WorksComputer Science
Red-Blue Hamiltonian Tomography via the Operator Lightcone Wavemap
Abstract
We introduce a model-free method for Hamiltonian tomography based on the operator spreading wavefront. Given two tensor network simulations of the same quantum lattice (one with the nominal Hamiltonian and one with one or more perturbed bonds), the signed cross-entropy of the non-identity Pauli weight field identifies each perturbed bond by its causal signature: a red-blue sign flip at the two endpoints, visible from the first wavefront crossing. Combining measurements from strategically placed observation centers via an ensemble log-likelihood sum, a single broken bond is identified with an 82-fold margin over the next candidate on a lattice. Extending to realistic coupling degradation () and up to 8 simultaneous faults, the integrated log-ratio achieves a per-bond detection rate above 90% in a single seismic pass using only 6 simulations. An iterative re-scan strategy progressively resolves multi-bond cases as the signal clears with each recalibration pass. The method requires no model assumptions and no knowledge of the perturbation type or magnitude; the detection is deliberately blind.
Introduction
Identifying miscalibrated or anomalous couplings in a quantum processor is a central challenge in quantum hardware characterization. Standard approaches require explicit Hamiltonian learning: fitting a parameterized coupling model to measurement data. This scales poorly with system size and depends on model assumptions about the noise and interaction structure.
We propose a different approach based on a clear separation of roles. The simulator knows the nominal Hamiltonian and produces two Pauli weight fields: from the nominal evolution, and from an evolution where one bond has been silently modified. The detection algorithm receives only these two arrays of numbers and must identify the modified bond from their ratio alone, with no knowledge of the Hamiltonian, the perturbation type, or its magnitude. The detection is deliberately blind.
A perturbed bond creates a causal signature in the wavefront: a sign flip in the cross-entropy field between its two endpoints that propagates no further. The broken bond is exactly where the wavefront sign changes, identifiable from the observables alone.
The method is inspired by seismic tomography: place sensors (observation centers) at strategic positions, emit a wave (initialize a local operator), and read the arrival pattern. The perturbed bond is the anomaly in the medium.
The red-blue color scheme is a deliberate reference to the red-blue pebble game of Hong and Kung [2], which characterizes I/O complexity as a bottleneck in computational graphs. Here the roles are analogous: red sites (upstream, wave blocked) and blue sites (downstream, wave starved) identify the bond that limits operator propagation, exactly as red and blue pebbles identify the I/O bottleneck in a computation.
Setup
We simulate the kicked isotropic Heisenberg model on a rectangular lattice (49 sites, 84 bonds) using a GPU-accelerated belief propagation tensor network simulator with Trotter step , bond dimension , and truncation cutoff . The non-identity Pauli weight at site and cycle is:
where is the evolved operator initialized at center . By normalization , is a probability.
To simulate miscalibrated bonds, we weaken their couplings by , with each bond potentially taking a different value. In the experiments below we use (, a 20% degradation) as a representative choice; the detection method does not depend on knowing in advance and works for any coupling deviation. The Pauli transfer matrix entries scale as , attenuating operator propagation without blocking it entirely. For the multi-bond experiments, bonds are selected uniformly at random and perturbed simultaneously.
We place observation centers using a greedy geodesic bond cover: the minimum set of sites such that every bond is within hops of at least one center. For the lattice this gives centers at sites , covering all 84 bonds.
Signal Vocabulary and Measurement Definitions
All definitions are expressed in terms of two observables per site per cycle: and , the non-identity Pauli weights under the nominal and perturbed Hamiltonian respectively. The BFS frontier crossing time defines a site-specific reference time; it marks the physical arrival of the wavefront at , set by the Lieb-Robinson lightcone, not by the numerical cutoff.
At each site in the 3-cycle window we observe one of three frontier patterns. A site is White [W] when : the baseline wave has not yet arrived and no information is available. A site is Blue [B] when : the perturbed wave accumulates at because drainage is blocked by a broken bond downstream; the site is a wall face. A site is Red [R] when with : the perturbed wave is attenuated because a broken bond on the path from center blocked transmission; the site is in the shadow.
For a bond with on the center-side and on the shadow-side, these patterns combine into three observable signatures. When both endpoints are white [W, W, W] the bond has not yet been reached or is healthy. When the wall face is detected but the shadow is not yet visible [W, B, W], one more cycle is needed. The broken bond signature [W, B, R] is the primary target: appears blue at cycle and appears red at cycle , identifying the bond between them as anomalous. This pattern rests on a single assumption: both waves reach at the same frontier time , so any amplitude difference at is attributable to the bond, not to a timing offset. The temporal order (blue before red by one cycle) is then universal, following from the BFS causal structure: is reached before , so the wall accumulates before the shadow appears.
