Inverse Z-Transform Methods | DSP-101

Why We Need the Inverse Z-Transform

The Z-transform converts a discrete-time sequence x[n] into a function X(z) in the complex plane, making analysis and filter design algebraically tractable. But the real payoff comes when we need to go the other direction: given a transfer function H(z) or an output spectrum Y(z), recover the time-domain sequence y[n]. This is the inverse Z-transform problem, and it arises constantly — whenever you design a filter from a pole-zero specification, when you solve difference equations, or when you implement a system in hardware.

Three practical methods dominate: partial fraction expansion (the most systematic), power-series / long division (the most direct), and for completeness the formal contour integral (seldom computed by hand but important for theory). Each method illuminates a different facet of the Z-transform relationship.

Method 1 — Partial Fraction Expansion

Partial fractions work whenever X(z) is a ratio of polynomials with simple poles. The strategy is to decompose X(z)/z (or X(z) itself) into a sum of first-order terms whose inverse transforms are individually recognizable from the Z-transform table.

Partial Fraction Decomposition

X(z) = \sum_{k=1}^{N} \frac{A_k\,z}{z - p_k}, \quad A_k = \left.\frac{X(z)}{z}\,(z-p_k)\right|_{z=p_k}

X(z) is expanded into first-order terms. Each residue A_k = (z − p_k) X(z)/z evaluated at z = p_k. The ROC determines whether each term corresponds to a causal (right-sided) or anti-causal (left-sided) sequence.

The key Z-transform pair to remember is: a causal exponential aⁿu[n] transforms to z/(z−a) with ROC |z| > |a|. So once X(z) is in partial fraction form, each term A_k z/(z−p_k) maps directly to A_k p_kⁿu[n] in the causal case.

ROC Determines Causality

The same algebraic form of X(z) can correspond to different time sequences depending on the ROC. Always specify the ROC alongside X(z) — the inverse transform is not unique without it.

ROC: |z| > |p| → causal sequence | ROC: |z| < |p| → anti-causal sequence

For repeated poles of order r, the decomposition requires higher-order terms and corresponding table entries. In practice, most DSP filter designs use simple poles (no repeated roots), so the basic partial fraction recipe handles the vast majority of real-world cases.

Method 2 — Power Series / Long Division

If you only need the first few samples of x[n] — or if X(z) does not factor nicely — long division of the numerator polynomial by the denominator polynomial yields a power series in z⁻¹. The coefficients of this series are directly the samples x[0], x[1], x[2], …

Power Series Expansion

X(z) = \frac{b_0 + b_1 z^{-1} + \cdots}{1 + a_1 z^{-1} + \cdots} = x[0] + x[1]z^{-1} + x[2]z^{-2} + \cdots

Dividing the numerator by denominator in ascending powers of z⁻¹ yields x[0] + x[1]z⁻¹ + x[2]z⁻² + … Reading off the coefficients gives the time-domain sequence sample by sample.

Long division is especially useful for causal sequences (power series in z⁻¹) and for verifying partial-fraction results. It is also the computational basis for converting a transfer function into a direct-form difference equation — the recipe a processor follows sample by sample.

Step 1

Arrange Polynomials

Write numerator and denominator in descending powers of z (or ascending powers of z⁻¹ for causal systems). Divide the leading term of the numerator by the leading term of the denominator.

Step 2

Multiply and Subtract

Multiply the current quotient term by the full denominator, subtract from the running numerator, and bring down the next term. Repeat to generate as many coefficients as needed.

Step 3

Read Off Samples

Each quotient coefficient in the power series in z⁻¹ is a direct time-domain sample: the coefficient of z⁻ⁿ is x[n].

Step 4

Stop When Satisfied

Long division can continue indefinitely for IIR systems. Stop once you have enough samples for your application, or use it to verify the first few values from partial fractions.

Method 3 — The Contour Integral (Formal Definition)

The formal definition of the inverse Z-transform is a contour integral in the complex plane. While rarely computed by hand, it underpins the theory and explains why the ROC matters so fundamentally.

