The Gold Standard of Z-Transform

A Z-transform represents a sequence of discrete samples by a complex-domain series X(z)=Σx[n]z⁻ⁿ. It can turn filtering and recurrence questions into algebra. But the series is meaningful only with a definition of which integer indices are included and which complex z values give convergence.

Try beginning with three samples at indices 0, 1 and 2. The polynomial is x[0]+x[1]z⁻¹+x[2]z⁻², not a polynomial in continuous time. This small reconstruction catches common sign errors.

Contents · Next

The bilateral transform sums over every integer n. A one-sided transform begins at a specified boundary, commonly n=0, and is especially useful when a recurrence has nonzero initial conditions. The definitions must not be interchanged silently because time-shift identities acquire boundary contributions in unilateral calculations.

If a digital filter was running before observation began, its y[−1] may be nonzero. A zero-state transfer function cannot reconstruct that prior state; the one-sided transform or state-space equations must retain it.

Contents · Previous · Next

For x[n]=aⁿu[n] the transform is Σ_n=0^∞(az⁻¹)ⁿ. It becomes 1/(1−az⁻¹) only when |a/z|<1, or |z|>|a|. This convergence requirement is part of the answer, not decoration.

Let a=0.5 and z=1. The direct sum is 1+0.5+0.25+…=2, matching the expression. At z=0.5 the series diverges. A student who omits the inequality loses this distinction.

Contents · Previous · Next

The region of convergence, or ROC, is the set of complex values where the Z-transform’s defining infinite series exists. Rational expressions can have interior, exterior or annular ROCs depending on the underlying sequence. Poles themselves are excluded.

The same fraction 1/(1−2z⁻¹) can correspond to different right- or left-sided sequences under different convergence conditions. Request the ROC before selecting an inverse.

Contents · Previous · Next

Poles are singularities of a reduced rational transform, whereas zeros are roots of its numerator. Their locations influence frequency and transient response, but a pole-zero plot does not give initial conditions, and an input-output ratio does not necessarily reveal every internal state in a nonminimal system.

Write the rational function consistently in z or z⁻¹ and display cancellations openly. Treat a removed common factor as an algebraic simplification, not proof that every internal physical mode vanished.

Contents · Previous · Next

For bilateral transforms under the relevant conditions, delaying x[n] by k samples to x[n−k] multiplies X(z) by z^−k. An advance to x[n+k] introduces z^k. A delay does not automatically decrease amplitude, though it can change phase.

A unit pulse at n=0 has transform 1. Move it to n=2 and obtain z⁻². If your transform contains z², you have advanced the pulse instead.

Contents · Previous · Next

For an LTI filter under valid convergence conditions, y[n]=Σh[k]x[n−k] becomes Y(z)=H(z)X(z). This saves computation and makes filter transfer functions easier to study. It does not say that ordinary pointwise multiplication of samples is convolution.

Take a two-tap filter h[0]=1 and h[1]=1. Its output is x[n]+x[n−1], while its transfer function is 1+z⁻¹. Check both forms on a short pulse sequence.

For the finite-support boundary of transform multiplication, see short DFT products and folded linear-convolution tails with signed examples, residue-class identities and a support-based verification route.

The multiplication rule needs an existing convolution and a justified domain. Work through the bilateral convolution and convergence guide to compare convergent and divergent opposite-sided sequences, remainder bounds and a finite-kernel cancellation.

Contents · Previous · Next

The illustrative recurrence y[n]−0.5y[n−1]=x[n] gives H(z)=1/(1−0.5z⁻¹) for the zero-state response. Under the causal interpretation its pole lies at 0.5 and the ROC is outside the pole circle.

The first impulse-response samples are 1, 0.5, 0.25 and 0.125. These test values verify the transform independently of a software-generated pole-zero display.

Contents · Previous · Next

Causality concerns whether current outputs depend only on current and past inputs. BIBO stability concerns whether every bounded input yields a bounded output; for an LTI system, absolute summability of h[n] is the standard test.

For a causal reduced rational filter, poles strictly inside the unit circle establish BIBO stability. Do not apply that condition to an unspecified noncausal ROC or use it to certify unobserved internal dynamics.

For a worked extension of the time-direction distinction, see stable forward filters and unstable causal inverses with causal and anti-causal impulse responses, finite-delay reconstruction and explicit error bounds.

Contents · Previous · Next

Where the unit circle is contained in the ROC, substituting z=e^jω yields the discrete-time Fourier transform. The digital angular frequency ω has units of radians per sample. This relates pole-zero geometry to filter magnitude and phase.

Substituting the unit circle into a formal rational expression outside the ROC need not produce a convergent ordinary Fourier transform. Verify the convergence premise first.

Contents · Previous · Next

Zeros can suppress a response near their angles on the unit circle, while poles near that circle can enhance sensitivity. The exact outcome depends on all zeros, poles and gain. Frequency magnitude and phase are separate properties.

An ideal zero at z=−1 can cancel a component at digital frequency π. It does not imply complete removal of every high-frequency disturbance or identical time-domain behaviour across filters.

A nominal pole plot leaves coefficient uncertainty unresolved. The second-order coefficient-set stability guide proves the quadratic criterion, finds exact worst margins over boxes and distinguishes fixed coefficients from switching.

