Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2611.11748

Model-family selection and recovery-gate design in generator-based noise tomography

Long Huang

Published Nov 9, 2026• 0 citations
arXivHTMLPDF

Abstract

Family selection in generator-model noise tomography, usually quoted as a hardware property, is a hypothesis test whose shot floor is closed form under a purely coherent noise model, $N^{*}=λ_{\mathrm{req}}/(r_{\mathrm{b}} S)^{2}$, controlled by the absolute bias $r_{\mathrm{b}} S$ rather than by the relative residual $r_{\mathrm{b}}$, hence computable for any device from a calibration, before a single shot. The same computation prices the recovery gates: the fingerprint that fixes a family's residual and the Fisher information that ranks designs come from the same map, so there is no design-independent verdict and no family-independent design optimum. Whether a design can move that price is structural: a bias confined to the block every design measures identically is design-blind, while one that reaches the code blocks can be repriced by a factor of sixty-eight; the two platforms are its extremes, at $98.1\%$ and $48.0\%$. At the reference point the floor is $1.1\times10^{7}$ and $6.7\times10^{6}$ shots per component, and that point is the easy corner: the median of its own direction over the sampled points is $8.9\times10^{8}$, $83$ times dearer. At the $10^{5}$ shots a run spends the test has power $0.053$ at a level of $0.05$. Rebuilding the fingerprint under each of the sixty-four designs moves the floor by $68$ on one platform and $1.7$ on the other, and the most precise design costs ten times more shots for the verdict than the cheapest to test. All sixty-four placements are full rank: at three gates the rms varies by factors of $2.7$ and $3.9$ with placement alone, while the optimum over counts is flat. At five qubits the same rule moves the floor by $21$ while the verdict costs $10^{9}$ shots; against a real three-qubit calibration the floor is a requirement on systematic control before it is a shot count, and where that control is missing no budget supplies it.

DOI
10.48550/arxiv.2611.11748

Related papers

A Hard Worst-Case Floor for All-Parameter Recoverability in Three-Qubit Coherent-Noise Fingerprinting

Long Huang

Supplementary note to arXiv:2611.08310 and arXiv:2610.19779. A fingerprint of coherent noise built from a stabilizer code is only as informative as the recovery gates that produce it, and every fixed choice of those gates is blind somewhere. This note isolates what can nevertheless be guaranteed, and states it as a worst-case floor rather than an average improvement. Two elementary facts about Fisher information make the guarantee exact. Because information matrices add, the identifiable set of a collection of designs is the union of theirs, so running several designs in parallel can never cost a direction that any member sees. Because the inverse of a sum is Loewner-dominated by the inverse of any summand, splitting a fixed shot budget over S designs can never cost more than sqrt(S) in standard deviation, that is, S in shots. The premium is therefore fixed in advance and independent of the working point, while the gain is not. Both statements are certified over the 64-design space of the three-qubit repetition code, under the 48- and 112-component fingerprints: the Loewner inequality is evaluated rather than assumed at every point tested, the worst case over 300 working points is 1.25 against the certified sqrt(2) = 1.4142, the bound is attained to five figures at one structured point, and at the worst points of the individual designs the parallel pair sits below the better single design by factors of 2.6 to 9.5 under the 48-component fingerprint and 4.3 to 17 under the 112-component one. Two three-gate placements suffice; a third adds 3.5 to 11.5 per cent. The floor buys recoverability and worst-case precision, and it buys nothing else: it does not buy accuracy under common-mode calibration bias, and it does not repair a model degeneracy that no design can see. Converting the floor into a shot count: at a unit per-component error the pair brings the worst parameter to a 10 mrad target in a median of 7.5e4 shots under the 112-component fingerprint and 5.5e5 under the 48-component one, with 95th percentiles of 4.6e5 and 2.9e6. These are sample statistics and are labelled as such, because the parameter box contains a degenerate locus at which all 64 designs fall to rank 5: no design has a finite worst-case count over the whole box. A directed search from the worst sampled point raises the sample maximum by factors of 4.7e3 and 1.1e4 while leaving all three Jacobians at full rank, which is the honest size of that tail; over the whole domain the only statement that survives is the sqrt(2) inequality itself. Contents: the two propositions with proofs, the cost corollary, the numerical certification (six tables, one figure), the relation to the two companion papers, and a reproduction map. The archive repro_v1.0.zip in this record contains the eight scripts, five data files, the figure and a self-test that re-derives every quoted number; run python scripts/selftest.py from the archive root and compare against MANIFEST.sha256. Companion papers: arXiv:2611.08310 (recovery gates as a resource) and arXiv:2610.19779 (coherent-noise identifiability).