The signed cross-entropy evaluated in the frontier window is:
This is a cross-correlation fixed in space: both waves evaluated at the same site at the same frontier moment , with the window providing context (white before, red after). Positive identifies red (shadow), negative identifies blue (wall face). The broken bond is the edge between consecutive blue and red sites in the BFS ordering.
A complementary observable measures the wave through a bond rather than at a site. Consider the bond in the BFS path from the center. The baseline reaches at frontier time and at . The baseline transfer ratio is:
This measures how much of the wave at propagates through the bond to in one BFS step, directly encoding the gate’s coupling strength . For the perturbed run (with potentially delayed frontier times ):
The ratio signals a broken bond: less weight crosses the bond in the perturbed run.
Two important caveats apply. First, there is a path distortion. The perturbed wave arriving at has already traversed all upstream bonds. If any of those are also broken, is already weakened, not by bond but by upstream damage. The ratio then conflates the upstream distortion with the target bond signal. The ratio is not a pure measurement of bond . Second, there is a measurable Sensitivity. problem. The perturbed weight may be close to , the threshold, making the ratio numerically unreliable. The local cross-correlation avoids both caveats by using the baseline as the reference; it measures the relative change at the frontier moment, not the absolute transfer.
The Cross-Entropy Wavemap Observable
The previous section introduced the signal vocabulary for a wave with a single origin: one center fires, the wavefront spreads, and the red-blue sign flip at the bond endpoints is the detection signal. Here we extend to the case where centers fire independently, each producing its own wavemap. The key quantities are defined per-center, then combined into an ensemble.
For each center and site , we evaluate the signed cross-entropy in a 3-cycle window centered on the BFS frontier crossing :
This is a cross-correlation fixed in space: both waves are evaluated at the same site and the same frontier moment , defined by the baseline. The window captures three states — before arrival (WHITE), wall-face signal (BLUE or WHITE), and shadow signal (RED) — and is the moment of maximum signal and minimum -saturation noise.
Consider a bond with closer to center . At time , the baseline wave has just crossed the bond and ; both waves have reached with equal amplitude. One cycle later at , the weaker bond slows drainage from : weight accumulates, , so (blue, wall face). Simultaneously, in the baseline weight flows through bond and arrives at , while in the perturbed run less weight crosses: , so (red, shadow). The temporal order is universal: blue at the wall face first, red at the shadow one cycle later. The broken bond sits between the blue and red sites, identified by the sign change (Figure 1). The contrast is maximum at , providing the detection signal.

Figure 1. Signed cross-entropy for a single broken bond (bond 9, sites 11↔12, ), rect , , 3 cycles. Rows show the three seismic scans: corners+center, quadrants, mid-edges. Columns show cycles 1–3. Cycle 1: no signal in any scan. Cycle 2: yellow (BLUE) accumulation visible at the wall-face site. Cycle 3: red (RED) shadow appears at the downstream site; sign flip across the bond identifies it in all three scans independently. Gold circles: bond endpoints. Stars: scan centers.
Each center is initialized and evolved independently: a separate baseline and perturbed simulation pair per center. The signed cross-entropy is computed for each center individually, then summed into an ensemble log-likelihood:
Centers on the same side of the broken bond reinforce each other; centers on opposite sides contribute opposite signs that partially cancel. Silent centers (wave never reaches site ) contribute zero and do not corrupt the sum.
A more robust alternative to the pointwise accumulates the signal over all cycles:
Positive identifies red (shadow), negative identifies blue (wall face). The integral captures both the delay and the sustained amplitude difference: a broken bond delays the perturbed wave and reduces its steady-state weight, making more robust than when the pointwise signal is weak. In the multi-bond experiments, is the primary detection signal.
The Log-Transfer Vector Field
The cross-entropy is a scalar field on sites. We now show that its natural spatio-temporal gradient is a vector field defined on bonds, and that this gradient field directly reveals the broken bond as a physical singularity.
Since is a scalar field on the graph, its discrete spatial gradient at site in the direction of neighbor is . Since is a probability, weight flows multiplicatively through bonds: , so the natural gradient is in log-space. The spatio-temporal gradient of at directed bond from cycle to is:
At each site , collecting this quantity over all neighbors (up to 4 for a rectangular lattice) gives a vector field:
one component per bond direction. Each component is a log-likelihood ratio: positive when the baseline transfers more than the perturbed run along that bond, negative when the perturbed transfers more.
