WorksPhysics and Astronomy
Observable Spectral Framework for Noisy Amplitude Amplification: Stopping Stability and Correction
Abstract
Stopping rules for noisy amplitude amplification often rely on channel-specific success-probability formulas, limiting their transfer across noise mechanisms. We develop an observable spectral framework for stationary noisy standard amplitude amplification under explicit leading-mode, spectral-separation, and residual-control conditions. Within this class, the first amplification lobe has a damped-oscillatory leading component and a bounded non-leading contribution. We derive how frequency shift, damping, and observable phase offset displace the leading peak, and establish sampled-gap criteria for retaining the ideal integer stopping depth, certifying recalibration, or selecting a robust candidate window. In single-target Grover search, frequency and damping errors have stopping-coordinate sensitivities proportional to the search-space size, whereas phase offsets scale with its square root; residual modes can independently reverse the ordering of neighbouring depths. Analytical constructions, circuit simulations, finite-shot inference, and calibration-derived device-like trajectories examine these predictions and their operational limits. For a 64-element search, an a posteriori sampled-set check selects depth five over six, with success probability increasing from 0.595 to 0.626. The resulting theory links observable spectral dynamics to conditional integer stopping guarantees without requiring a separate microscopic channel solution.
Introduction
Grover search and amplitude amplification are canonical quantum algorithms whose performance depends on the success-probability trajectory and the number of iterations executed [1, 2, 3]. Under noise, that trajectory is determined jointly by the noisy iteration map, the initial state, and the success measurement. Analytic expressions obtained for particular phase-flip and depolarizing channels [4, 5] generally cannot be transferred unchanged to a different noise process. This creates a modeling problem for noise-aware algorithm execution: one needs more structure than an arbitrary noisy channel supplies, but a useful stopping theory should not depend on a separate closed-form solution for each microscopic noise model.
We address this problem through an intermediate, observable-level description of stationary noisy standard amplitude amplification. The framework identifies a class of trajectories whose first amplification lobe is governed by an observable-relevant, semisimple conjugate eigenvalue pair of the noisy iterate, with non-leading contributions controlled on the same lobe. Spectral separation, overlap, and residual bounds make the leading oscillation and its error allowance explicit. The resulting damped-cosine normal form separates frequency shift, damping, and observable phase offset from residual dynamics. § 4.1 specifies the spectral conditions defining this dynamics class, and the Supplement develops the reduction and sufficient realizability routes. For near-term quantum devices, where a complete microscopic noise model may be unavailable but calibration data can constrain the observed success-probability trajectory, this observable-level formulation provides a route to noise-aware stopping decisions under the stated spectral and residual conditions.
The algorithmic consequence is an integer stopping decision. The continuous first success peak and the implementable sampled maximizer are distinct, and a useful peak may survive after the ideal integer depth ceases to be optimal. Choosing an inappropriate depth can waste oracle calls, accumulate additional noise, or provide a suboptimal amplification power to iterative and QPE-free amplitude estimation [6, 7]. Within the spectral framework, continuous peak motion determines candidate depths, while discrete probability gaps and bounded residuals determine whether their ordering is reliable.
Two scales are therefore relevant. First-peak preservation asks whether a useful amplification lobe survives; same-step stability asks whether the ideal integer depth remains its best sampled choice. Frequency and damping perturbations can shift the stopping coordinate on the search-space-size scale, while an observable phase offset acts on its square-root scale in single-target Grover search. A visible first peak can consequently admit substantially larger perturbations than a fixed stopping decision. The analysis also distinguishes horizontal displacement of the leading peak from direct value-level reordering by non-leading modes. These mechanisms lead to retaining the ideal depth, recalibrating to a certified nearby depth, or returning a robust local window when a unique decision is not resolved.
To assess operational relevance, we examine an amplitude-estimation workflow in which a coarse estimator localizes the candidate depth region before local stopping calibration. We evaluate calibration cost, bootstrap-supported unique outputs, and observed robust-window coverage under two backend-derived noise models.
The main contributions are as follows.
We formulate an observable spectral framework for a class of stationary noisy amplitude-amplification trajectories with a dominant oscillatory eigenvalue pair and controlled non-leading modes. We derive the resulting normal form and residual bounds, and give sufficient perturbative and exact channel realizations of the assumptions.
Within this class, we derive a mechanism-resolved first-peak displacement law for effective-angle shift, damping, and observable phase offset, including its local remainder. Its single-target specialization quantifies the distinct stopping sensitivities and separates first-peak survival from fixed same-step stability.
We connect the leading dynamics to integer decisions using sampled gaps, directional margins, and curvature-scaled residual allowances. The resulting rules retain the ideal stopping depth, certify a recalibrated depth, or return a robust local candidate window as the available information permits.
We characterize residual-induced sampled reversal, in which non-leading observable contributions change the ordering of neighbouring depths without removing the leading first peak.
We validate the framework across analytical and simulated noise settings, quantify finite-shot decision reliability and matched stopping baselines, and assess calibration and reuse costs through an amplitude-estimation-localized application.
The theory assumes a known ideal success fraction and a controlled, spectrally structured first lobe; single-target Grover search provides the main asymptotic specialization. Sections 2–4 position the approach, define its setting, and develop the normal form, stability, correction, and residual-reversal results. Sections 5 and 6 present the validation and operational implications before Section 7 concludes.
Related Work
Noise-dependent success probabilities and stopping depths have been studied under several explicit error models. Azuma developed a higher-order perturbation expansion for Grover search with independently occurring single-qubit phase-flip errors, examining success-probability degradation, thresholds, and an earlier first peak under noise [4]. Cohn et al. studied both total (full-register) and local (per-qubit) depolarizing channels [5]. For total depolarization they obtained an analytic success trajectory and a weak-noise correction of approximately to the optimal iteration count, where is their per-iterate full-register error rate; for local depolarization they developed bounds and approximations. Our encoded-subspace realization in Eq. (15) instead gives the continuous first-peak displacement for weak . The corrections share an scale, but the full-register success probability includes a depth-dependent maximally mixed contribution, unlike the constant background in our encoded-subspace observable; the optimum and the continuous peak are also distinct objects. These examples establish noise-adjusted stopping in specified channel models. Our starting point is instead an observable-relevant oscillatory spectral pair and controlled non-leading contributions, defining a class on which common displacement, sampled-gap, and correction statements can be made. The conditions needed for this reduction are specified in § 4.1.
Further studies treat stochastic error models [8, 9], unitary gate imperfections [10], and locally acting noise with time correlations [11]. Shapira et al. analyze imperfect unitary operations, providing a related context for the coherent circuit-error test in Subsection 5.4, although the perturbed gates and error models differ. The time-correlated setting of Mandel et al. also illustrates a distinct noise regime relative to the fixed per-iterate map adopted here. Experimental realizations include NMR, trapped-ion, and programmable-processor searches [12, 13, 14, 15], with near-term and QSVT-oriented circuit studies providing further implementation contexts [16, 17].
