Fourier, Laplace and Z-transforms

In short

Transforms move a signal from the time domain, where an LTI system acts by convolution and by differential or difference equations, to a frequency or complex-variable domain, where the same action becomes simple multiplication. The Fourier family (series, transform, DTFT, DFT) gives the spectrum; the Laplace transform handles continuous signals that grow, and carries initial conditions, so it is the tool for circuits and control; the Z-transform does the same job for sequences and is the basis of all digital filter design. This is the highest-yield block in GATE EC/EE/IN (typically 6-10 marks a year, mostly on ROC, poles, stability and standard pairs) and a fixed part of AICTE-model "Signals and Systems" and "Digital Signal Processing" papers. In interviews at BEL, ISRO, DRDO, C-DOT, Tejas Networks, Qualcomm India, Texas Instruments India, Analog Devices and Synopsys, the questions are nearly always about ROC, the unit circle, the FFT's speed advantage and why Laplace is preferred over Fourier.

Explained

Fourier series — periodic signals and line spectra

• A periodic signal x(t) with period T (fundamental ω₀ = 2π/T rad/s) is a sum of harmonically related sinusoids. Three equivalent forms: trigonometric (a₀, aₙ, bₙ), polar (Aₙ, φₙ) and exponential (Cₙ) — the exponential form is the one used in exams.
• Dirichlet conditions (sufficient, not necessary): x(t) absolutely integrable over one period, finitely many maxima/minima and finitely many finite discontinuities per period.
• Symmetry shortcuts: even signal → only cosine terms (bₙ = 0); odd → only sine terms (a₀ = aₙ = 0); half-wave odd symmetry, x(t ± T/2) = −x(t) → only odd harmonics (square wave, triangular wave).
• Line spectrum: |Cₙ| versus nω₀ is discrete. For a real x(t), C₋ₙ = Cₙ* — magnitude even, phase odd (conjugate symmetry).
• Parseval: average power equals the sum of powers of the harmonics, so a few harmonics can be shown to carry most of the power.
• Gibbs phenomenon: truncating the series at N terms leaves about 9% overshoot near a discontinuity; the overshoot does not shrink with N, it only moves closer to the jump.
• A periodic signal has no Fourier transform in the ordinary sense; its CTFT is a train of impulses of area 2πCₙ at ω = nω₀.

Continuous-time Fourier transform (CTFT)

• Extends the series to aperiodic signals by letting T → ∞: the line spectrum becomes a continuous spectrum X(ω). Exists (sufficient condition) if x(t) is absolutely integrable; energy signals qualify. Power signals (u(t), cos ω₀t, dc) are handled using impulses.
• Key properties (used far more than the defining integral):
  • Linearity; time shift x(t−t₀) ↔ X(ω)e^−jωt₀ — magnitude unchanged, linear phase added.
  • Frequency shift / modulation x(t)e^{jω_c t} ↔ X(ω−ω_c); multiplying by cos ω_c t splits the spectrum into two half-amplitude copies.
  • Scaling x(at) ↔ (1/|a|)X(ω/a) — compress in time, expand in frequency (the time-bandwidth trade-off).
  • Duality: every pair read backwards is another pair (rect ↔ sinc, sinc ↔ rect).
  • Differentiation dx/dt ↔ jωX(ω) (a high-pass action); integration ↔ X(ω)/jω + πX(0)δ(ω).
  • Convolution x₁*x₂ ↔ X₁X₂, and multiplication ↔ (1/2π)(X₁*X₂).
  • Parseval/Rayleigh: energy is the same computed either domain; |X(ω)|² is the energy spectral density.
• Limitations: cannot represent signals that grow (e^atu(t), a > 0), gives no systematic way to include initial conditions, and treats stable and unstable systems alike — which is exactly why Laplace exists.

