10 Hz apart requires at least 100 ms of data, no matter how expensive the instrument. This is not a limitation of technology but of mathematics, and Problem 19 shows the DFT obeys it exactly.

Why the Gaussian is optimal in so many settings follows directly:

Set 30's Bessel filter approximates a Gaussian response, which is why it shows almost no step overshoot — the same optimality, reached from the filter-design direction.

The ideal response is a rectangle in frequency:

By duality (Problem 4), its inverse transform is a sinc:

Read its properties:

The third row is fatal. \(h(t) \ne 0\) for \(t < 0\) means the filter responds before the input arrives — Set 30, Problem 2's causality argument, now with the offending function in hand.

Quantify the non-causality. The energy in the pre-arrival part:

Exactly half the impulse response occurs before the impulse. This is not a small correction to be neglected — the response is symmetric about \(t = 0\), so half its activity is in the future's past.

The practical fix: delay it. Truncate \(h(t)\) to a window of length \(2T_d\) and shift right by \(T_d\):

This is exactly how an FIR digital filter is designed. The delay is the price of causality, and it is a price a digital filter can pay because it can buffer the signal — an analogue circuit cannot.

The truncation ripple is Gibbs. Cutting the sinc off abruptly is multiplying by a rect, which convolves the spectrum with a sinc (Problem 8):

producing 8.9% overshoot at the band edge however long the truncation — Set 33, Problem 9's result in the continuum. Tapering the window instead of cutting abruptly removes it, at the cost of a wider transition, which is Problem 18's trade.

The three impossibilities, now quantified:

Set 30 could state these; the transform explains them. The brick wall's non-causality is not an abstract obstacle but a specific function extending symmetrically into negative time.

Model sampling as multiplication by an impulse train of spacing \(T_s\):

By the multiplication theorem (Problem 8), this convolves the spectra:

Insert the impulse train's transform from Problem 9:

Sampling replicates the spectrum at every multiple of the sampling rate. Convolving with an impulse relocates (Problem 8), so an impulse train produces an infinite set of copies.

The condition for recovery. The copies must not overlap. If \(X(\omega) = 0\) for \(|\omega| > \omega_m\), adjacent copies stay clear provided

The Nyquist criterion. Sample at more than twice the highest frequency present and the baseband copy survives intact, recoverable with an ideal low-pass — which is Problem 16's sinc, and why the reconstruction formula involves sinc interpolation.

What aliasing does. Content above \(\omega_s/2\) folds back:

Aliasing is irreversible. Once 600 Hz has become 400 Hz there is nothing in the samples distinguishing it from a genuine 400 Hz component — no algorithm can undo it, which is why the anti-aliasing filter must precede the sampler.

The folding rule. A component at \(f\) appears at

which maps the whole frequency axis onto \([0, f_s/2]\) by repeated reflection. The spectrum folds like a concertina at every multiple of the Nyquist frequency.

The anti-aliasing requirement, connecting to Set 30, Challenge C1:

for a 20 kHz passband. The stopband edge is \(f_s-20\) kHz, not Nyquist — only content above that folds into the audio band. Set 30 found this turns a 44th-order filter problem into a 4th-order one.

The origin of the problem. No measurement is infinitely long, so what is actually transformed is

The true spectrum is convolved with the window's transform. Even doing nothing applies a rectangular window, whose sinc has \(-13.3\) dB sidelobes — there is no such thing as an unwindowed measurement.

What leakage looks like. A pure sinusoid should give one line. Convolved with a sinc it gives a peak plus sidelobes spreading across the whole spectrum:

Two findings, both important. When the signal fits a whole number of cycles into the record, the rectangular window is perfect and a Hann window makes things worse. When it does not — the normal case — the rectangular window is catastrophic and the Hann recovers 30 dB.

Why coherent sampling is special. With an exact number of cycles, the sinc's zeros land precisely on every other bin:

so all the leakage falls in the nulls and vanishes. Half a bin of offset destroys this completely — which is why leakage is the normal condition and coherent sampling a laboratory luxury.

The window catalogue:

The trade is resolution against leakage. A Blackman window buries the sidelobes 45 dB deeper than a rectangular one and costs a factor of three in the ability to separate close peaks.

Choosing a window is choosing which error to tolerate:

The second row is the common case in practice: a harmonic 60 dB below a fundamental is invisible under rectangular sidelobes at \(-13\) dB, and plainly visible under Blackman's at \(-58\) dB.

The connection to filter design. This is Set 30's trade in another guise:

Designing a window and designing a filter are the same problem — which is why FIR filter design is done by windowing the ideal sinc of Problem 16.

The definition — the integral replaced by a finite sum:

Both domains are now discrete and finite — the only version a computer can evaluate.

The parameter relationships, which are what one actually needs:

Bin spacing is the reciprocal of the record length — Problem 15's uncertainty relation appearing as an exact identity. Finer resolution requires a longer record and nothing else.

The hidden assumption. The DFT treats the record as one period of a periodic signal:

Both consequences follow from Problem 9's duality. Discreteness in time forces periodicity in frequency (aliasing, Problem 17); discreteness in frequency forces periodicity in time — which is why a record whose ends do not match produces a discontinuity at the wrap-around, and hence leakage (Problem 18).

Two elementary checks:

Exactly the discrete analogues of \(\delta \leftrightarrow 1\) and \(1 \leftrightarrow 2\pi\delta(\omega)\) — duality survives discretisation.

The FFT. Direct evaluation costs \(N^2\) complex multiplications; the fast algorithm exploits symmetry to reduce this to \(O(N\log_2N)\):

The FFT is not a different transform — it computes exactly the same DFT, only faster. That speed is what made real-time spectrum analysis, digital filtering and modern communications practical at all.

What must be watched when using it:

The third deserves emphasis. Zero-padding a short record increases the number of bins without adding information — it interpolates the same smeared spectrum more finely. Resolution comes from record length alone, exactly as Problem 15 requires.

The debts settled. Reciprocal relations noticed across five sets, all now consequences of one theorem:

Every one required a continuum to state. A discrete spectrum has no notion of "duration times bandwidth" because there is no bandwidth, only a set of lines.

The three great dualities the set established:

The division of labour across Part 3, now complete:

All six evaluate the same \(H\) — on a point, a line, a region or the whole plane. They differ in what question is being asked, not in the object being examined.

The one structural gap that survives. Set 32, Challenge C1 showed that two \(RC\) sections cascaded directly give \(1/(s^2+3s+1)\), not \(1/(1+s)^2\):

No transform fixes this, because it is not a transform problem. A single \(H(s)\) describes a network only when nothing is drawn from its output — and a directly connected second stage always draws current.

What Set 35 must supply:

Four parameters, not one. The \(z\), \(y\), \(h\) and \(ABCD\) descriptions of Set 35 each carry enough information to predict what happens when a network is loaded — and the \(ABCD\) form makes cascading an exact matrix product.