•Nov 15, 2026•0 citations

Experimental supplement: three-qubit, three-gate repetition-code noise tomography

Long Huang

Experimental supplement to arXiv:2611.08310, a companion manuscript on gate-optimal recovery designs (2026). This supplement specifies a concrete experiment for measuring the eleven-parameter small-angle coherent noise of a three-qubit register with three single-qubit Pauli gates, using a distance-three repetition code in both orientations. It gives the branch structure, the syndrome-to-Pauli look-up table for the three-gate design, the fixed sixteen-setting measurement table, shot and wall-clock budgets for three platform classes, a calibration and systematics checklist, an analysis pipeline, and a symptom-to-cause table. Two sections price the alternatives: a fixed recovery that does not resolve the syndrome saves a factor three in circuits and loses a factor five in precision floor, and a coherent unitary recovery keeps the circuit saving but replaces each single-qubit Pauli by a multi-controlled gate that every branch must compile. At three qubits no design in the scanned budget loses rank, so identifiability does not separate the gate counts; what depends on the working point is which three gates to use, not how many. The organising point is that the recovery gate count is a resource: three gates attain the lowest precision floor, the standard six-gate decoder is 22 times worse at a small-angle operating point and 2.3 times worse at the well-conditioned point of arXiv:2610.19779, and the ordering is stable over forty sampled working points while the size of the factor is not. The ordering survives when the look-up is compiled rather than measured, and the margin then widens from 5.6 to 7.7 in wall clock, because a compiled table charges every entry in every branch. A section added in version 1.5 asks whether that answer depends on the platform. A second device archetype -- a tunable coupler or iSWAP-style gate, whose always-on term is the exchange J(XX+YY)/2 -- carries one more parameter per bond, thirteen in all, and the two clusters are not nested: XX and YY commute, and ZZ is orthogonal to their plane. Enumerating all 64 design pairs at n=3 and all 3969 at n=5 gives the same gate budget for both archetypes, zero gates to reach full rank at n=3 and one at n=5; that count is a lower bound on the budget, not a recommended design, since the precision optimum is the three gates of the main text, where zero gates is 18 times worse at the median and 379 times worse at the certified point. At matched per-layer coherent infidelity the thirteen-parameter cluster is not worse (101.0 against 88.9 mrad on the local angles at 10^5 shots), and at the three-gate optimum the comparison widens (41.7 against 32.2 mrad). Choosing the wrong cluster is a bias, not a variance: the correct family leaves a residual of exactly zero, the wrong one leaves 0.4-0.6% in the fingerprint and moves the nine local angles by a quarter to a half, while its Cramer-Rao bound looks better. The decoded logical error cannot arbitrate at that bias level -- 200 random perturbations of the same total variation spread it over a factor 43 -- so the model family must be chosen by the residual. The same section closes with a composition law: a per-gate composition of Pauli probabilities is elementwise in the Walsh eigenvalues and therefore under-counts the coherent part of the depth-L noise by exactly L (measured slopes 1.98 for the generator against 0.99 for the per-gate rule). A closing section fixes the size of the gain and asks how far it extends. An exact prior removes 19.3% of the logical error at d=3, which is 4.3e-5 errors per code cycle; a three-gate fit captures 83, 96.5 and 99.5% of that share on the three platform classes. That share, not the accuracy of the estimate, is where every prior-based protocol stops, because 94% of the residual is weight-two aliasing. It is not a limit set by using Pauli probabilities: the same prior decodes both the coherent channel and its twirled version, so the drop from 19.3% to 11.6% under twirling prices the device rather than the estimate, and composing per-gate twirls changes at most two of the 256 decoder decisions. A generator rather than a channel is what additionally allows the coherent part to be inverted and the prior to be extrapolated to unmeasured working points. Carrying the same three-qubit fit to d=5 or d=7 obeys a closed form, G_d = G_3^((d+1)/4), so no extra precision is needed at larger distance: a 6.4 mrad rms on the eleven parameters moves the d=7 gain by 8.5%. What does not transport is the truncated part of the model---the 2(d-1)^2 three-body positions a d x d patch adds, none of which a chain fit covers---which costs 5% of the gain at d=5 and 10% at d=7; site-to-site spread costs another 3%. The law itself is a property of the decoder and applies to any prior, so this is a bound on what a three-qubit run can claim, not an advantage over Pauli-statistics protocols. Measuring at the target size instead of transporting avoids the truncation count but pays conditioning: a factor 68 from three to seven qubits. All numbers are model-level: no device data are reported, and the floors assume a calibration residual of 1e-3. Contents (v1.5): 1 What the experiment measures · 2 The three branches · 3 Syndrome decoding · 4 Settings and shot arithmetic · 5 Budget on three platforms · 6 Why the syndromes must be resolved · 7 What does and does not depend on the gate count · 8 Does the platform change the answer? · 9 How large the gain is, and how far it extends · 10 Calibration and systematics · 11 Analysis pipeline · 12 Diagnostics · 13 Checklist.

