The Gold Standard of Z-Transform | Why a Stable Causal Filter Can Require an Unstable Causal Inverse

Let x[n] be a two-sided input sequence and define y[n]=x[n]−2x[n−1]. The current output uses the current and previous input only. It is causal: it needs no input from a future index.

If |x[n]|≤M for every n, then |y[n]|≤3M by the triangle inequality. Its impulse response has h[0]=1, h[1]=−2 and all other samples zero. The sum of absolute impulse-response values is 3, so the forward filter is bounded-input bounded-output stable.

The word stable here describes the mapping from x to y. It does not promise that solving backward from y to x has the same property. Multiplication by a number can be easy while division near zero is sensitive; a filter has a time-dependent version of that distinction. We will identify the exact obstruction rather than infer it from the size of the coefficient 2 alone.

Contents · Next chapter

An inverse G should recover x from the complete filtered sequence y for the declared class of signals. In transform notation we want G(z)H(z)=1, where H(z)=1−2z⁻¹. Algebra therefore suggests G(z)=1/(1−2z⁻¹).

That rational expression is not yet a complete sequence specification. A bilateral Z-transform requires a region of convergence, or ROC. Different expansions of the same fraction can represent different time directions.

We will first require a causal inverse and then require a stable inverse. Those requirements lead to different expansions because the reciprocal has a pole at z=2, outside the unit circle. Our signal class will also matter: a formula that recovers every bounded two-sided input can fail for an unbounded exponential. Stating the allowed sequences keeps algebraic inversion connected to a well-defined operation on actual sample records.

Contents · Previous chapter · Next chapter

Rearrange the forward rule as x[n]=y[n]+2x[n−1]. Starting from a declared initial value, this computes the current recovered input from the current observation and the previous recovered value. It is a causal recursion.

From zero state, an impulse at y[0] produces recovered samples 1, 2, 4, 8 and so on. Thus g_c[n]=2ⁿ for n≥0 and zero for n<0. Its Z-transform is the geometric series Σₙ≥₀(2z⁻¹)ⁿ, converging only when |z|>2.

The inverse is algebraically correct on sequences and initial conditions for which the recursion is defined. Its impulse response is not absolutely summable. A bounded observation consisting of one impulse gives an unbounded recovered sequence. Therefore it fails bounded-input bounded-output stability as an operator from arbitrary bounded y to the recovered result. Exact algebra has not made that failure disappear.

Contents · Previous chapter · Next chapter

Suppose two causal reconstructions see exactly the same y but begin with a difference e[−1] in their previous recovered value. Subtracting the two recursions gives e[n]=2e[n−1], hence e[n]=2ⁿ⁺¹e[−1].

If e[−1]=10⁻⁶, the difference at n=19 is about 1.049. No additional observation noise is needed after initialization. The numbers are dimensionless teaching values, not a proposed hardware precision requirement.

This is a clean way to see why a recursion can reconstruct a particular noiseless record correctly and still be unstable. The correct initial value selects one trajectory exactly; small discrepancies select different trajectories that separate exponentially. A successful demonstration using specially prepared exact data does not establish robustness to uncertainty in the boundary value or the recorded samples. To judge an inverse, examine the operator and its admissible perturbations as well as one ideal trajectory.

Contents · Previous chapter · Next chapter

Rewrite G(z)=−(z/2)/(1−z/2). Expanding in positive powers of z gives G(z)=−Σₖ≥₁2⁻ᵏzᵏ, valid when |z|<2.

The corresponding sequence is g_s[n]=−2ⁿ for n≤−1 and zero for n≥0. Every nonzero impulse-response sample lies at a negative index. This is an anti-causal inverse: its output at time n depends on future values of its input y.

Its absolute impulse-response sum is Σₖ≥₁2⁻ᵏ=1, so it is a stable convolution operator on bounded sequences. Its ROC includes the unit circle. The causal and stable expressions are the same rational fraction but different sequences, with disjoint convergence regions. Writing the ROC is what prevents these different implementations from being confused. A pole outside the unit circle forbids a stable causal choice here; it does not forbid every stable two-sided choice.