Spectral approaches provide another perspective on noisy quantum dynamics. Reitzner and Hillery used the spectral structure of localized-dephasing Grover evolution to characterize its success-probability dynamics [18], while Gu et al. developed channel-spectrum benchmarking methods based on experimentally accessible repeated-channel signals [19]. These studies establish complementary uses of spectral information in noisy search dynamics and quantum-device characterization.
Our related response-threshold study [20] develops a model-specific predictive theory on the standard Grover iteration grid. For a Drude–Lorentz environment, linearized HEOM provides a priori timing and sampled-probability responses, while finite-coupling HEOM, a same-bath NZ2 comparator, and circuit-level validation test the threshold predictions. Here the dynamical input is instead a leading observable spectral pair and a controlled residual over the first lobe. The resulting statements concern mechanism-resolved displacement, conditional sampled stability, recalibration, robust windows, and residual-induced reversal across that defined class. Both approaches use the same underlying integer-sampling geometry but obtain the relevant dynamical information differently.
A second line of work changes the amplification protocol itself. Arbitrary-phase and phase-matching methods analyze reflection-phase choices [21, 22, 23, 24]; Leng, Yang, and Wang proposed a phase-modified D2p procedure under phase noise [25]; and fixed-point amplification avoids overshooting through modified sequences [26, 27, 28]. Our analysis keeps the standard amplification sequence fixed and studies the integer depth decision induced by its noisy observable trajectory.
Noise-aware depth selection provides a complementary operational perspective. Leng, Yang, and Wang use predicted success probabilities to reduce restart-aware noisy-search cost [29]. Ganguly and Ahn use depolarizing-noise information to choose hardware-aware amplification depths [30]. The present work targets the sampled maximizer within a surviving first lobe and specifies when this decision is stable or certifiably corrected under its spectral and residual assumptions. The objectives and information required by these approaches are therefore different, even where they lead to similar depth choices.
Robust and QPE-free amplitude estimation develops likelihoods, measurement schedules, or noise treatments for amplitude inference [6, 7, 31, 32, 33, 34]. In the application of Subsection 6.1, coarse amplitude estimation localizes the candidate region and the stopping framework makes the subsequent local depth decision. Table 1 summarizes the different dynamical inputs and algorithmic outputs.
| Work | Dynamics / information | Principal result | Analysis scope |
| Azuma [4] | Independent single-qubit phase flips | Perturbative success probabilities, threshold behavior, peak advance | Specified phase-flip process |
| Cohn et al. [5] | Total and local depolarizing channels | Noise-adjusted optimal depth; local-noise bounds and estimates | Channel-dependent probability and complexity analysis |
| Zhong et al. [20] | Drude–Lorentz bath; linearized HEOM | A priori response and threshold screening | Microscopic structured-noise model |
| Leng et al. [29] | Predicted noisy success probabilities | Restart-aware depth selection | Expected search cost |
| Ganguly and Ahn [30] | Gate-level depolarizing information | Hardware-aware depth choice | Noise-informed compilation / execution |
| Ramôa and Santos [34] | Bayesian amplitude and noise inference | Adaptive amplitude estimation | Parameter estimation |
| Present work | Dominant observable spectral pair; bounded residual | Displacement, sampled certification, recalibration, robust window | Conditional first-lobe stopping across a defined dynamics class |
Table 1. Representative noisy-Grover and amplitude-estimation approaches, organized by their dynamical inputs and principal results.
Problem Setting
We specify the ideal integer-sampling geometry and the stationary noisy iterate that generates the observable success trajectory. These definitions provide the inputs for the spectrally structured first-lobe regime developed in § 4.
Ideal Sampled Stopping Geometry
The ideal sampling grid provides the geometric reference for subsequent noise-dependent decisions. Let be the initial success probability and define
The ideal success probability after iterations is
For single-target Grover search, and . Treating as continuous, the first maximum occurs at
The implementable ideal stopping step is therefore
where nearest-integer rounding uses a fixed tie-breaking convention.
Write , with . The distance from the continuous peak to the nearest boundary of its integer rounding cell is
The standard rounding-cell geometry is shared with the response-threshold framework in Ref. [20]. We record the corresponding observation here only to fix notation: if a perturbed continuous first peak is and
then . This rounding-cell geometry fixes the continuous-peak notation used below; sampled stopping stability is treated separately in § 4.6.
The exact right and left margins are
with . Thus a positive displacement is allowed up to , whereas a negative displacement is allowed up to . These directional margins give the exact continuous rounding boundaries.
Although in single-target Grover search, is determined by the fractional part of and is therefore not a smooth asymptotic function of . Search sizes with the same asymptotic scale can consequently have different fixed-step robustness.
Running example: . To keep the geometric, spectral, and correction quantities concrete, we use one worked example throughout the paper. For single-target Grover search with ,
Hence , the early (left) directional margin is , the late (right) margin is , and the early rounding-cell boundary is 5.5. § 5.5 returns to this same instance under backend-derived device-like noise, so that the displacement, sampled-step change, correction certificate, and calibration cost can be followed without changing examples.
Noisy Iterate Model
Let denote the ideal standard amplitude-amplification iterate and its channel representation. For an initial state and good-subspace measurement , the ideal trajectory is from Eq. (1). We model stationary per-iterate noise by
where is CPTP in the main physical setting. After noisy iterations,
Time homogeneity permits spectral analysis of repeated noisy iterates. The leading-mode and residual conditions in Subsection 4.1 then specify the subclass of observable first-lobe trajectories treated by the stopping theory.
The principal symbols are introduced locally; a consolidated notation table is provided in Supplementary tbl:S1.
Formulation
We first identify the observable spectral conditions under which stationary noisy iterations admit a controlled first-lobe normal form. We then derive the peak-displacement, sampled-stability, and correction results that follow from that structure. Figure 1 summarizes the resulting operational hierarchy.

Figure 1. Diagnostic hierarchy for noisy sampled stopping. A measured or calibrated trajectory is first checked for a reliable first-lobe normal-form regime. If the fixed same-step certificate holds, the ideal sampled depth is retained. Otherwise, the fitted leading lobe is used to test whether a unique recalibrated sampled depth can be certified; if not, the displacement uncertainty is converted into a robust local candidate window. The same hierarchy can be evaluated using point estimates or finite-shot bootstrap regions.
Observable First-Peak-Preserving Regime
Observable-relevant leading oscillatory pair
The spectral class is defined by requiring the measured trajectory on the first-peak window to be dominated by an observable-relevant conjugate pair
Observable relevance means that the pair has non-negligible overlap with the chosen initial state and success measurement. Exactly stationary observable modes (eigenvalue 1) contribute to the constant background; slowly varying modes that are not stationary remain in the residual and are subject to its bound.
Since
we parameterize the noisy frequency and observable phase as
Thus , , and represent the effective-angle shift, damping scale, and observable phase offset. They are fitted normal-form parameters rather than independent algorithmic controls.