•Nov 6, 2026•3 citations

Certification cost of quantum models: measurement correlation, not parameter count

Pavel Sulimov, Claude Lehmann

Reporting the Fisher geometry of a trained variational quantum model is routine; quoting the shot budget that would establish it is not. Certifying an empirical Fisher matrix to relative Frobenius error $\varepsilon$ under coordinate-wise parameter shift costs $Θ(B p^{2} V/(\varepsilon^{2} G))$ circuit executions, where $V$ is the measured readout variance and $G$ the measured squared gradient norm, with uniform allocation optimal in that class. One constant reproduces the cost of two circuit families whose exponents differ by a full power of $p$. The exponent is an identity in how $nV$ and $nG$ scale with the register, holding family by family to $0.001$ across 624 matrix-product-state cells once the finite-$p$ prefactor is removed. The cubic cost is therefore a finite-size window, set by whether the readout light cone grows with the register. A product family to 256 qubits gives $1.966$ (95% CI $1.934$--$1.997$); a brickwork entangler falls from $2.853$ below ten qubits to $1.715$ beyond sixty-four; a blocked entangler gives $1.984$ at a fixed cone width against $3.034$ at a proportional one. Fixed device connectivity fixes the cone, so a cubic budget from a small simulation overestimates a large machine, on top of hardware multipliers $2.07\times$ ($1.41$--$3.02$), $2.38\times$ and $1.91\times$ on ibm_marrakesh, ibm_fez and ibm_kingston. Cost-optimal readout weights cut the measured shot budget by $2.67\times$ ($1.33$--$4.00$) on hardware, flat from four to twelve qubits. A discrepancy model fitted on cheap circuits transfers its mean inside the calibration grid and, at six larger sizes named before the data, does not: nominal 90% intervals cover 36%, and split conformal is the only rung that stays near nominal.

quant-pharXiv:2609.14424•Sep 15, 2026•0 citations

Recovery gates as a resource in repetition-code process tomography: exhaustive gate-optimal designs for three, five and seven qubits