Contents · Previous chapter · Next chapter

Convolution with g_s gives x[n]=−Σₖ≥₁2⁻ᵏy[n+k]. To recover the current x[n], the formula reads y[n+1], y[n+2], and every later output, with progressively smaller weights.

It uses no y[n] term. That may initially look strange because y[n] contains x[n]. The next sample also contains x[n], with coefficient −2, and successive future samples allow the unwanted future x terms to cancel in a telescoping sum.

If |y[n]|≤B, the absolute size of the recovered result is at most B because the absolute weights sum to one. This operator bound is valid for arbitrary bounded two-sided observation sequences. It does not say that every such observation sequence came from a particular physical experiment or a particular unknown input. It states exactly how the declared inverse acts on the sequence supplied to it.

Contents · Previous chapter · Next chapter

Take only K future terms and substitute y[n+k]=x[n+k]−2x[n+k−1]. Then −Σₖ₌₁ᴷ2⁻ᵏy[n+k]=x[n]−2⁻ᴷx[n+K]. Every intermediate x term cancels.

If x is bounded by M, the remaining term has absolute size at most M2⁻ᴷ, which tends to zero as K increases. This proves that the stable anti-causal sum recovers every bounded two-sided input exactly in the infinite-preview limit.

The boundary condition is essential. For x[n]=2ⁿ, the forward output is zero for every n, yet x is not the zero sequence. That input is unbounded over the two-sided record, and the terminal term 2⁻ᴷx[n+K] stays equal to 2ⁿ instead of vanishing. The proof fails at precisely the point where the signal violates the assumption. A formula cannot establish uniqueness on a larger class than its boundary argument supports.

Contents · Previous chapter · Next chapter

Suppose two bounded input sequences give exactly the same y. Their difference d obeys d[n]=2d[n−1]. Any solution has the form d[n]=c2ⁿ after fixing an index origin.

For c≠0 this grows without bound as n tends to positive infinity. Therefore no nonzero difference of this form belongs to the bounded two-sided signal class. The two bounded inputs must be identical.

This distinguishes two claims that sound similar: the forward recurrence has a nonzero homogeneous exponential solution, and the forward map has a nontrivial kernel among bounded two-sided inputs. The first is true; the second is false. A signal class and boundary condition determine which homogeneous solutions are admitted. The stable anti-causal inverse selects the bounded solution by its condition at the future end, whereas a causal recursion selects a solution through an initial value at the past end.

Contents · Previous chapter · Next chapter

Let x[0]=1, x[1]=−1, x[2]=2 and x[n]=0 at every other integer index. Compute y using the forward rule, including the sample after the last nonzero x.

For n=0, the stable inverse gives −[y[1]/2+y[2]/4+y[3]/8]=−[−3/2+1−1/2]=1. For n=1 it gives −[y[2]/2+y[3]/4]=−[2−1]=−1. For n=2 it gives −y[3]/2=2.

If y[3] is omitted, the reconstruction is wrong near the end. The last nonzero input leaves a delayed contribution in the next output. Recording only through the last nonzero x sample silently discards part of the observation needed for this inverse.

Contents · Previous chapter · Next chapter

Define the K-term approximation x̂_K[n]=−Σₖ₌₁ᴷ2⁻ᵏy[n+k]. It requires K future samples rather than infinitely many. The telescoping identity already gives its exact error: x̂_K[n]−x[n]=−2⁻ᴷx[n+K].

For the original short sequence at n=0, K=1 gives 1.5, K=2 gives 0.5, and K=3 gives 1. These errors match −x[1]/2=0.5, −x[2]/4=−0.5, and −x[3]/8=0.

The approximation error need not shrink monotonically at a fixed index as K increases, because x[n+K] changes. Its worst-case bound M2⁻ᴷ decreases monotonically when the same global M is valid. This is a useful distinction between a guaranteed envelope and the particular errors of one sequence. A moving value inside a shrinking envelope can still move in either direction.