The relevant first-peak window is
with independent of . In single-target Grover search, this is an -iteration window. Its integer sampling set is . We choose the window within a single leading first lobe and require it to contain the ideal step when fixed-step stability is tested and both neighbouring integer candidates of the leading peak when recalibration is used. The relevant sampled gap is defined on a finite set with at least two candidate depths. The leading pair is assumed semisimple. Controlled non-leading Jordan contributions may be absorbed into the residual bound below, whereas a Jordan block in the leading pair would produce a polynomially modulated leading oscillation outside the present damped-cosine theory.
Spectral separation and residual control
Let index the observable-relevant residual modes after removing the leading conjugate pair and the exactly stationary modes absorbed into . Define
with if . Nonstationary slowly varying modes belong to ; they are not part of the constant background. We assume
Writing the non-leading observable contribution as , define
A spectral decomposition gives the bound
where is determined by the beginning of , and collects spectral-projector conditioning and observable overlaps.
The admissible regime requires not to overwhelm the relative spectral decay; for example, uniformly conditioned blocks give , while polynomial growth is also compatible with a sufficiently persistent gap. Nearly defective or strongly non-normal dynamics may violate this control and then require model-specific resolvent or dynamical estimates. The residual bound and its conditioning assumptions are derived in Supplementary Sections S1.4–S1.6, while the perturbative realizability route is developed separately in Supplementary Section S1.8.
The leading pair and residual have distinct roles: the pair controls horizontal first-peak displacement, whereas controls value-level changes in the ordering of nearby sampled steps. The sampled-gap certificate therefore compares this residual with the local discrete probability gap, which is of order in the clean single-target Grover regime.
Perturbative realizability. The main results assume the observable leading-pair normal form, spectral separation, and residual control directly. A standard sufficient route from a small perturbation of the ideal iterate, based on Bauer–Fike and observable overlap control, is given in Supplementary Section S1.8. This provides a realizability certificate for the spectral framework; the stopping results below use the stated leading-mode and residual conditions.
Closed-form depolarizing realization. A useful exact realization of the assumptions is encoded-subspace depolarization on the marked–unmarked Grover plane. It gives , , and . The channel spectrum, observable-overlap calculation, and the distinction between the continuous-cell crossing and exact sampled equality are derived in Supplementary Section S1.9.
Spectral Normal Form and Peak Displacement
Observable-level cosine normal form
The spectral conditions of Subsection 4.1 translate the repeated noisy iterate into a description of the measured values . The leading observable-relevant conjugate pair produces a damped cosine. Stationary observable modes form the constant background; the other observable non-leading modes contribute the residual:
Here , and , , , , and are observable-level normal-form parameters. We write
with
The sign convention places the ideal first peak near a cosine minimum. The spectral-to-observable reduction is derived explicitly in Supplementary Sections S1.1–S1.3, where the leading pair, background contribution, and non-leading residual modes are separated.
The ideal curve is recovered from
which reduces Eq. (12) to Eq. (1). Using
the normal form separates the effective-angle shift , damping , and observable phase offset .
We assume that the leading first peak remains non-degenerate:
for some . This condition requires a visible oscillatory component with non-vanishing local curvature; the background may shift the curve vertically without moving the leading peak. A sufficient perturbative route for the persistence of the leading observable overlap is given in Supplementary Section S1.8.
Leading first-peak equation
Ignoring , the stationary branch of the leading damped cosine near the ideal first peak gives
The full differentiation and stationary-branch calculation are given in Supplementary Section S1.7. At this peak the local curvature is
When ,
Thus, under single-target Grover scaling and non-collapsing peak height, . Setting , , and in Eq. (13) recovers .
Displacement contributions from , , and
Substituting and into Eq. (13), and defining , gives the following expansion in the local regime
The remainder obeys
with an -independent local constant. Supplementary Sections S2.1–S2.3 give the exact displacement identity, the first-order expansion, and the explicit remainder envelope.
The three first-order terms have distinct origins:
The first term accumulates a frequency shift over the Grover-scale trajectory, the second translates the observable phase, and the third arises from the tilt of the damped envelope. The damping coefficient follows from the probability-level convention . Accordingly, and are parameters of the observable probability trajectory, not eigenangle parameters of the statevector Grover rotation.
For single-target Grover search, and , so
These are stopping-coordinate sensitivities, not peak-height scalings. A visible first peak may therefore cross an integer rounding boundary under perturbations that remain small at the one-step level.
Running example. At the shifted GenericBackendV2 operating point used later in Subsection 5.5, the following parameters are obtained from a 13-depth sampled-trajectory fit () reported in Subsection 5.5:
The three first-order contributions are , , and , giving . The fitted leading peak lies at , so the exact displacement is and the displacement remainder is . Thus the observable phase offset dominates the early shift, damping reinforces it, and the negative frequency shift partially compensates it. For this example the dimensionless perturbation size is approximately 0.523, so the local remainder estimate applies with, for example, . The fitted peak location above uses the exact leading-peak equation; the separated first-order terms provide a mechanism decomposition rather than an independent certificate of the stopping step.
As a concrete specialization, the encoded-subspace depolarizing realization has
Substituting these parameters into Eq. (13) gives the exact continuous first-peak displacement
For weak noise,
which reproduces the damping contribution in Eq. (14). The corresponding finite- sampled transition is determined by comparing neighbouring sampled probabilities, as derived in Supplementary Section S1.9.
Combined displacement condition
Define
Before accounting for the value-level residual , the rounding-cell margin condition gives
Under Eq. (16), . The signed sum is essential: different mechanisms may reinforce or partially cancel one another, so separate componentwise bounds do not characterize the rounding-cell boundary.
Using the directional margins of § 3.1, the exact directional condition for continuous-cell membership of the leading peak is
Equation (17) is the exact geometric condition for the leading peak to remain in the same nearest-integer cell, while Eq. (16) is its uniform sufficient version. Damping shifts the peak earlier and is therefore controlled by the left margin , whereas frequency and phase shifts may act in either direction. The sampled-value certificate is introduced once in § 4.6.
Fixed Sampled Stopping Stability
Sampled-gap certificate
A continuous rounding-cell condition locates the leading peak, whereas an integer stopping decision depends on the sampled probability ordering. For the observable normal form, this ordering is determined by the leading curve together with the non-leading residual. We express the sampled-gap condition on a general candidate set so that it can support retaining the ideal step, certifying a recalibrated step, and constructing a robust window.
Let be a finite candidate set and let be the unique sampled maximizer of the leading curve on . Define
Lemma 4.1 (Sampled-gap preservation). Assume for , with . If
then is also the unique sampled maximizer of on .
The gap-versus-residual criterion applies to any candidate set satisfying the lemma’s assumptions and supplies the common sampled-ordering condition for the stopping decisions below.