Long Huang

Process tomography built on stabilizer codes compares the logical channel of a repetition code with a reference channel, and the decoder that produces it is almost always fixed to the standard minimum-weight one. Under a fixed measurement basis, where the gate count is a resource rather than a bookkeeping artefact, we enumerate every recovery design for three, five and seven qubits and decide the rank of each exactly. The minimal counts are $0$, $1$ and $5$ single-qubit Pauli gates, against $6$, $50$ and $308$ for full decoding, and the count depends on how many sites are corrected, never on which. Correcting more sites is not monotonically better: at seven qubits full decoding has rank $24$ of $27$, proved in rational arithmetic. At the median of $40$ working points the three-gate design is $9.0$ times more precise than the six-gate decoder, and the ordering holds at every one.

arXiv:2611.08310•Nov 5, 2026•1 citation

Noise-Field Sampling and Regime Transitions in Qubit Layout Selection: Hardware Evidence on IBM Superconducting Processors

Saripudi Suneetha, Kakumani K. C. Deepthi

Qubit layout selection determining which physical qubits a logical circuit occupies on a superconducting processor is fundamentally a noise-field sampling problem. On IBM Heron r1 and Heron r2 devices, two-qubit gate error varies by nearly one order of magnitude (7– $9\times $ ) across qubit pairs on the same chip, causing measured fidelity differences up to 0.903 from layout choice alone. We introduce the Gate Noise Sensitivity Index (GNSI), a calibration-weighted linear combination of per-qubit readout error and per-edge two-qubit gate error, derived analytically as a first-order approximation of circuit fidelity under independent, heterogeneous noise channels. For 12 backend–circuit combinations on three IBM processors (133-qubit Heron r1 and 156-qubit Heron r2), GNSI scores are computed for the full pool of 1,320 computationally enumerated candidate layouts; the top-ranked (GNSI-BEST) and bottom-ranked (GNSI-WORST) layouts are then executed on hardware. GNSI-BEST ranks first by score in all 12 instances (Hit@1 = 100% within the enumerated candidate pool) and consistently achieves higher measured fidelity than GNSI-WORST, with Spearman rank correlation $\rho = 0.926 \pm 0.016$ and statistically significant fidelity separation in 9 of 12 paired tests. Initial hardware comparisons on ibm_fez ( $n=5$ ) show GNSI-BEST matching or exceeding SABRE-selected fidelity. We further identify a depth-dependent regime transition: gate-error heterogeneity appears to govern ranking in shallow circuits (depth $\leq 15$ ), while qubit-specific $T_{1}$ variation may dominate for deeper circuits (depth $\geq 20$ ). The analytic regime boundary $d^{*}(b)$ , computable from calibration data alone, shows agreement within ±2 depth layers across all three backends in initial comparisons; GNSI-Extended incorporating the decoherence term receives an initial hardware comparison on two circuits, with full validation left to future work. These results suggest that layout-dependent fidelity variation may be governed by a calibration-predictable noise field, with direct implications for compiler design and noise-adaptive quantum circuit execution.

IEEE Access•Jan 1, 2026•0 citations

Compact physical Gaussian-ket models for homodyne quantum-state tomography

Wataru Kawashima

Continuous-variable quantum-state tomography usually reconstructs a truncated Fock-basis density matrix. We study a compact alternative: finite mixtures of displaced, squeezed Gaussian kets whose ρ = BB† construction is positive semidefinite by construction, fitted by per-sample homodyne likelihood with closed-form gradients, and composed with physical loss and noise channels. On the public propagating-light GKP dataset of Konno et al. (Science 2024), a rank-4 model with 92 real parameters matches the empirical full-rank maximum-likelihood frontier (255 parameters) at confidence-interval resolution on held-out likelihood. On a synthetic thermal-noise target that we prove lies outside the model family — no detection efficiency and no finite rank reproduces it exactly — the channel-composed model fitted blind exceeds a full-rank MLE run under the pre-declared 900-second baseline budget, a verdict that holds across all five pre-declared seed and noise configurations. We do not claim a universally superior method: comparisons on real data reuse observations across splits, the strongest baseline is test-selected, and the blind result covers one target class. The contribution is a compact, physically constrained model family together with a fully falsification-first research record — negative results, superseded scorings, and pre-declared protocols are all preserved in the accompanying repository.