Contents · Previous chapter · Next chapter

The K-term operator is G_K(z)=−Σₖ₌₁ᴷ2⁻ᵏzᵏ. Multiplying by H(z)=1−2z⁻¹ gives G_K(z)H(z)=1−2⁻ᴷzᴷ.

The product is not exactly one. The extra advance term maps x[n] to the residual −2⁻ᴷx[n+K]. This polynomial identity is a second proof of the time-domain error formula, obtained without an infinite series.

It also shows why calling a finite filter an “exact inverse” without qualification is misleading. It is exact for a particular target sample if the relevant terminal input x[n+K] is known to be zero. Otherwise it is an approximation with a stated remainder. Exactness can follow from a valid finite-support condition, but it cannot be inferred merely from choosing a long filter or observing that the numerical error looks small in one test.

Contents · Previous chapter · Next chapter

A live system cannot read a sample before it arrives. Multiply the finite operator by z⁻ᴷ to get F_K(z)=−Σₖ₌₁ᴷ2⁻ᵏzᵏ⁻ᴷ. Its powers are nonpositive, so it is a causal finite impulse-response filter.

At current time n, its output is x̂_K[n−K]=−Σₖ₌₁ᴷ2⁻ᵏy[n−K+k]. The newest value used is y[n]; the oldest is y[n−K+1]. Its target is the input from K samples earlier.

The composed response is F_K(z)H(z)=z⁻ᴷ−2⁻ᴷ. Thus the output equals x[n−K]−2⁻ᴷx[n]. This delay formulation is implementable with a finite buffer and has the same magnitude error bound M2⁻ᴷ relative to its delayed target. Adding delay changes when an estimate becomes available; it does not turn the finite approximation into an exact same-time inverse.

Contents · Previous chapter · Next chapter

Suppose the input is known to satisfy |x[n]|≤8 and the allowed approximation error is at most 0.01 in the same units. Require 8·2⁻ᴷ≤0.01, or K≥log₂800. The smallest integer choice is K=10.

At K=9 the bound is 8/512=0.015625, which is too large. At K=10 it is 8/1024=0.0078125, which meets the stated requirement. The bound is conservative for many individual sequences but is attainable when the terminal value has magnitude 8.

If the sampling rate is 1000 samples per second, ten samples correspond to 0.01 seconds of latency. That is an arithmetic conversion for this fictional design requirement, not a recommendation for a particular application. A real choice must consider observation noise, coefficient uncertainty, boundaries and what delay the task permits. The truncation calculation addresses only the component of error it explicitly models.

Contents · Previous chapter · Next chapter

Let the recorded sequence be ỹ[n]=y[n]+v[n], with |v[n]|≤η. The finite-preview estimate acquires noise contribution −Σₖ₌₁ᴷ2⁻ᵏv[n+k]. Its absolute size is bounded by η(1−2⁻ᴷ).

Combining it with truncation gives |x̂_K[n]−x[n]|≤M2⁻ᴷ+η(1−2⁻ᴷ). The same formula applies to the delayed implementation relative to x[n−K]. This is a deterministic worst-case bound; it is not a confidence interval and does not require independent random noise.

Increasing K reduces the truncation term while the noise bound approaches η. There is no promise of unlimited improvement toward zero error if the measurement itself remains uncertain. For noise with a specified stochastic structure, a variance analysis could answer a different question, but it should not replace or be confused with this bound. Here every possible bounded noise sequence is allowed, including aligned errors that make the absolute-weight bound tight.

Contents · Previous chapter · Next chapter

The stable inverse for coefficient 2 has absolute gain bound one. That does not make the forward filter minimum phase or make its causal inverse stable. It says that the noncausal inverse is a bounded operator on the stated sequence class.

To see conditioning separately, replace 2 by a>1. The stable inverse weights are −a⁻ᵏ for k≥1, whose absolute sum is 1/(a−1). As a approaches 1 from above, this bound becomes large even though the inverse remains stable for each fixed a>1.