For fixed same-step stability, take and . Define
Corollary 4.2 (Fixed same-step stopping). If
then the noisy sampled first-peak maximizer is unchanged:
Curvature-scaled residual allowance
To make Eq. (21) operational, write
and let
When the leading peak remains in the ideal cell, . We impose the local sampled-gap regularity condition
For a non-degenerate damped-cosine lobe away from a sampled equality, a Taylor expansion with controlled local skew supplies this curvature-to-gap bound; its limitations near ties, asymmetric peaks, and flattened lobes are given in Supplementary Section S2.6. For the ideal single-target curve, and , so can be chosen close to 1 away from ties.
Combining Eq. (22) with Corollary 4.2 gives
Using the perturbative displacement formula yields
The first two terms consume margin through horizontal displacement, whereas the last term is the margin reserved for residual value perturbations. In the clean single-target regime, ; asymptotic same-step stability with non-vanishing remaining margin therefore requires
Directional and discrete boundaries. The symmetric margin can be conservative when the ideal peak is not centred in its rounding cell. After reserving for value-level uncertainty, a direction-aware sufficient condition is
The sign-resolved derivation, damping-only specialization, and the exact comparison of continuous and sampled boundaries are given in Supplementary Sections S2.8 and S1.9.
Speedup–Stability Duality
Stopping-displacement sensitivity
Let
so that ; in single-target Grover search, . From Eq. (14), the first-order displacement sensitivities are
Thus frequency and damping perturbations act at the square of the Grover stopping scale, while an observable phase offset acts at the linear scale. For single-target Grover search these sensitivities are , , and , respectively. The derivation of these sensitivities and the corresponding comparison between peak-survival and no-correction scales is given in Supplementary Subsection S2.7. Combining them with the directional continuous rounding margin gives the characteristic continuous-cell no-correction scales
Together with the sampled-gap regularity condition of § 4.6, these scales provide a sufficient route into the no-correction regime. The factor has a simple origin: a frequency or damping error accumulates over iterations and is then converted back from phase to iteration displacement by another factor . A direct observable phase offset requires only the second conversion.
Peak survival and the speedup–stability gap
Same-step stability is stronger than preservation of a useful first peak. In the observable normal form, the leading peak remains visible when
for some , and it preserves the Grover stopping scale when . A recalibrated rule then still uses iterations, which is in the single-target case.
For damping, peak-height survival only requires
whereas the continuous-cell no-correction scale is
Hence the intermediate regime
can retain a visible Grover-scale first peak after the continuous no-correction condition has ceased to apply. Frequency and observable-phase perturbations can produce the same separation through horizontal displacement. Under the non-collapsing peak and Grover-scale location assumptions just stated, retaining the ideal sampled step preserves the first-peak query scale; preserving a visible first peak at that scale does not imply same-step stability. An peak displacement can be negligible relative to the Grover stopping scale while still changing the integer decision, and residual modes provide a second route by directly reordering neighbouring sampled values. This separation is the speedup–stability duality. Correlations among , , , remain model-dependent.
Stopping-Rule Correction Theory
Recalibrated stopping
When the ideal step is not certified but the noisy first peak survives, the natural correction is to follow the sampled maximizer of the fitted leading first lobe. For a single-lobed , define the local candidate set (with both adjacent integers assumed to lie in )
and define
A fixed tie-breaking convention is used if the two neighbouring sampled values are equal. The displacement formula locates the candidate pair, while the sampled comparison selects the corrected integer. For a locally symmetric peak this reduces to nearest-integer rounding, but that reduction is not assumed for a general damped trajectory.
Define the leading recalibrated gap by
Corollary 4.3 (Recalibrated first-peak stopping). Assume , and that is the unique sampled maximizer of on . If
then
This is Lemma 4.1 with and . Recalibrated stopping does not require the noisy peak to remain in the ideal rounding cell. It only requires the peak to remain identifiable and the residual to be smaller than its local sampled gap. Since in the first-peak-preserving regime, ; thus recalibration preserves the Grover query scale even when . Finite-shot inference is specified in Subsection 5.4 and Supplementary Subsection S4.12; calibration design and amortized cost are quantified in Subsection 6.2 and Supplementary Subsection S4.15.
Running example. For , the displaced leading peak gives the candidate pair and selects . The numerical certificate and finite-shot fallback for this same instance are reported in Subsection 5.5.
Robust-window stopping
When the exact noisy peak is uncertain, suppose instead that
where one admissible bound is
Define the extremal sampled candidates compatible with this displacement bound by
A symmetric robust window is
with
This construction contains both integer neighbours of every admissible continuous peak location and therefore contains . The corresponding containment proof is given in Supplementary Section S3.1. Let
Corollary 4.4 (Robust-window first-peak stopping). Assume , and that is the unique sampled maximizer of in the window. Define
If
then
This is Lemma 4.1 restricted to the robust candidate set .
The window construction guarantees containment of the leading-curve floor–ceiling candidates, including , not automatically the true noisy maximizer outside the window. If , only a constant number of candidate depths is added; more generally, any is subleading relative to the amplification cost.
If sign information or asymmetric uncertainty bounds are available, the robust window can be tightened separately on its early and late sides, provided it contains every admissible floor–ceiling candidate pair and satisfies the same sampled-gap certificate. The complete direction-aware construction is given in Supplementary Section S3.2.
Residual-Induced Sampled Reversal
Value-to-step conversion
Let be the sampled maximizer of the leading curve , and let be a competitor. Define
Then
Hence only residual differentials matter for sampled stopping: a common vertical shift is irrelevant, whereas a competitor overtakes the leading step once its residual advantage exceeds the leading sampled gap.
Sampled reversal condition
For a competing step , sampled reversal occurs exactly when
In the fixed same-step setting, . The nearest competitor is typically one of , and local quadraticity gives
Using recovers the curvature-scaled allowance in Eq. (23).
The uniform condition rules out every reversal, but its failure only removes the certificate; an actual reversal additionally requires the destabilizing sign in Eq. (31). For the ideal single-target curve,
so the nearest-competing gap satisfies
Therefore the critical direct residual differential is
If a symmetric adversarial perturbation lowers and raises the competitor by the same amplitude , the direct differential is , giving
The direct residual construction, its spectral parametrization, and the large- validation protocol are given in Supplementary Sections S4.2–S4.4.
Spectral interpretation
The value-level reversal condition can be connected to spectral separation because non-leading observable modes decay relative to the leading pair at a rate governed by . A positive spectral gap alone is not sufficient: over the first-peak window, the surviving residual differential must be smaller than the local sampled gap, which is in clean single-target Grover scaling. Consequently, a slowly decaying non-leading mode can change the sampled maximizer even while the leading first peak remains visible.
The explicit residual-mode parametrization, critical spectral coefficient, and the resulting sufficient spectral-gap condition are given in Supplementary Section S4.3. The main-text point is therefore operational: displacement control locates the leading lobe, while residual control is separately required to preserve its sampled-value ordering.