•Jul 20, 2026•0 citations
Family selection in generator-model noise tomography, usually quoted as a hardware property, is a hypothesis test whose shot floor is closed form under a purely coherent noise model, N∗=λreq/(rbS)2N^{*}=λ_{\mathrm{req}}/(r_{\mathrm{b}} S)^{2}N∗=λreq​/(rb​S)2, controlled by the absolute bias rbSr_{\mathrm{b}} Srb​S rather than by the relative residual rbr_{\mathrm{b}}rb​, hence computable for any device from a calibration, before a single shot. The same computation prices the recovery gates: the fingerprint that fixes a family's residual and the Fisher information that ranks designs come from the same map, so there is no design-independent verdict and no family-independent design optimum. Whether a design can move that price is structural: a bias confined to the block every design measures identically is design-blind, while one that reaches the code blocks can be repriced by a factor of sixty-eight; the two platforms are its extremes, at 98.1%98.1\%98.1% and 48.0%48.0\%48.0%. At the reference point the floor is 1.1×1071.1\times10^{7}1.1×107 and 6.7×1066.7\times10^{6}6.7×106 shots per component, and that point is the easy corner: the median of its own direction over the sampled points is 8.9×1088.9\times10^{8}8.9×108, 838383 times dearer. At the 10510^{5}105 shots a run spends the test has power 0.0530.0530.053 at a level of 0.050.050.05. Rebuilding the fingerprint under each of the sixty-four designs moves the floor by 686868 on one platform and 1.71.71.7 on the other, and the most precise design costs ten times more shots for the verdict than the cheapest to test. All sixty-four placements are full rank: at three gates the rms varies by factors of 2.72.72.7 and 3.93.93.9 with placement alone, while the optimum over counts is flat. At five qubits the same rule moves the floor by 212121 while the verdict costs 10910^{9}109 shots; against a real three-qubit calibration the floor is a requirement on systematic control before it is a shot count, and where that control is missing no budget supplies it.
Supplementary note to arXiv:2611.08310 and arXiv:2610.19779. A fingerprint of coherent noise built from a stabilizer code is only as informative as the recovery gates that produce it, and every fixed choice of those gates is blind somewhere. This note isolates what can nevertheless be guaranteed, and states it as a worst-case floor rather than an average improvement. Two elementary facts about Fisher information make the guarantee exact. Because information matrices add, the identifiable set of a collection of designs is the union of theirs, so running several designs in parallel can never cost a direction that any member sees. Because the inverse of a sum is Loewner-dominated by the inverse of any summand, splitting a fixed shot budget over S designs can never cost more than sqrt(S) in standard deviation, that is, S in shots. The premium is therefore fixed in advance and independent of the working point, while the gain is not. Both statements are certified over the 64-design space of the three-qubit repetition code, under the 48- and 112-component fingerprints: the Loewner inequality is evaluated rather than assumed at every point tested, the worst case over 300 working points is 1.25 against the certified sqrt(2) = 1.4142, the bound is attained to five figures at one structured point, and at the worst points of the individual designs the parallel pair sits below the better single design by factors of 2.6 to 9.5 under the 48-component fingerprint and 4.3 to 17 under the 112-component one. Two three-gate placements suffice; a third adds 3.5 to 11.5 per cent. The floor buys recoverability and worst-case precision, and it buys nothing else: it does not buy accuracy under common-mode calibration bias, and it does not repair a model degeneracy that no design can see. Converting the floor into a shot count: at a unit per-component error the pair brings the worst parameter to a 10 mrad target in a median of 7.5e4 shots under the 112-component fingerprint and 5.5e5 under the 48-component one, with 95th percentiles of 4.6e5 and 2.9e6. These are sample statistics and are labelled as such, because the parameter box contains a degenerate locus at which all 64 designs fall to rank 5: no design has a