For a=1.01, the bound is 100. Observation noise bounded by η can therefore produce inverse error up to 100η in the unrestricted worst case. Stability asks whether a finite bound exists; conditioning asks how large or sensitive that bound is for the problem at hand. Treating those questions separately helps explain why a mathematically stable operation can still be practically unhelpful with noisy data.

Contents · Previous chapter · Next chapter

For H_a(z)=1−az⁻¹ with a>1, the stable anti-causal inverse is −Σₖ≥₁a⁻ᵏzᵏ. Its K-term reconstruction obeys x̂_K[n]=x[n]−a⁻ᴷx[n+K]. The bounded-input truncation error is at most Ma⁻ᴷ.

With bounded observation noise η, the K-term noise bound is η(1−a⁻ᴷ)/(a−1). For delayed implementation, multiply by z⁻ᴷ exactly as before. The combined bound is Ma⁻ᴷ+η(1−a⁻ᴷ)/(a−1).

These formulas follow by the same finite telescoping argument, so they can be tested without trusting a software inverse routine. Notice that the two terms respond differently as a approaches one: the remainder decays more slowly with K, while the infinite inverse gain increases. A zero close to the unit circle can therefore require both a longer preview and better evidence about noise. The simple coefficient 2 was chosen to make those quantities easy to compute, not because it describes every filter.

Contents · Previous chapter · Next chapter

Use y[n]=x[n]−3x[n−1], input bound M=9 and K=4 future samples. Before reading further, find the stable inverse weights, causal inverse ROC, delayed target and truncation bound.

The stable weights are −3⁻ᵏ for k≥1, with ROC |z|<3 and absolute sum 1/2. The causal inverse has impulse response 3ⁿ for n≥0 and ROC |z|>3, so it is unstable. The delayed finite output targets x[n−4] and equals x[n−4]−x[n]/81.

The truncation bound is 9/81=1/9. If observation noise is bounded by η=0.02, the finite noise bound is 0.02(1−1/81)/2=0.009876543… . Their sum is approximately 0.120988. Recompute these values from the geometric sum, then verify the product of the four-term inverse polynomial and the forward filter. Changing the coefficient tests the reasoning rather than the ability to remember the original powers of two.

Contents · Previous chapter · Next chapter

Change the forward filter to H_m(z)=1−(1/2)z⁻¹. Its zero lies at z=1/2. The causal reciprocal has impulse response (1/2)ⁿ for n≥0, ROC |z|>1/2 and absolute sum 2. It is both causal and stable.

The relevant difference is the location of the forward zero, which becomes a pole of the reciprocal. A stable forward FIR filter can have a zero outside the unit circle because its stability depends on its finite impulse response, not on putting every zero inside the circle.

The standard minimum-phase condition provides a stable causal inverse under the usual rational-filter conventions, including attention to pure delay. Our original H fails that condition. This comparison changes one coefficient and directly checks the inverse response, avoiding the misleading shortcut that all stable filters are automatically safe to invert causally.

Contents · Previous chapter · Next chapter

A pure delay H_d(z)=z⁻¹ is a stable causal forward filter. Its exact reciprocal is z, which advances by one sample: x[n]=y[n+1]. There is no unstable exponential in that inverse, yet it is noncausal for same-time recovery.

This example shows why a statement about inverse poles alone may omit timing requirements. A delayed recovery can be exact here: output y[n] directly estimates x[n−1]. There is no truncation remainder because the filter only shifted the sequence.

For the original H=1−2z⁻¹, finite delay helps produce a controlled approximation, but no fixed finite delay removes its infinite anti-causal tail for arbitrary bounded inputs. Multiplying the exact inverse by a finite delay still leaves some future dependence relative to the current time. Distinguishing pure transport delay from the tail caused by an outside-unit-circle zero prevents two different obstacles to same-time inversion from being merged.

Contents · Previous chapter · Next chapter

Suppose y is recorded only through index T. For a target x[n] with n<T, the maximum available preview is K=T−n. The finite formula leaves remainder −2⁻⁽ᵀ⁻ⁿ⁾x[T]. Near the end, the bound is larger because less future information is available.