Numerical Validation
The numerical study examines the predictions of the observable normal form and the sampled decisions supported by its residual conditions. Five tests address mechanism-resolved displacement, the separation of peak survival and same-step stability, residual-induced reordering, parameter inference from circuit and finite-shot data, and calibration-derived device-like trajectories with matched baselines. Normal-form and residual constructions directly probe the analytical conditions; circuit and backend-derived studies assess their use through fitted observables and sampled residual checks. Detailed protocols and secondary scans are given in Supplementary Sections S4 and S5.
Mechanism and margin checks
We test the mechanism decomposition within the observable normal form: whether the three perturbation parameters produce the displacement scales predicted by Eq. (14), and whether their joint effect follows the signed displacement coordinate. For single-target Grover search, the single-mechanism scans use , , and , one at a time. A separate scan varies jointly, and a dense search-size scan with tests the effect of the directional rounding margin. The full parameter choices are given in Supplementary Subsections S4.5–S4.7.
Figure 2 shows that the normalized exact displacements approach the predicted constants for all three mechanisms. The joint scan collapses onto a diagonal band of nearly constant signed displacement, while the dense -scan shows that the same asymptotic perturbation law can either preserve or change the sampled step depending on the arithmetic position of the ideal peak inside its rounding cell. The joint and margin-sensitive scans are shown in Supplementary Figures S3 and S4.
Figure 2. Single-mechanism large- scaling. The normalized displacements for , , and approach the predicted limits , , and , respectively.
Speedup–stability separation
We test the distinct noise scales predicted for a surviving leading first peak and for preservation of the ideal integer sampling decision. We use damping-only trajectories over increasing search size and compare the first-peak survival scale with the no-correction scale predicted by § 4.7 analysis. The sampled stopping step is evaluated directly from the noisy trajectory.
Figure 3 shows the two scales separating with : a non-collapsed first peak persists on an -type damping scale, whereas the no-correction scale follows the sharper margin-modulated -type behavior. The two asymptotic envelopes differ, but their finite- separation is modulated by the arithmetic margin , so the plotted no-correction threshold is non-monotone and does not define a single crossover search size.
Figure 3. Peak-height survival versus the geometric no-correction scale under damping. A constant fractional peak-height loss occurs at the scale, whereas the rounding-margin threshold follows the sharper margin-modulated -type scale. Sampled decisions are evaluated directly as specified in § 4.6.
Residual-induced sampled reversal
This experiment isolates the value-level failure mode that is not captured by horizontal peak displacement: a non-leading residual can reverse the ordering of neighbouring sampled depths while the leading peak remains geometrically well located. At , the leading curve is held fixed and the residual amplitude is varied through the local sampled-gap threshold. A second scan varies the residual spectral ratio to control how strongly the residual survives to the first-peak window.
Figure 4 resolves the predicted transition: the sampled step changes when the residual differential exceeds the leading local gap, and modes with closer to one require a smaller initial coefficient to produce the reversal. A larger- supporting scan is reported in Supplementary Figure S2.
Figure 4. Residual-induced reversal at . The sampled step changes when the residual differential exceeds the leading local gap; the spectral ratio controls how strongly the residual survives to the first peak.
Circuit and finite-shot robustness
We test whether the same stopping quantities remain observable in an explicit Grover circuit and whether the retain–recalibrate–window hierarchy survives finite measurement noise. For , , the marked-state reflection is given a coherent phase over-rotation , while the diffuser remains ideal. The resulting statevector trajectories are fitted to the observable normal form. We then add binomial sampling noise at shots per sampled depth and use parametric-bootstrap refits. Complete fitting and bootstrap details are given in Supplementary Subsections S4.10 and S4.12.
The exact circuit trajectories fit with root-mean-square residual below . The observed sampled transition occurs near , close to the first-order prediction near 0.150. In the finite-shot bootstrap decisions, supports retaining the ideal depth 6, returns the robust window , and supports the recalibrated depth 5. The full circuit and finite-shot plots are reported in Supplementary fig:S5 and S6; bootstrap percentile intervals are shown in Supplementary fig:S8.
Backend-derived device-like validation and matched baselines
The final main-text experiment assesses observable normal-form fitting and sampled-depth certification under combined, hardware-like error mechanisms in a GenericBackendV2 simulation. A calibration-derived FakeOslo study in the Supplement provides a second backend-derived test with an archived calibration snapshot. Six-qubit Grover circuits are transpiled to a GenericBackendV2 basis and coupling graph with backend-derived gate errors, relaxation, and gate durations. The simulated success probabilities are extracted from premeasurement density matrices; readout-error metadata is available but readout noise is not applied to these probabilities. At the ideal reference depth , the transpiled circuit has depth 2866 and 1915 CX gates. Because the unscaled backend-derived model destroys the useful first lobe, we use the controlled strength interpolation to resolve the first-peak-preserving regime. The backend construction, strength scan, normal-form fit, matched baselines, and calibration accounting are specified in Supplementary Subsections S4.13–S4.15.
Figure 5 summarizes the running example. At , the sampled and recalibrated depths remain 6. At , both are 5, with the neighbouring sampled values exchanging order at . The shifted normal-form fit has , and
Figure 5. GenericBackendV2 device-like validation for , . Panel (a) compares the ideal first lobe with reference and shifted noise strengths. Panel (b) resolves the sampled-step transition through the sign change of . Panel (c) decomposes the shifted leading-peak displacement into frequency, phase, and damping contributions and compares their sum with the fitted displacement. Panel (d) compares the D13 fitted-discrepancy scale (labeled in the panel) with . This provides an a posteriori sampled-set check rather than a bound on the theoretical spectral residual.
For the sampled diagnostics, let denote the 13-depth evaluation set and let denote the sparse eight-depth calibration design. Both are distinct from the theoretical first-peak window . For a normal form fitted on , define the fitted discrepancy
This maximum bounds the fitted discrepancy at every depth in and yields an a posteriori sampled-ordering check on that finite set. It is not by itself a bound on the theoretical spectral residual outside the sampled set, nor a proof of the leading-pair assumptions. The leading recalibrated gap is evaluated on this same numerical set , using the sampled-gap minimization of Section 4.5 with as its competitor domain. For the D13 fit it is and , so depth 5 passes the a posteriori ordering check on the tested sampled depths; the robust fallback is . Supplementary Subsection S4.16 repeats the certificate using a normal form fitted only on D8, with the remaining five depths held out. The recalibration gap is still evaluated over using that D8-fitted normal form. For GenericBackendV2, the D8 fit gives and a positive but narrow all-depth margin of about ; the calibration-derived FakeOslo case has an all-depth margin of . The difference between the two displayed values comes from the D13 and D8 fitting designs; both gaps use the same D13 candidate domain.