Contents · Previous · Next

A finite-impulse-response filter has finitely many nonzero impulse coefficients; an infinite-impulse-response filter generally has a continuing recursive tail. A response can become numerically tiny while remaining mathematically infinite.

A three-sample moving average is FIR; y[n]=0.5y[n−1]+x[n] is IIR. Both may be stable, but their transient, phase and implementation properties differ.

Contents · Previous · Next

State-space models track internal state variables, while transfer functions usually describe zero-state input-output relationships. Different state-space realisations can generate the same reduced transfer function. Hidden or cancelled modes can make internal stability a separate issue.

An engineer must choose the representation appropriate to the question. Testing a filter’s frequency response is not a complete safety test for a physical feedback system.

For an explicit state-space counterexample, study cancelled transfer poles and hidden internal modes with original state trajectories, rank tests and small-coupling bounds.

Contents · Previous · Next

A one-sided Z-transform of a recurrence must preserve initial conditions. For y[n]=a y[n−1]+x[n], the pre-start y[−1] influences y[0]. Forgetting it silently changes the mathematical problem to initial rest.

Suppose a=0.5, y[−1]=4 and x[0]=1. Then y[0]=3, not the initial-rest value 1. This small calculation catches a lost boundary term.

Contents · Previous · Next

Partial fraction decomposition breaks a rational transform into terms with familiar inverses. A/(1−pz⁻¹) corresponds to A pⁿu[n] for an appropriate right-sided ROC. Repeated poles require additional time factors.

The inverse is not unique if the ROC is omitted. Check the first reconstructed samples against the given recurrence or direct summation before accepting the result.

Contents · Previous · Next

For the fictional sequence x[n]=0.5ⁿu[n], X(z)=1/(1−0.5z⁻¹) with ROC |z|>0.5. At z=1 the sum is 2. The same impulse sequence is absolutely summable and therefore supports a stable causal LTI filter.

This computation links time samples, a pole, the convergence domain and a stability test. Omitting any one piece makes the interpretation less reliable.

Contents · Previous · Next

For y[n]=0.5y[n−1]+1 with y[−1]=0, the first values are 1, 1.5, 1.75 and 1.875. The closed form is y[n]=2(1−0.5ⁿ⁺¹) for n≥0.

The long-run value is 2, matching the DC transfer gain. Recheck n=0: the expression produces 1. An incorrect exponent n instead of n+1 fails the very first sample.

Contents · Previous · Next

Let fictitious H(z)=(1−0.25z⁻¹)/(1−0.5z⁻¹). Equivalently H(z)=(z−0.25)/(z−0.5). The zero lies at 0.25 and the pole at 0.5.

The DC gain H(1) is 1.5. Under the causal ROC outside 0.5, the unit circle lies within the ROC, supporting the ordinary stable input-output interpretation.

Contents · Previous · Next

For causal h[n]=0.8ⁿu[n], Σ|h[n]|=1/(1−0.8)=5. Because the sum is finite, the LTI system is BIBO stable. Its single pole lies inside the unit circle.

Replace 0.8 by 1.1 and the corresponding causal sum diverges. A similar-looking rational expression is not sufficient evidence of stability without the time interpretation.

Contents · Previous · Next

The fictional filter y[n]=(x[n]+2x[n−1]+x[n−2])/4 has H(z)=0.25+0.5z⁻¹+0.25z⁻². At z=1 its value is 1, preserving a constant input.

At z=−1 its value is zero, cancelling the alternating signal at digital frequency π in this ideal model. The result describes specific frequencies, not universal noise removal.

Contents · Previous · Next

Digital frequency ω is in radians/sample; conversion to hertz requires a sample rate. At 1000 samples/second, ω=π/2 corresponds to 250 Hz, because (π/2)/(2π)×1000=250.

Aliasing means continuous frequencies separated by the sampling rate may become indistinguishable in sampled data. Never report a physical frequency from an unlabeled discrete-time plot.

Contents · Previous · Next

SciPy’s tf2zpk and freqz can calculate poles, zeros and frequency responses from filter coefficients. Their documentation specifies ordering and numerical conventions. A computed curve can be mathematically correct for the wrong coefficient array supplied by the user.

Compare at least one hand-calculated DC gain, the first few impulse samples and pole locations before treating the software result as validated engineering evidence.

For explicit arithmetic-state checks, study rounding in a stable stored-state recurrence with nearest-rounding fixed points, two-cycles, error bounds and signed truncation comparisons.

Contents · Previous · Next

Common mistakes include changing the sign of a delay, dropping the ROC, using a zero-state transfer function with nonzero initial conditions, confusing the digital frequency axis with hertz and assuming all rational filters are stable.

Physical systems also introduce quantisation, numerical precision, clock variability and unmodelled dynamics. A nominal Z-domain model is one piece of engineering evidence, not an exhaustive safety or performance assessment.

Contents · Previous · Next

The gold standard is a reproducible sequence: state x[n], define X(z), establish the ROC, interpret poles and zeros, invert with correct initial conditions and check the physical or simulated output. Another competent analyst should be able to reconstruct the important steps.

Continue with Signal Processing, Control Theory and Fourier Analysis. The method exemplifies Learning Transfers into Life.

Contents · Previous · Sources