If x[T] is known to be zero from a justified stopping condition, the formula is exact for earlier targets. If it is merely assumed zero by padding the record, the reconstruction inherits that assumption. Padding is not an additional measurement.

A finite-record report should label which target indices were estimated, how much preview each used and what boundary value or bound was supplied. A single error number for the whole record can hide severe edge uncertainty. In the original short sequence, including y[3] was decisive because it recorded the delayed effect of the final nonzero input x[2]. The bookkeeping is part of the inverse problem.

Contents · Previous chapter · Next chapter

From y[n+1]=x[n+1]−2x[n], solve x[n]=(x[n+1]−y[n+1])/2. A backward recursion starts from a terminal value and walks toward smaller n. An error in that terminal value is halved at each backward step.

This is the finite-record counterpart of the stable anti-causal inverse. Starting with a terminal guess of zero gives the finite weighted sum and its terminal remainder. Starting with the true terminal value removes that remainder for the observed interval.

The forward recursion doubles initial error; the backward recursion halves terminal error. Neither is magically free of boundary information. They place the boundary at opposite ends of the record and propagate its uncertainty differently. This comparison can be reproduced with the same short sequence, so a learner can see time direction as an operational choice rather than a notation attached to a rational fraction.

Contents · Previous chapter · Next chapter

If the true initial value is known exactly and observations are noiseless, causal recursion can recover the corresponding exact finite sequence. That useful special case does not make the causal inverse stable to arbitrary bounded perturbations.

If future observations can be collected, delayed recovery becomes possible with the error bounds already derived. If a valid terminal condition is known, finite backward recursion can be exact. If neither boundary is known, a model or prior assumption may select a reconstruction, but its result remains conditional on that choice.

For a physical measurement chain, independent calibration can establish whether the proposed forward filter is a reasonable approximation. The equation y[n]=x[n]−2x[n−1] was stipulated here. A good inverse of an incorrect forward model can still mislead. More elaborate inverse algorithms do not eliminate the need to validate the original relation between input and observation.

Contents · Previous chapter · Next chapter

Julius O. Smith’s chapter on minimum-phase filters explains the relationship between zero locations and a stable inverse under its filter conventions. It also highlights timing issues such as poles at infinity and the treatment of delay. The worked coefficients, sequences and finite-preview remainder in this article are independently constructed.

When comparing terminology across books, check whether “minimum phase” includes causality, stability, invertibility and an exclusion of pure delay in the stated definition. Do not lift a pole-location slogan out of a definition that supplied other conditions.

Our example avoids that ambiguity by writing the actual sequences, their absolute sums and their convergence regions. Even if a source chooses a different naming convention, the computations still specify which inverse uses past or future data and whether a bounded observation perturbation has a bounded effect.

Contents · Previous chapter · Next chapter

Undoing a known convolution is the setting of this inverse example. Real deconvolution can add uncertain wavelets, incomplete bandwidth, multidimensional propagation and noise. The Archaeology article on seismic deconvolution, wavelet estimation and ringing artefacts considers those evidence problems.

The transferable lesson is to specify the forward operator, boundary information and permitted time direction before interpreting a sharpened result. The coefficient 2 and the exact geometric remainder should not be imported into a seismic record whose wavelet has not been established.

Regularization can trade exact inversion for controlled behaviour under a chosen criterion. A useful comparison must state that criterion, the noise model and the validation evidence. This article’s finite-delay construction makes its own tradeoff explicit through a closed-form remainder. It is one transparent example of conditional reconstruction, not a universal replacement for domain-specific inversion methods.

Contents · Previous chapter · Next chapter

The expression 1/H(z) alone leaves the ROC unresolved. For our example, one expansion is causal and unstable; the other is anti-causal and stable. A software routine that returns coefficients may also assume an initial state and a time direction that must be checked.

A complete implementation description therefore includes the recurrence or convolution, boundary condition, data window and target timestamp. If a delayed output estimates an earlier input, label that earlier index. Otherwise a numerically good delayed estimate can be misreported as a same-time prediction.