Matched quantitative baselines. Table 2 compares the same trajectories under three depth-selection rules. The “ideal reference” is the known-fraction integer Grover depth , not the randomized Boyer–Brassard–Høyer–Tapp (BBHT) schedule for an unknown marked fraction [2]. The present framework checks the sampled first peak using the simulated trajectory-derived discrepancy, whereas the Leng et al. rule uses a success-probability predictor with the restart-aware objective ; the in hindsight minimum of that objective is shown only as a diagnostic.
| Regime | Rule | Selected | Peak retention | Cost regret | ||
| Reference () | Ideal Grover step (known fraction) | 6 | 0.913853 | 1.000000 | 6.565609 | 26.25% |
| Reference () | Present framework | 6 | 0.913853 | 1.000000 | 6.565609 | 26.25% |
| Reference () | Leng et al. | 4 | 0.769176 | 0.841684 | 5.200373 | 0.00% |
| Shifted () | Ideal Grover step (known fraction) | 6 | 0.594629 | 0.949841 | 10.090323 | 54.56% |
| Shifted () | Present framework | 5 | 0.626030 | 1.000000 | 7.986832 | 22.34% |
| Shifted () | Leng et al. | 4 | 0.572021 | 0.913728 | 6.992745 | 7.11% |
Table 2. Matched quantitative comparison on the backend-derived device-like trajectories. Peak retention is normalized by the actual noisy first-peak probability in each regime; cost regret is relative to the in-hindsight minimum of over the matched domain .
In the reference regime, the present framework retains the first-peak depth 6, whereas the Leng rule selects depth 4. In the shifted regime, the present framework restores the noisy first-peak depth 5, while Leng selects 4 and the in-hindsight optimum is 3. The first-peak and restart-aware objectives yield different choices: the spectral stopping framework selects an a posteriori checked first-peak depth with a robust fallback, whereas the restart-aware rule prioritizes lower expected execution cost.
A calibration-derived FakeOslo snapshot (17 July 2022) gives the same qualitative 6 → 5 first-peak transition under controlled strength scaling, with . At the frozen shifted point , the fit has , , and the a posteriori finite-set residual-gap check remains positive when the normal form is fitted only on the D8 depths and tested on the five excluded depths . The complete protocol and stress-test figure are reported in Supplementary Subsection S4.16 and Figure S10.
Practical Implications and Limitations
Operational use and amplitude-estimation connection
The observable spectral framework yields the retain–recalibrate–window hierarchy in Figure 1. Under finite shots, the application uses joint bootstrap support for fitted stopping decisions. This empirical support is not, by itself, a rigorous confidence bound on the unobserved structural residual ; theorems requiring a valid residual allowance retain that condition separately.
Amplitude estimation provides a natural application because it must choose amplification powers even when the underlying success amplitude is initially unknown. In the tested workflow, a coarse QPE-free amplitude-estimation stage maximizes the ideal binomial success likelihood over a small pre-specified set of amplification depths to infer a provisional ideal reference depth. This is a likelihood-based QPE-free estimator of the type discussed in [7]. The estimator has no access to the true success fraction, ideal Grover angle, ideal stopping depth, noisy optimum, or fitted normal-form parameters. The provisional reference defines a seven-depth guard band, within which the stopping framework fits the noisy first lobe and returns either a unique recalibrated depth or a robust local window. The complete application protocol is provided in Supplementary Subsection S4.17.
For the running example, offline screening against the known GenericBackendV2 simulated optimum selects a localization configuration, using the fixed depth schedule with 512 shots per depth. Under the shifted GenericBackendV2 trajectory, this coarse estimator returns provisional depths 7 or 8 rather than the true noisy optimum 5, but every resulting guard band contains both the ideal reference depth 6 and the noisy optimum. The local stopping stage then uses 8192 shots per depth and 200 bootstrap refits per end-to-end dataset. Across 100 independent datasets, the bootstrap modal depth is 5 in every case: 60 datasets support a unique bootstrap-selected depth 5, while the remaining 40 return robust windows. There are no incorrect decisive outputs, no fit failures, and every fallback window contains the true noisy optimum. The coarse estimator localizes the candidate region, while the fitted stopping rule resolves the subsequent sampled-depth decision.
The same configuration is transferred without retuning to the calibration-derived FakeOslo trajectory, providing the independent-backend transfer test after the GenericBackendV2 design screening. It again attains complete observed relevant-window coverage, with 78 unique correct recalibrations and 22 safe fallbacks out of 100 datasets. Table 3 summarizes the resulting cost– reliability tradeoff. Under conservative accounting in which the amplitude-estimation and local-correction measurements are independent, the GenericBackendV2 and FakeOslo configurations cost about 72% and 71% less, respectively, than the shifted D8 reference configuration. The D8 reference uses 32768 shots per depth versus 8192 for local calibration, with higher unique-depth return rates; the observed cost difference is primarily a shot-budget comparison, not a saving isolated to AE localization. Reusing overlapping observations gives only a small additional credit.
| Route | Correct unique | Wrong decisive | Fallback coverage | Mean | Mean shots | reduction vs D8 |
| Shifted D8 (GenericBackendV2) | 0.99 | 0.00 | 1.00 | 1,474,560 | 262,144 | – |
| AE-localized GenericBackendV2 | 0.60 | 0.00 | 1.00 | 407,286 | 59,392 | 72.38% |
| AE-localized FakeOslo | 0.78 | 0.00 | 1.00 | 421,622 | 59,392 | 71.41% |
Table 3. AE-localized stopping application versus the shifted D8 calibration baseline. “Correct unique” is the observed unconditional rate of returning the true single depth; “wrong decisive” counts incorrect bootstrap-supported single-depth outputs; fallback coverage is conditional on returning a window. The D8 reference uses 32768 shots per depth, versus 8192 for local calibration in the AE-localized routes; these are not equal-shot or equal-reliability comparisons. All rates are empirical over 100 datasets per route.
Calibration cost and when it pays
The correction rules require the observable normal form to be calibrated or otherwise inferred. Let be the shots at sampled depth , and define the Grover-iterate-equivalent calibration cost
For each reported design/regime, the reliability study uses 100 independent calibration datasets with 200 bootstrap refits per dataset. The six-depth stress design is cheaper but visibly distorts the shifted fit: it gives and an RMSE of 0.019880 evaluated over the 13-depth set , compared with 5.169850 and 0.007342 for the D13 reference fit. The recommended D8 design defined in Subsection 5.5 gives and RMSE 0.007595, nearly matching D13 at substantially lower cost.
Across the 300 tested datasets there are no incorrect decisive single-depth outputs, every fallback window contains the true sampled optimum, and the fit-failure rate is zero. These are empirical frequencies, not zero-error guarantees.