finite worst-case count over the whole box. A directed search from the worst sampled point raises the sample maximum by factors of 4.7e3 and 1.1e4 while leaving all three Jacobians at full rank, which is the honest size of that tail; over the whole domain the only statement that survives is the sqrt(2) inequality itself. Contents: the two propositions with proofs, the cost corollary, the numerical certification (six tables, one figure), the relation to the two companion papers, and a reproduction map. The archive repro_v1.0.zip in this record contains the eight scripts, five data files, the figure and a self-test that re-derives every quoted number; run python scripts/selftest.py from the archive root and compare against MANIFEST.sha256. Companion papers: arXiv:2611.08310 (recovery gates as a resource) and arXiv:2610.19779 (coherent-noise identifiability).
Experimental supplement to arXiv:2611.08310, a companion manuscript on gate-optimal recovery designs (2026). This supplement specifies a concrete experiment for measuring the eleven-parameter small-angle coherent noise of a three-qubit register with three single-qubit Pauli gates, using a distance-three repetition code in both orientations. It gives the branch structure, the syndrome-to-Pauli look-up table for the three-gate design, the fixed sixteen-setting measurement table, shot and wall-clock budgets for three platform classes, a calibration and systematics checklist, an analysis pipeline, and a symptom-to-cause table. Two sections price the alternatives: a fixed recovery that does not resolve the syndrome saves a factor three in circuits and loses a factor five in precision floor, and a coherent unitary recovery keeps the circuit saving but replaces each single-qubit Pauli by a multi-controlled gate that every branch must compile. At three qubits no design in the scanned budget loses rank, so identifiability does not separate the gate counts; what depends on the working point is which three gates to use, not how many. The organising point is that the recovery gate count is a resource: three gates attain the lowest precision floor, the standard six-gate decoder is 22 times worse at a small-angle operating point and 2.3 times worse at the well-conditioned point of arXiv:2610.19779, and the ordering is stable over forty sampled working points while the size of the factor is not. The ordering survives when the look-up is compiled rather than measured, and the margin then widens from 5.6 to 7.7 in wall clock, because a compiled table charges every entry in every branch. A section added in version 1.5 asks whether that answer depends on the platform. A second device archetype -- a tunable coupler or iSWAP-style gate, whose always-on term is the exchange J(XX+YY)/2 -- carries one more parameter per bond, thirteen in all, and the two clusters are not nested: XX and YY commute, and ZZ is orthogonal to their plane. Enumerating all 64 design pairs at n=3 and all 3969 at n=5 gives the same gate budget for both archetypes, zero gates to reach full rank at n=3 and one at n=5; that count is a lower bound on the budget, not a recommended design, since the precision optimum is the three gates of the main text, where zero gates is 18 times worse at the median and 379 times worse at the certified point. At matched per-layer coherent infidelity the thirteen-parameter cluster is not worse (101.0 against 88.9 mrad on the local angles at 10^5 shots), and at the three-gate optimum the comparison widens (41.7 against 32.2 mrad). Choosing the wrong cluster is a bias, not a variance: the correct family leaves a residual of exactly zero, the wrong one leaves 0.4-0.6% in the fingerprint and moves the nine local angles by a quarter to a half, while its Cramer-Rao bound looks better. The decoded logical error cannot arbitrate at that bias level -- 200 random perturbations of the same total variation spread it over a factor 43 -- so the model family must be chosen by the residual. The same section closes with a composition law: a per-gate composition of Pauli probabilities is elementwise in the Walsh eigenvalues and therefore under-counts the coherent part of the depth-L noise by exactly L (measured slopes 1.98 for the generator against 0.99 for the per-gate rule). A closing