Laplace transform — ROC, poles and stability

• Bilateral: X(s) = ∫ x(t)e^−stdt over all t, s = σ + jω. Unilateral (engineering default) integrates from 0⁻, so an impulse or a jump at t = 0 is captured and initial conditions appear naturally.
• Region of convergence (ROC): the set of σ for which the integral converges. Rules: the ROC is a vertical strip, contains no poles; right-sided signal → Re s > σ_max (right of the rightmost pole); left-sided → Re s < σ_min; two-sided → a strip between poles; finite-duration → whole s-plane.
• Fourier transform exists iff the ROC includes the jω axis; then X(jω) is the Fourier transform. So Laplace is the general tool and Fourier is its restriction to the imaginary axis.
• Properties: L{x′(t)} = sX(s) − x(0⁻); L{x″} = s²X − sx(0⁻) − x′(0⁻); integration → X(s)/s; time shift (t₀ > 0) → e^−st₀X(s); s-shift e^−atx(t) → X(s+a); convolution → product; time scaling; multiplication by t → −dX/ds.
• Initial value theorem: x(0⁺) = lim_{s→∞} sX(s), valid only if X(s) is a strictly proper rational function (no impulse at t = 0). Final value theorem: lim_{t→∞} x(t) = lim_{s→0} sX(s), valid only if all poles of sX(s) lie strictly in the left half-plane — applying it to an oscillatory or unstable system is the classic exam trap.
• Transfer function H(s) = Y(s)/X(s) with zero initial conditions = Laplace transform of the impulse response. Poles fix the natural response (e^pt terms), zeros shape the forced response.
• Stability (continuous): a causal LTI system is BIBO stable iff all poles are in the open left half-plane (ROC includes the jω axis). Poles on the jω axis → marginally stable; repeated jω poles or any RHP pole → unstable.
• Inversion: partial fractions plus a table (distinct poles, repeated poles, complex-conjugate pairs handled as damped sinusoids); formally the Bromwich contour integral or the residue method.

DTFT, DFT and the FFT

• DTFT: X(e^jω) = Σ x[n]e^−jωn; it is continuous in ω and 2π-periodic; exists if x[n] is absolutely summable. Inverse is an integral over one period (−π to π).
• DFT: the computable transform of a finite N-point sequence — N frequency samples of the DTFT at ω_k = 2πk/N. Implicitly treats x[n] as periodic with period N, so its natural convolution is circular.
• Linear convolution of sequences of length N₁ and N₂ needs zero-padding both to L ≥ N₁ + N₂ − 1 before multiplying DFTs, otherwise wrap-around (aliasing in time) corrupts the start of the result.
• Frequency resolution = f_s/N Hz; zero-padding interpolates the display but does not improve true resolution. A frequency not exactly at a bin centre causes spectral leakage and the picket-fence effect; windows (Hanning, Hamming, Blackman) trade main-lobe width against side-lobe level.
• FFT: radix-2 decimation-in-time/frequency algorithms exploit symmetry and periodicity of W_N = e^−j2π/N to cut N² complex multiplications to about (N/2)log₂N. Needs N a power of 2 (otherwise mixed-radix); DIT takes bit-reversed input, DIF gives bit-reversed output.
• DFT properties: circular shift ↔ linear phase, circular convolution ↔ product, Parseval Σ|x[n]|² = (1/N)Σ|X[k]|², and for real x[n], X[N−k] = X*[k] (so only N/2 + 1 bins are independent).

Z-transform — the discrete counterpart

• Definition: X(z) = Σ x[n]z⁻ⁿ, z complex. ROC is an annulus R₁ < |z| < R₂ centred at the origin, containing no poles. Right-sided (causal) → |z| > largest |pole|; left-sided → |z| < smallest |pole|; two-sided → a ring; finite-length → all z except possibly 0 and/or ∞.
• The algebraic expression alone does not determine the signal: (0.5)ⁿu[n] and −(0.5)ⁿu[−n−1] have the identical X(z) = z/(z−0.5); only the ROC separates them.
• X(z) evaluated on z = e^jω gives the DTFT, so the DTFT exists iff the ROC includes the unit circle.
• Properties: linearity; shift x[n−k] ↔ z^−kX(z) (bilateral); scaling aⁿx[n] ↔ X(z/a); time reversal x[−n] ↔ X(1/z); differentiation n·x[n] ↔ −z dX/dz; convolution ↔ product; accumulation ↔ X(z)/(1−z⁻¹); initial value x[0] = lim_{z→∞}X(z); final value lim(z−1)X(z) as z→1 (only if (z−1)X(z) has all poles inside the unit circle).
• Unilateral shift with initial conditions: Z{x[n−1]} = z⁻¹X(z) + x[−1] — this is how difference equations with stored states are solved.
• Inversion: (i) long division / power-series expansion (gives x[n] term by term, good for the first few samples); (ii) partial fractions plus a table, with the ROC deciding whether each term is right- or left-sided; (iii) residues of X(z)zⁿ⁻¹.
• Stability (discrete): a causal LTI system is BIBO stable iff all poles lie strictly inside the unit circle; a pole at |z| = 1 is marginally stable (an accumulator has a pole at z = 1), and |z| > 1 is unstable. FIR filters have poles only at the origin, so they are always stable.
• Mapping z = e^sT (T = sampling period): jω axis → unit circle, left half-plane → inside, right half-plane → outside, and the mapping is many-to-one (every horizontal strip of width 2π/T maps onto the same z-plane) — the algebraic face of aliasing. The bilinear transform is the practical, one-to-one alternative used to convert analogue prototypes into IIR filters.