For the shifted trajectory, keeping gives , whereas depth , selected by an a posteriori sampled-set check, gives 7.986832, a 20.85% reduction. Against the recommended D8 cost , that saving is recovered only after about 701,007 repeated uses under approximately stationary noise. The D8 calibration cost thus requires amortized reuse of an approximately stationary setting and is not recovered for a one-off workload. The AE-localized application in Subsection 6.1 addresses this cost bottleneck differently: a cheap coarse localization stage first restricts the relevant depth region, after which only that local window is calibrated. Its measured end-to-end costs and reliability are reported in Table 3. The tested lower-shot AE-localized configuration has lower calibration cost than the higher-shot D8 reference, without isolating a localization-specific saving. Dividing its mean query cost by the same per-use difference gives approximately repeated uses for GenericBackendV2. Since 40% of its datasets return a window, this is only an illustrative normalization, not an observed break-even point or an equal-reliability comparison.
| Regime/design | Decisive | Correct | Wrong decisive | Fallback coverage | ||
| Reference D8 | 2048 | 92160 | 0.98 | 0.98 | 0.00 | 1.00 |
| Shifted D8 | 32768 | 1474560 | 0.99 | 0.99 | 0.00 | 1.00 |
| Shifted D13 | 32768 | 2555904 | 0.99 | 0.99 | 0.00 | 1.00 |
Table 4. Finite-shot calibration reliability for the running example. Rates are measured over 100 independent calibration datasets per case. “Correct” is the unconditional observed rate of returning the true single depth; “wrong decisive” counts incorrect bootstrap-supported single-depth outputs; fallback coverage is conditional on returning a robust window. Rates are empirical, not statistical coverage guarantees.
Scope and limitations
The framework applies to stationary noisy standard amplitude amplification with a known ideal success fraction and a fixed reference step, subject to an observable-relevant semisimple leading conjugate pair, spectral separation, and controlled non-leading residuals on the first lobe. The asymptotic scaling laws and encoded-subspace channel specialization are developed for single-target Grover search. Phase-modified, fixed-point, and QSP/QSVT-based amplification schemes have different trajectory geometries; nearly defective or strongly non-normal dynamics may also fall outside the controlled damped-cosine class.
The GenericBackendV2 and FakeOslo studies use strength-scaled simulations and an archived calibration snapshot, not live hardware executions. Independently transpiled circuits at each depth need not implement powers of one stationary per-iterate map; their normal-form fits are effective trajectory diagnostics. Fits and held-out sampled residual checks do not establish the spectral assumptions of § 4.1 for those device-like trajectories. The unscaled backend-derived noise also suppresses the useful first lobe. The calibration-cost analysis assumes a noise setting stable enough to amortize calibration over repeated uses. The AE-localized application supplies a provisional depth region from finite-shot data; it does not alter the known-reference assumptions of the stopping theorems. A deployment-level statistical certificate would additionally require a valid residual allowance, including measurement and model-mismatch uncertainty; that was not established in these simulations. The sampled first-peak objective is also distinct from restart-aware runtime minimization and from the end-to-end objective of an amplitude-estimation protocol.
Conclusion
We established an observable spectral framework for stationary noisy standard amplitude amplification with a leading observable-relevant oscillatory pair, spectral separation, and controlled residual contributions. The resulting normal form connects frequency, damping, and phase perturbations to first-peak motion and integer stopping decisions across the defined dynamics class. In single-target Grover search, frequency and damping have stopping-coordinate sensitivities, while observable phase offsets act at scale. Sampled-gap and residual bounds distinguish first-peak survival from same-step stability, identify residual-induced reversals, and support retention, certified recalibration, or a robust window.
The numerical studies test these conclusions using analytic constructions, circuit simulations, finite-shot fits, and strength-scaled backend-derived models. In the example, an a posteriori sampled-set check selects depth 5 over 6, increasing success probability from 0.595 to 0.626. The tested AE-localized configurations cost over 70% less than the higher-shot D8 reference, principally because of their lower calibration shot budget, and have lower unique-depth return rates. Across 100 datasets per backend, a correct unique depth was returned in 60 GenericBackendV2 and 78 FakeOslo runs; the remaining runs returned windows containing the true optimum, with no incorrect decisive outputs observed. The spectral assumptions and simulation-only validation scope are detailed in Subsection 6.3.
Research statements
Availability of data and materials
The numerical data, scripts, and environment specifications supporting the reported results are provided in the accompanying reproducibility package. Supplementary Section S4 documents the parameter settings, fitting procedures, calibration designs, and numerical protocols used to reproduce the figures and tables. No human-subject or proprietary datasets were used.
Use of generative AI tools
The authors used ChatGPT to assist with language polishing, manuscript restructuring, LaTeX editing, consistency checks, selected derivation cross-checks, and preliminary drafting and debugging of numerical code. All mathematical arguments, scripts, simulations, figures, citations, interpretations, and conclusions were reviewed and verified by the authors. No AI-generated numerical data or unverified scientific claims were used. The authors take full responsibility for the content of the work.
Competing interests
The authors declare that they have no competing interests.
Funding
No specific funding was received for this research.
Authors’ contributions
Ge Zhong conceived the research, developed the theoretical framework and mathematical derivations, designed and conducted the numerical experiments, performed the computational analyses, and wrote the original manuscript.
Charith Jayasekara, Nisal De Silva, Muhammad Usman, Muhammad Aamir Cheema, and Chetan Arora provided substantive scientific guidance and critical review, including evaluation of the theoretical arguments, assessment of the numerical validation, and recommendations for additional analyses and verification. They also contributed to the scientific interpretation and revision of the manuscript.
References
- [1]Lov K. Grover. A fast quantum mechanical algorithm for database search. In Proceedings of the Twenty-Eighth Annual ACM Symposium on Theory of Computing, STOC ’96, pages 212–219, New York, NY, USA, 1996. Association for Computing Machinery. doi: 10.1145/237814.237866.DOI
- [2]Michel Boyer, Gilles Brassard, Peter Høyer, and Alain Tapp. Tight bounds on quantum searching. Fortschritte der Physik, 46(4–5):493–505, 1998. doi: 10.1002/(SICI)1521-3978(199806)46:4/5<493::AID-PROP493>3.0.CO;2-P.
- [3]Gilles Brassard, Peter Høyer, Michele Mosca, and Alain Tapp. Quantum amplitude amplification and estimation. In Samuel J. Lomonaco, Jr. and Howard E. Brandt, editors, Quantum Computation and Information, volume 305 of Contemporary Mathematics, pages 53–74. American Mathematical Society, Providence, RI, 2002. doi: 10.1090/conm/305/05215.DOI
- [4]Hiroo Azuma. Higher-order perturbation theory for decoherence in Grover’s algorithm. Physical Review A, 72(4):042305, 2005. doi: 10.1103/PhysRevA.72.042305.DOI
- [5]Ilan Cohn, André L. Fonseca de Oliveira, Efrain Buksman, and Jesús García López de Lacalle. Grover’s search with local and total depolarizing channel errors: Complexity analysis. International Journal of Quantum Information, 14(02):1650009, 2016. doi: 10.1142/S021974991650000X.DOI
- [6]Dmitry Grinko, Julien Gacon, Christa Zoufal, and Stefan Woerner. Iterative quantum amplitude estimation. npj Quantum Information, 7:52, 2021. doi: 10.1038/s41534-021-00379-1.DOI
- [7]Yohichi Suzuki, Shumpei Uno, Rudy Raymond, Tomoki Tanaka, Tamiya Onodera, and Naoki Yamamoto. Amplitude estimation without phase estimation. Quantum Information Processing, 19:75, 2020. doi: 10.1007/s11128-019-2565-2.DOI
- [8]B. Pablo-Norman and M. Ruiz-Altaba. Noise in Grover’s quantum search algorithm. Physical Review A, 61(1):012301, 1999. doi: 10.1103/PhysRevA.61.012301.