section fixes the size of the gain and asks how far it extends. An exact prior removes 19.3% of the logical error at d=3, which is 4.3e-5 errors per code cycle; a three-gate fit captures 83, 96.5 and 99.5% of that share on the three platform classes. That share, not the accuracy of the estimate, is where every prior-based protocol stops, because 94% of the residual is weight-two aliasing. It is not a limit set by using Pauli probabilities: the same prior decodes both the coherent channel and its twirled version, so the drop from 19.3% to 11.6% under twirling prices the device rather than the estimate, and composing per-gate twirls changes at most two of the 256 decoder decisions. A generator rather than a channel is what additionally allows the coherent part to be inverted and the prior to be extrapolated to unmeasured working points. Carrying the same three-qubit fit to d=5 or d=7 obeys a closed form, G_d = G_3^((d+1)/4), so no extra precision is needed at larger distance: a 6.4 mrad rms on the eleven parameters moves the d=7 gain by 8.5%. What does not transport is the truncated part of the model---the 2(d-1)^2 three-body positions a d x d patch adds, none of which a chain fit covers---which costs 5% of the gain at d=5 and 10% at d=7; site-to-site spread costs another 3%. The law itself is a property of the decoder and applies to any prior, so this is a bound on what a three-qubit run can claim, not an advantage over Pauli-statistics protocols. Measuring at the target size instead of transporting avoids the truncation count but pays conditioning: a factor 68 from three to seven qubits. All numbers are model-level: no device data are reported, and the floors assume a calibration residual of 1e-3. Contents (v1.5): 1 What the experiment measures · 2 The three branches · 3 Syndrome decoding · 4 Settings and shot arithmetic · 5 Budget on three platforms · 6 Why the syndromes must be resolved · 7 What does and does not depend on the gate count · 8 Does the platform change the answer? · 9 How large the gain is, and how far it extends · 10 Calibration and systematics · 11 Analysis pipeline · 12 Diagnostics · 13 Checklist.
Reporting the Fisher geometry of a trained variational quantum model is routine; quoting the shot budget that would establish it is not. Certifying an empirical Fisher matrix to relative Frobenius error ε\varepsilonε under coordinate-wise parameter shift costs Θ(Bp2V/(ε2G))Θ(B p^{2} V/(\varepsilon^{2} G))Θ(Bp2V/(ε2G)) circuit executions, where VVV is the measured readout variance and GGG the measured squared gradient norm, with uniform allocation optimal in that class. One constant reproduces the cost of two circuit families whose exponents differ by a full power of ppp. The exponent is an identity in how nVnVnV and nGnGnG scale with the register, holding family by family to 0.0010.0010.001 across 624 matrix-product-state cells once the finite-ppp prefactor is removed. The cubic cost is therefore a finite-size window, set by whether the readout light cone grows with the register. A product family to 256 qubits gives 1.9661.9661.966 (95% CI 1.9341.9341.934--1.9971.9971.997); a brickwork entangler falls from 2.8532.8532.853 below ten qubits to 1.7151.7151.715 beyond sixty-four; a blocked entangler gives 1.9841.9841.984 at a fixed cone width against 3.0343.0343.034 at a proportional one. Fixed device connectivity fixes the cone, so a cubic budget from a small simulation overestimates a large machine, on top of hardware multipliers 2.07×2.07\times2.07× (1.411.411.41--3.023.023.02), 2.38×2.38\times2.38× and 1.91×1.91\times1.91× on ibm_marrakesh, ibm_fez and ibm_kingston. Cost-optimal readout weights cut the measured shot budget by 2.67×2.67\times2.67× (1.331.331.33--4.004.004.00) on hardware, flat from four to twelve qubits. A discrepancy model fitted on cheap circuits transfers its mean inside the calibration grid and, at six larger sizes named before the data, does not: nominal 90% intervals cover 36%, and split conformal is the only rung that stays near nominal.