Inverse Z-Transform — Formal Definition

x[n] = \frac{1}{2\pi j}\oint_{C} X(z)\,z^{n-1}\,dz = \sum_{\text{poles inside }C}\operatorname{Res}\bigl[X(z)z^{n-1}\bigr]

The contour C is a counterclockwise circle in the ROC that encloses the origin. By Cauchy's residue theorem, x[n] equals the sum of residues of X(z)z^{n−1} at all poles inside C — which recovers the partial-fraction result automatically.

In practice, engineers evaluate this integral by residues rather than numerical integration. The contour integral perspective clarifies two things: (1) the ROC determines which poles are enclosed and therefore which partial-fraction terms contribute, and (2) closed-form evaluation is possible precisely because rational X(z) has isolated poles with computable residues.

Worked Example — Second-Order Filter

Consider the transfer function H(z) = z² / (z² − 0.9z + 0.81) with ROC |z| > 0.9. The denominator has roots at z = 0.9e^±j60°, giving a conjugate pole pair at radius 0.9 and angle ±π/3. We want the impulse response h[n].

Partial Fraction of Example H(z)

H(z) = \frac{z^2}{z^2 - 0.9z + 0.81} = \frac{A\,z}{z - 0.9e^{j\pi/3}} + \frac{A^*\,z}{z - 0.9e^{-j\pi/3}}

With poles at p = 0.9e^{jπ/3} and p* = 0.9e^{−jπ/3}, the residues are complex conjugates. Combining the conjugate pair gives a real-valued sinusoidal impulse response h[n] = (2/√3) · (0.9)ⁿ · cos(nπ/3 − π/6) u[n]. The amplitude 2/√3 and the phase −π/6 are not decoration: with a bare z² numerator the residues are not purely real, and dropping them would give h[1] = 0.45 where the true value is 0.9.

This example illustrates the general pattern: conjugate pole pairs at radius r and angle θ always produce impulse responses of the form rⁿcos(nθ + φ)u[n] — damped sinusoids. The radius controls the decay rate, and the angle sets the oscillation frequency. This is the prototype response of every resonant IIR section.

Choosing the Right Method

Use When

Partial Fractions

You need a closed-form expression for x[n] for all n. Best for rational X(z) with simple poles. Most systematic; directly links poles to exponential components.

Use When

Long Division

You only need the first few samples, or X(z) is hard to factor. Also great for verification and understanding how a filter builds its output recursively.

Use When

Z-Transform Tables

X(z) matches a standard form directly (unit step, exponential, ramp, sinusoid). Always check the table first — it may save significant algebra.

Use When

Contour Integral

You need theoretical justification or are working in the complex analysis framework. In practice, residue evaluation equals the partial-fraction result.

With Module 4 complete, you have the full Z-transform toolkit: forward transform, ROC, poles and zeros, and now the inverse transform. Module 5 will apply these tools to the design and implementation of digital filters.

Key Takeaways

The inverse Z-transform recovers the time-domain sequence x[n] from its Z-domain representation X(z); the ROC is essential — the same X(z) can correspond to different sequences for different ROCs.
Partial fraction expansion decomposes X(z)/z into first-order terms whose inverses are recognizable table entries (exponentials, sinusoids, ramps).
Long division generates x[n] sample by sample as the power-series coefficients of X(z) in z⁻¹; useful when only a finite number of samples are needed.
The contour integral is the formal definition; in practice it reduces to summing residues, which equals the partial-fraction result.
Conjugate pole pairs at radius r and angle θ produce damped sinusoidal responses rⁿcos(nθ + φ)u[n] — the signature of every resonant IIR section.
Always specify the ROC: ROC |z| > |p| selects the causal sequence; ROC |z| < |p| selects the anti-causal sequence for the same pole.
When in doubt, use long division to verify the first few values produced by partial fractions.