- [9]Pedro J. Salas. Noise effect on Grover algorithm. The European Physical Journal D, 46(2):365–373, 2008. doi: 10.1140/epjd/e2007-00295-1.DOI
- [10]D. Shapira, S. Mozes, and O. Biham. Effect of unitary noise on Grover’s quantum search algorithm. Physical Review A, 67(4):042301, 2003. doi: 10.1103/PhysRevA.67.042301.DOI
- [11]Sheikh Parvez Mandal, Ahana Ghoshal, Chirag Srivastava, and Ujjwal Sen. Invariance of success probability in Grover’s quantum search under local noise with memory. Physical Review A, 107(2):022427, 2023. doi: 10.1103/PhysRevA.107.022427.DOI
- [12]Isaac L. Chuang, Neil Gershenfeld, and Mark Kubinec. Experimental implementation of fast quantum searching. Physical Review Letters, 80(15):3408–3411, 1998. doi: 10.1103/PhysRevLett.80.3408.DOI
- [13]K.-A. Brickman, P. C. Haljan, P. J. Lee, M. Acton, L. Deslauriers, and C. Monroe. Implementation of Grover’s quantum search algorithm in a scalable system. Physical Review A, 72(5):050306, 2005. doi: 10.1103/PhysRevA.72.050306.DOI
- [14]Caroline Figgatt, Dmitri Maslov, Kevin A. Landsman, Norbert M. Linke, Shantanu Debnath, and Christopher Monroe. Complete 3-qubit Grover search on a programmable quantum computer. Nature Communications, 8:1918, 2017. doi: 10.1038/s41467-017-01904-7.DOI
- [15]Avik Mitra, Avatar Tulsi, and Anil Kumar. Experimental NMR implementation of a robust quantum search algorithm, 2009. arXiv:0912.4071.arxiv.org/abs/0912.4071
- [16]Muhammad Abdullah Ijaz and Muhammad Faryad. Noise analysis of Grover and phase estimation algorithms implemented as quantum singular value transformations for a small number of noisy qubits. Scientific Reports, 13:20144, 2023. doi: 10.1038/s41598-023-47246-x.DOI
- [17]Kun Zhang, Kwangmin Yu, and Vladimir Korepin. Quantum search on noisy intermediate-scale quantum devices. Europhysics Letters, 140(1):18002, 2022. doi: 10.1209/0295-5075/ac90e6.DOI
- [18]Daniel Reitzner and Mark Hillery. Grover search under localized dephasing. Physical Review A, 99(1):012339, 2019. doi: 10.1103/PhysRevA.99.012339.DOI
- [19]Yanwu Gu, Wei-Feng Zhuang, Xudan Chai, and Dong E. Liu. Benchmarking universal quantum gates via channel spectrum. Nature Communications, 14(1):5880, 2023. doi: 10.1038/s41467-023-41598-8.DOI
- [20]Ge Zhong, Charith Jayasekara, Nisal De Silva, and Chetan Arora. Response-threshold framework for first-peak preservation in noisy Grover search, 2026. Manuscript under review at ACM Transactions on Quantum Computing, manuscript TQC-2026-0105.arxiv.org/abs/2611.20339
- [21]Peter Høyer. Arbitrary phases in quantum amplitude amplification. Physical Review A, 62(5):052304, 2000. doi: 10.1103/PhysRevA.62.052304.DOI
- [22]Gui Lu Long, Yan Song Li, Wei Lin Zhang, and Li Niu. Phase matching in quantum searching. Physics Letters A, 262(1):27–34, 1999. doi: 10.1016/S0375-9601(99)00631-3.DOI
- [23]Gui-Lu Long, Xiao Li, and Yang Sun. Phase matching condition for quantum search with a generalized initial state. Physics Letters A, 294(3–4):143–152, 2002. doi: 10.1016/S0375-9601(02)00055-5.DOI
- [24]F. M. Toyama, W. van Dijk, Y. Nogami, M. Tabuchi, and Y. Kimura. Multiphase matching in the Grover algorithm. Physical Review A, 77(4):042324, 2008. doi: 10.1103/PhysRevA.77.042324.DOI
- [25]Jian Leng, Fan Yang, and Xiang-Bin Wang. Improving D2p Grover’s algorithm to reach performance upper bound under phase noise. Physical Review Research, 5(2):023202, 2023. doi: 10.1103/PhysRevResearch.5.023202.DOI
- [26]Theodore J. Yoder, Guang Hao Low, and Isaac L. Chuang. Fixed-point quantum search with an optimal number of queries. Physical Review Letters, 113(21):210501, 2014. doi: 10.1103/PhysRevLett.113.210501.DOI
- [27]Alexander M. Dalzell, Theodore J. Yoder, and Isaac L. Chuang. Fixed-point adiabatic quantum search. Physical Review A, 95(1):012311, 2017. doi: 10.1103/PhysRevA.95.012311.DOI
- [28]Bao Yan, Shijie Wei, Haocong Jiang, Hong Wang, Qianheng Duan, Zhi Ma, and Gui-Lu Long. Fixed-point oblivious quantum amplitude-amplification algorithm. Scientific Reports, 12:14339, 2022. doi: 10.1038/s41598-022-15093-x.DOI
- [29]Jian Leng, Fan Yang, and Xiang-Bin Wang. Noise-tolerant Grover’s algorithm via success-probability prediction. Physical Review Research, 7(1):L012017, 2025. doi: 10.1103/PhysRevResearch.7.L012017.DOI
- [30]Debashis Ganguly and Wonsun Ahn. A depolarizing noise-aware transpiler for optimal amplitude amplification, 2022. arXiv:2210.14335.arxiv.org/abs/2210.14335
- [31]Tomoki Tanaka, Shumpei Uno, Tamiya Onodera, Naoki Yamamoto, and Yôhichi Suzuki. Noisy quantum amplitude estimation without noise estimation. Physical Review A, 105(1):012411, 2022. doi: 10.1103/PhysRevA.105.012411.DOI
- [32]Dax Enshan Koh, Guoming Wang, Peter D. Johnson, and Yudong Cao. Foundations for bayesian inference with engineered likelihood functions for robust amplitude estimation. Journal of Mathematical Physics, 63(5):052202, 2022. doi: 10.1063/5.0042433.DOI
- [33]Archismita Dalal and Amara Katabarwa. Noise tailoring for robust amplitude estimation. New Journal of Physics, 25(2):023015, 2023. doi: 10.1088/1367-2630/acb5bc.DOI
- [34]Alexandra Ramôa and Luis Paulo Santos. Bayesian quantum amplitude estimation. Quantum, 9:1856, 2025. doi: 10.22331/q-2025-09-11-1856.DOI