Choosing the right transform

Key points

• Fundamental idea - convolution in time becomes multiplication in the transform domain; differential/difference equations become algebra.
• Exponential Fourier series - Cₙ = (1/T)∫x(t)e^−jnω₀tdt; real signals give C₋ₙ = Cₙ*.
• Dirichlet conditions - absolute integrability over a period plus finitely many discontinuities and extrema.
• Gibbs phenomenon - ~9% overshoot at a jump, independent of the number of terms retained.
• Duality - rect ↔ sinc both ways; narrow in one domain means wide in the other.
• CTFT of a periodic signal - impulses of area 2πCₙ at nω₀, never an ordinary function.
• ROC (Laplace) - a vertical strip with no poles; right-sided signals give a right half-plane ROC.
• Fourier from Laplace - substitute s = jω, legitimate only if the ROC contains the jω axis.
• Final value theorem - valid only when every pole of sX(s) is in the open left half-plane.
• Continuous stability - all poles in the left half-plane; discrete stability - all poles inside the unit circle.
• Transfer function - Laplace (or Z) transform of the impulse response, taken with zero initial conditions.
• ROC decides the signal - the same X(z) is causal, anticausal or two-sided depending on which annulus is specified.
• DTFT is periodic in ω with period 2π; DFT samples it at N equally spaced points.
• DFT convolution is circular - zero-pad to N₁ + N₂ − 1 to get linear convolution.
• Resolution = f_s/N; zero-padding interpolates, it does not resolve closer tones.
• FFT cost ≈ (N/2)log₂N complex multiplications against N² for the direct DFT.
• z = e^sT maps the LHP inside the unit circle and is many-to-one (aliasing); the bilinear transform avoids that.
• FIR is always stable (poles only at z = 0); IIR stability must be checked against the unit circle.

Formulas and examples

Example 1 — Fourier series of a square wave (harmonic power). A bipolar square wave of amplitude 1 V and period T = 2 ms.

• f₀ = 1/T = 500 Hz. Half-wave odd symmetry ⇒ odd harmonics only: x(t) = (4/π)[sin ω₀t + (1/3)sin 3ω₀t + (1/5)sin 5ω₀t + …].
• Fundamental amplitude = 4/π = 1.273 V at 500 Hz; third harmonic = 4/(3π) = 0.4244 V at 1,500 Hz; fifth = 0.2546 V at 2,500 Hz.
• Total power (1 Ω) = (1 V)² = 1 W. Fundamental power = 1.273²/2 = 0.811 W (81.1%); adding the third = 0.4244²/2 = 0.090 W gives 90.1%; adding the fifth = 0.0324 W gives 93.3%.
• A square wave therefore keeps only about 81% of its power in the fundamental — about 19% sits in the harmonics (THD = √(π²/8 − 1) ≈ 48.3%), which is why square-wave inverter output is unacceptable for sensitive loads.

Example 2 — CTFT of a rectangular pulse (bandwidth). x(t) = 1 V for |t| ≤ 0.5 µs, i.e. τ = 1 µs.

• X(f) = Aτ sinc(fτ) ⇒ X(0) = 1 × 10⁻⁶ V·s.
• First spectral null at f = 1/τ = 1 MHz; nulls repeat every 1 MHz. The one-sided 3 dB bandwidth is 0.443/τ ≈ 443 kHz (0.886/τ ≈ 886 kHz between the two −3 dB points).
• Energy from time domain: E = A²τ = 1 × 10⁻⁶ V²·s (1 µJ into 1 Ω). Parseval check: ∫|X(f)|²df = A²τ²·(1/τ) = A²τ = 1 × 10⁻⁶ ✓.
• Shortening the pulse to 0.1 µs moves the first null to 10 MHz — ten times the bandwidth for one-tenth the duration.

Example 3 — Laplace: step response, initial and final value. H(s) = 10/(s² + 3s + 2) = 10/[(s+1)(s+2)], unit-step input.