Process tomography built on stabilizer codes compares the logical channel of a repetition code with a reference channel, and the decoder that produces it is almost always fixed to the standard minimum-weight one. Under a fixed measurement basis, where the gate count is a resource rather than a bookkeeping artefact, we enumerate every recovery design for three, five and seven qubits and decide the rank of each exactly. The minimal counts are 000, 111 and 555 single-qubit Pauli gates, against 666, 505050 and 308308308 for full decoding, and the count depends on how many sites are corrected, never on which. Correcting more sites is not monotonically better: at seven qubits full decoding has rank 242424 of 272727, proved in rational arithmetic. At the median of 404040 working points the three-gate design is 9.09.09.0 times more precise than the six-gate decoder, and the ordering holds at every one.
Qubit layout selection determining which physical qubits a logical circuit occupies on a superconducting processor is fundamentally a noise-field sampling problem. On IBM Heron r1 and Heron r2 devices, two-qubit gate error varies by nearly one order of magnitude (7– $9\times $ ) across qubit pairs on the same chip, causing measured fidelity differences up to 0.903 from layout choice alone. We introduce the Gate Noise Sensitivity Index (GNSI), a calibration-weighted linear combination of per-qubit readout error and per-edge two-qubit gate error, derived analytically as a first-order approximation of circuit fidelity under independent, heterogeneous noise channels. For 12 backend–circuit combinations on three IBM processors (133-qubit Heron r1 and 156-qubit Heron r2), GNSI scores are computed for the full pool of 1,320 computationally enumerated candidate layouts; the top-ranked (GNSI-BEST) and bottom-ranked (GNSI-WORST) layouts are then executed on hardware. GNSI-BEST ranks first by score in all 12 instances (Hit@1 = 100% within the enumerated candidate pool) and consistently achieves higher measured fidelity than GNSI-WORST, with Spearman rank correlation ρ=0.926±0.016\rho = 0.926 \pm 0.016ρ=0.926±0.016 and statistically significant fidelity separation in 9 of 12 paired tests. Initial hardware comparisons on ibm_fez ( n=5n=5n=5 ) show GNSI-BEST matching or exceeding SABRE-selected fidelity. We further identify a depth-dependent regime transition: gate-error heterogeneity appears to govern ranking in shallow circuits (depth ≤15\leq 15≤15 ), while qubit-specific T1T_{1}T1​ variation may dominate for deeper circuits (depth ≥20\geq 20≥20 ). The analytic regime boundary d∗(b)d^{*}(b)d∗(b) , computable from calibration data alone, shows agreement within ±2 depth layers across all three backends in initial comparisons; GNSI-Extended incorporating the decoherence term receives an initial hardware comparison on two circuits, with full validation left to future work. These results suggest that layout-dependent fidelity variation may be governed by a calibration-predictable noise field, with direct implications for compiler design and noise-adaptive quantum circuit execution.
Continuous-variable quantum-state tomography usually reconstructs a truncated Fock-basis density matrix. We study a compact alternative: finite mixtures of displaced, squeezed Gaussian kets whose ρ = BB† construction is positive semidefinite by construction, fitted by per-sample homodyne likelihood with closed-form gradients, and composed with physical loss and noise channels. On the public propagating-light GKP dataset of Konno et al. (Science 2024), a rank-4 model with 92 real parameters matches the empirical full-rank maximum-likelihood frontier (255 parameters) at confidence-interval resolution on held-out likelihood. On a synthetic thermal-noise target that we prove lies outside the model family — no detection efficiency and no finite rank reproduces it exactly — the channel-composed model fitted blind exceeds a full-rank MLE run under the pre-declared 900-second baseline budget, a verdict that holds across all five pre-declared seed and noise configurations. We do not claim a universally superior method: comparisons on real data reuse observations across splits, the strongest baseline is test-selected, and the blind result covers one target class. The contribution is a compact, physically constrained model family together with a fully falsification-first research record — negative results, superseded scorings, and pre-declared protocols are all preserved in the accompanying repository.