A useful repair is to ask for the first three nonzero impulse-response samples and their indices. The causal inverse gives 1,2,4 at indices 0,1,2. The stable inverse gives −1/2,−1/4,−1/8 at indices −1,−2,−3. The numbers and indices expose a difference that the common rational fraction conceals.

Contents · Previous chapter · Next chapter

More terms shrink the worst-case truncation bound when the bounded-input assumption is valid. They do not remove observation error, model mismatch or an unknown boundary that violates the signal class. The noise bound in our example approaches η rather than zero.

A longer filter may also be unavailable near the edge of a finite record, and a live application may not permit its added delay. These are constraints on the experiment and task, not failures of the telescoping identity.

Judge improvement against a stated target: worst-case approximation error, average squared error under a specified stochastic model, latency or robustness to model uncertainty. Those targets need not select the same K. Our explicit bound answers one of them. Calling a larger K universally better would quietly add claims that the derivation never established.

Contents · Previous chapter · Next chapter

Is the forward filter unstable because its zero is at 2?

No. Its impulse response is finite and has absolute sum 3. The outside-unit-circle zero obstructs a stable causal inverse, not forward bounded-input bounded-output stability.

Can the stable inverse run live without delay?

Not as an exact same-time operation on arbitrary bounded inputs. It requires future samples. A finite delayed approximation is causal and has the stated remainder.

Does an exact initial value rescue causal inversion?

It selects the exact noiseless trajectory for that experiment. It does not make the operator robust to arbitrary bounded observation errors or initialization discrepancies.

Why does the anti-causal inverse omit y[n]?

Future outputs contain the required x[n] term, and their other input contributions telescope. The finite identity shows the cancellation directly.

Are these formulas confidence intervals?

No. They are deterministic equalities and worst-case bounds under explicit magnitude assumptions. A statistical interval would require an additional probability model and a stated estimation procedure.

Contents · Previous chapter · Next chapter

First compute the original five-row sequence and verify the three reconstructions at n=0. Next multiply G_KH for K=1,2 and3, checking the surviving terminal term. Then implement the delayed form and label its output as an estimate of x[n−K].

After those checks, change the coefficient to 3 and repeat the exercise without copying the powers of two. Finally change it to 1/2 and explain why the causal inverse now has a summable impulse response. Each change tests a distinct part of the argument: coefficient size, time direction and convergence.

If a result disagrees, check sample indexing before adjusting a numerical tolerance. A one-sample shift changes which terminal input controls the error. Keeping the finite telescoping sum beside the transform algebra makes such errors easier to locate and ensures that the final conclusion belongs to the operation actually performed.

Contents · Previous chapter · Next chapter

For a bounded constant input x[n]=c at every integer index, the original forward output is y[n]=−c. The finite-preview estimate is c(1−2⁻ᴷ). Its error is exactly −c2⁻ᴷ, attaining the truncation bound when |c|=M. This is a simple test of both the sign and the claim that the worst-case bound can be reached.

For x[n]=c(−1)ⁿ, the forward output is 3c(−1)ⁿ. The finite error is −2⁻ᴷc(−1)ⁿ⁺ᴷ, so its sign changes with the parity of K as well as the target index. The magnitude still follows the same bound.

These two infinite bounded inputs are legitimate tests even though their ordinary sums over all indices do not converge. The inverse convolution converges absolutely because its weights are summable and the supplied sequences are bounded. Do not impose a finite-energy assumption halfway through a proof that was stated for bounded sequences; it would unnecessarily exclude the very tests that make the worst-case calculation transparent.

Contents · Previous chapter · Next chapter

On the unit circle, the reciprocal magnitude is 1/|1−2e⁻ⁱω|. The denominator magnitude ranges from 1 at zero frequency to 3 at ω=π. Thus the stable inverse has magnitude one on constant signals and one third on alternating signals, agreeing with the direct sequence calculations.

For general a>1, the largest reciprocal magnitude is 1/(a−1), attained at zero frequency. It agrees with the absolute-weight bound because the stable inverse weights all have the same sign, allowing a constant observation perturbation to align their effects.