• Y(s) = 10/[s(s+1)(s+2)]. Residues: at s = 0, 10/(1×2) = 5; at s = −1, 10/[(−1)(1)] = −10; at s = −2, 10/[(−2)(−1)] = 5.
• y(t) = (5 − 10e^−t + 5e^−2t)u(t). Check y(0) = 5 − 10 + 5 = 0 ✓ and y′(0) = 10 − 10 = 0 ✓ (second-order system, zero initial slope).
• Final value: lim_{s→0} sY(s) = 10/(1×2) = 5 — legitimate because the poles of sY(s) are at −1 and −2, both in the LHP. Equivalently H(0) = 10/2 = 5 (dc gain).
• Poles at −1 and −2 are real and in the LHP ⇒ stable, overdamped, no overshoot; the slower pole (−1) gives the dominant time constant 1 s, so settling takes roughly 4-5 s.
• If instead H(s) = 10/(s² − 3s + 2), poles are at +1 and +2: unstable, and the final value theorem would wrongly report 5.

Example 4 — Z-transform: ROC and inversion. (a) x[n] = (0.5)ⁿu[n] + 2(3)ⁿu[−n−1].

• First term: z/(z−0.5), ROC |z| > 0.5. Second term: 2(3)ⁿu[−n−1] ↔ −2z/(z−3), ROC |z| < 3.
• Overall X(z) = z/(z−0.5) − 2z/(z−3), ROC 0.5 < |z| < 3. The ring contains the unit circle, so the DTFT exists and the two-sided sequence is absolutely summable — even though it has a pole at z = 3, because that pole is "used" as a left-sided term.
• (b) Invert X(z) = z/[(z−1)(z−0.5)], ROC |z| > 1 (causal).
• X(z)/z = 1/[(z−1)(z−0.5)] = A/(z−1) + B/(z−0.5) with A = 1/(1−0.5) = 2 and B = 1/(0.5−1) = −2.
• X(z) = 2z/(z−1) − 2z/(z−0.5) ⇒ x[n] = 2(1 − 0.5ⁿ)u[n]: x[0] = 0, x[1] = 1, x[2] = 1.5, x[3] = 1.75 → 2.
• Long-division check: X(z) = z⁻¹/(1 − 1.5z⁻¹ + 0.5z⁻²) gives x[0] = 0, x[1] = 1 ✓. The pole at z = 1 means the system behind this is an accumulator — marginally stable, not stable.

Example 5 — 4-point DFT and the FFT saving. x[n] = {1, 2, 3, 4}, N = 4, W₄ = −j.

• X[0] = 1+2+3+4 = 10.
• X[1] = 1 + 2(−j) + 3(−1) + 4(+j) = (1−3) + j(−2+4) = −2 + 2j (magnitude 2.828, phase +135°).
• X[2] = 1 − 2 + 3 − 4 = −2.
• X[3] = 1 + 2(j) + 3(−1) + 4(−j) = −2 − 2j = X*[1] ✓ (real-input symmetry).
• Parseval: Σ|x[n]|² = 1+4+9+16 = 30; (1/N)Σ|X[k]|² = (1/4)(100 + 8 + 4 + 8) = (1/4)(120) = 30 ✓.
• If f_s = 8 kHz, the bins sit at 0, 2, 4 and 6 kHz (resolution f_s/N = 2 kHz).
• Cost comparison at N = 1024: direct DFT ≈ N² = 1,048,576 complex multiplications; radix-2 FFT ≈ (N/2)log₂N = 512 × 10 = 5,120 — a saving of about 205×.

Where it is used

• Digital filter and DSP design: IIR filters from analogue prototypes via the bilinear transform, FIR design by windowing the inverse DTFT; implemented on TI C6000/C2000 and ADI SHARC DSPs and on FPGAs at BEL, Bharat Dynamics, Saankhya Labs and defence DSP houses.
• Communication baseband: FFT/IFFT blocks are the core of 4G/5G and Wi-Fi modems and of DSL; channel estimation and equalisation are done bin-by-bin in the DFT domain — relevant to Tejas Networks, Qualcomm India, Nokia/Ericsson India R&D and C-DOT.
• Control and avionics: Laplace transfer functions, pole placement and discrete-time redesign for launch-vehicle and satellite attitude loops (ISRO/VSSC), process control at BHEL and L&T, and motor-drive controllers.
• Test, measurement and condition monitoring: FFT spectrum analysers, vibration analysis of bearings and gearboxes, and power-quality/THD measurement on the Indian grid — harmonic limits for grid connectivity are specified by CEA regulations, which follow IEEE 519-style limits, and equipment emission limits follow IEC 61000-3-2 (verify the current edition before quoting numbers).
• Biomedical and imaging: ECG/EEG denoising and QRS detection, MRI reconstruction (the scanner samples k-space, and the image is its inverse Fourier transform), ultrasound Doppler.
• Audio, speech and compression: STFT/MFCC front ends in speech recognition for Indian languages, and transform coding in MP3/AAC and the DCT in JPEG/H.264.
• Exams and selection tests: GATE EC/EE/IN, ISRO and BARC scientist/engineer papers, IES/ESE, and the first technical round for VLSI, embedded and DSP roles.

Interview questions

• Q: Why use the Laplace transform when the Fourier transform already gives the spectrum? A: Laplace converges for signals that grow with time, carries initial conditions through the derivative property, and shows stability directly through pole locations; Fourier is only its restriction to the jω axis and needs the system to be stable to exist.
• Q: What does the region of convergence tell you? A: It tells you which time-domain signal a given algebraic transform represents (causal, anticausal or two-sided) and whether the Fourier/DTFT exists — the ROC must contain the jω axis (Laplace) or the unit circle (Z).
• Q: Two signals have the identical X(z) = z/(z−2). How do you distinguish them? A: By the ROC: |z| > 2 means the causal, unstable sequence 2ⁿu[n]; |z| < 2 means the anticausal sequence −2ⁿu[−n−1], which is absolutely summable.
• Q: State the stability condition for continuous and discrete causal LTI systems. A: All poles strictly in the left half of the s-plane, and all poles strictly inside the unit circle of the z-plane, respectively.
• Q: When does the final value theorem fail? A: When sX(s) has poles on or to the right of the jω axis — for example an undamped oscillator or an unstable system, where no final value exists even though the algebra returns a number.
• Q: What is the difference between the DTFT and the DFT? A: The DTFT is continuous in frequency and 2π-periodic for an infinite sequence; the DFT is N samples of the DTFT of a finite N-point sequence, and it implicitly assumes periodic extension — hence circular convolution.
• Q: Why is the FFT faster than the DFT, and by how much? A: It reuses the symmetry and periodicity of the twiddle factors in a divide-and-conquer structure, reducing about N² complex multiplications to (N/2)log₂N — roughly 205× at N = 1024.
• Q: Does zero-padding improve frequency resolution? A: No. Resolution is set by the observation time (f_s/N of real data); zero-padding only interpolates the spectrum so peaks are displayed more smoothly.
• Q: How does z = e^sT explain aliasing? A: The map is many-to-one: every horizontal strip of the s-plane of height 2π/T maps onto the same z-plane, so analogue frequencies separated by multiples of the sampling rate land on the same point of the unit circle.
• Q: Why is the bilinear transform preferred over impulse invariance for IIR design? A: It is a one-to-one map from the whole LHP into the unit circle, so it cannot alias; the price is frequency warping, corrected by pre-warping the critical frequencies.
• Q: What is the Gibbs phenomenon, and what does it imply for filter design? A: The ~9% overshoot when a discontinuity is approximated by a truncated Fourier series; truncating an ideal filter's impulse response gives the same ripple, which is why windows (Hamming, Blackman) are used instead of a plain rectangular cut.
• Q: An FIR filter is always stable — why? A: Its system function is a polynomial in z⁻¹, so all poles sit at the origin, well inside the unit circle, and its impulse response is finite, hence absolutely summable.

Common mistakes

• Quoting a Z- or Laplace-transform expression without its ROC, and so getting causality and stability wrong.
• Applying the final value theorem to a system with poles on or right of the jω axis (or to z-domain poles on/outside the unit circle).
• Using the bilateral shift property z^−kX(z) on a unilateral problem that has non-zero initial conditions x[−1], x[−2].
• Confusing the two sinc conventions — sin(πx)/(πx) versus sin(x)/x — and placing the first null of a pulse spectrum at the wrong frequency.
• Multiplying DFTs without zero-padding to N₁ + N₂ − 1 and calling the wrapped circular result a linear convolution.
• Writing a Fourier transform for a periodic signal as an ordinary function instead of a weighted impulse train, or forgetting the πX(0)δ(ω) term in the integration property.
• Dropping the 1/|a| in time scaling, or the 1/2π in the inverse CTFT and in the multiplication property; and in the s-plane, confusing poles (natural response, stability) with zeros (blocked frequencies).

Related topics

Signals and systems · Sampling and digital modulation · Digital signal processing and digital filter design · Control systems: stability and frequency response · Network analysis and circuit transients · OFDM and MIMO.