This frequency calculation describes the stable inverse whose ROC contains the unit circle. It does not make the causal inverse stable by assigning it the same formal rational values there. The causal impulse series does not converge on the unit circle. A frequency expression detached from its convergence conditions can therefore conceal the very time-direction distinction this example was designed to reveal.

Contents · Previous chapter · Next chapter

Set a=1, so y[n]=x[n]−x[n−1]. Every constant input produces y[n]=0. Two different bounded constant inputs are now indistinguishable. Consequently an inverse recovering every bounded two-sided input from y alone cannot exist.

The candidate causal inverse has unit weights extending indefinitely into the past; the candidate anti-causal inverse has negative unit weights extending indefinitely into the future. Neither is absolutely summable. The stability argument used for a>1 fails at the boundary, and the uniqueness argument fails too because the homogeneous difference can be a nonzero bounded constant.

Additional boundary information can still determine a finite reconstruction: a known initial value allows cumulative summation over the available record. That is a conditional finite problem, not a stable universal inverse on bounded two-sided sequences. Checking the equality case a=1 prevents a strict inequality from being silently weakened. A zero close to the unit circle is sensitive; a zero on it can remove information from the stated signal class altogether.

Contents · Previous chapter · Next chapter

Suppose the true forward rule uses b>1, but the stable inverse assumes a>1. Write Δ=b−a. Since H_b=H_a−Δz⁻¹, the composed operator is G_aH_b=1−ΔG_az⁻¹.

For a bounded true input |x[n]|≤M, the infinite-preview reconstruction error is therefore at most |Δ|M/(a−1). A constant input attains this bound in magnitude: its reconstructed value is [(b−1)/(a−1)]x, rather than x. Increasing the preview length cannot eliminate this coefficient mismatch.

For example, if the assumed a is 2 and the true b is 2.02, a constant true input of 5 is reconstructed as 5.1 by the infinite stable inverse. The discrepancy 0.1 equals |Δ|M/(a−1) for M=5. This teaching calculation separates model error from truncation and measurement noise. A validation plan should not attribute every residual to insufficient future samples when the forward coefficient itself may be uncertain.

Contents · Previous chapter · Next chapter

Return to the target error 0.01 and input bound M=8, now admitting observation error bounded by η=0.005. The combined bound becomes 0.005+7.995·2⁻ᴷ. Meeting the target requires 2ᴷ≥1599, so the smallest integer choice is K=11.

At K=10 the combined bound is 0.0128076171875, exceeding the target even though truncation alone met it. At K=11 the bound is 0.00890380859375. This is a concrete reason to revise an apparently successful delay calculation when the evidence assumptions change.

If the required worst-case error were smaller than η, this particular bound could not certify it for arbitrary bounded input and noise under these values. The next step would be to improve measurement quality, change the criterion or justify additional structure. Simply increasing K would not establish the missing guarantee. If a coefficient mismatch is also admitted, its contribution must be assessed rather than quietly absorbed into the observation-noise allowance.

Contents · Previous chapter · Next chapter

For the foundation, return to the Z-Transform subject guide, especially its sequence, ROC and inversion discussions. For a domain application with uncertain wavelets and physical evidence, follow the seismic deconvolution and ringing guide. Use the two readings for different questions: one explains the mathematical operator; the other examines what a measured signal can support about a site.

For wider systems questions, How Mathematics Works: Control Theory connects dynamic models to implementation and feedback. The Gold Standard of Education: Learning Transfers into Life explains why a changed coefficient and a changed data window are valuable tests of understanding.

The How X Works hub offers further subjects where a valid inverse depends on the information available, the assumptions admitted and the question the reconstructed quantity is meant to answer.

For another worked implementation check, the guide to cancelled transfer poles and hidden internal modes constructs a stable selected output alongside an unstable internal state. It compares the external map with the state realization, complementing this article’s comparison of inverse time directions.

Contents · Previous chapter · Return to Z-Transform