A broken bond is a singularity in this gradient field. At the first wavefront crossing (cycle ), the baseline flows through the bond freely while the perturbed run is impeded: (red, baseline transfers more through the broken bond). One cycle later, weight has accumulated at the wall-face site in the perturbed run; grows, making the reverse ratio larger: (blue, perturbed transfers more in reverse because the shadow site is depleted). The two directed components of the same bond have opposite signs: the gradient field has a discontinuity exactly at the broken bond.
By subsequent cycles the perturbation diffracts: the wall-face accumulation radiates excess weight along alternative paths, enriching bonds that would normally receive less, while the shadow deepens. The gradient field develops a characteristic diffraction pattern centered on the broken bond, with two diagonal fringes whose orientation encodes the bond direction directly (Figure 2).

Figure 2. , broken bond 9 (sites 1112), . Red: baseline transfers more. Blue: perturbed transfers more.
The gradient field has several properties that distinguish it from the scalar . It is defined on bonds rather than sites, so the broken bond appears directly as a gradient discontinuity rather than through its effect on neighboring sites. It carries directional information: the sign asymmetry between forward and reverse components identifies which side is the wall face and which is the shadow, and the diffraction pattern orientation encodes the bond direction. The normalization by suppresses the dominant-site bias that affects site-based metrics. Systematic artifacts from the simulation (Trotter color ordering, BP approximation, or any other effect present in both runs) cancel in the ratio , whether and come from two simulations or from a simulation and hardware shots.
Seismic Multi-Center Scanning
Rather than running each center independently, we initialize all centers simultaneously in a single seismic scan: the initial operator is , firing all centers at once. At each site , the first-arrival signal is dominated by the nearest center, automatically selecting the strongest directional component. The sign flip at the broken bond endpoints emerges without any explicit path tracing or per-center bookkeeping, and the cost drops from simulations to just 2 (one baseline, one perturbed).
To achieve broad coverage with no geometric blind spots, we use three complementary center placements on the lattice: corners and center , quadrant centers , and mid-edge sites . Each placement covers a different region of the grid; together the three scans ensure every bond is within hops of at least one center in at least one scan. At hops/cycle, 2 hops requires 5 simulation cycles. The total cost is 6 simulations (3 baselines + 3 perturbed), run in parallel on 3 GPUs.
Bond scores are computed using the integrated log-ratio across all three scans. For each bond connecting sites and , the score is the maximum contrast over all scans. The top- bonds by score are declared broken, where is the number of expected faults. On 20 single-bond trials (, two independent datasets), this procedure achieves 20/20 correct identifications. On 10 two-bond trials it achieves 9/10, with the single failure corresponding to two bonds sharing an endpoint whose wavefronts cancel (the BLUE-BLUE-RED interference pattern discussed in § 5).
Bond Detection
For each bond connecting sites and we define a bond score:
A negative product indicates a sign flip across the bond, the causal signature of the broken bond. Rank bonds by most negative .
Table 1 shows the individual and ensemble results for bond 12 (sites 15 16, identity PTM perturbation, lattice, , 5 cycles).
| Center | Side | Result | ||
| 10 | upstream | +0.87 | −0.02 | ✓ CORRECT |
| 30 | upstream | +0.37 | −0.04 | ✓ CORRECT |
| 25 | downstream | +0.03 | −0.04 | × WRONG |
| 32 | downstream | +0.00 | −0.03 | × WRONG |
| 6 | too far | 0 | 0 | silent |
| 34 | too far | 0 | 0 | silent |
| Ensemble | — | +1.23 | −0.13 | ✓ CORRECT |
Table 1. Signed cross-entropy at bond 12 endpoints (sites 15,16). Ensemble score: , next bond (82× margin).
The ensemble identifies bond 12 with an 82-fold margin over the next candidate, despite only 2 of 6 individual centers giving the correct answer. The wrong individual answers: centers 25 and 32 both correctly identify site 16 as a bond endpoint; their error is only in declaring the wrong adjacent bond. The ensemble product cancels these partial errors.
The gradient field of § 5 provides a complementary bond score. For each directed bond , summing above a threshold over all scans and cycles gives:
with threshold . The broken bond is the one with the largest asymmetry between forward and reverse directions. Table 2 compares the two methods across bond counts.
| bonds | from | from |
| 1 | 20/20 | 20/20 |
| 2 | 9/10 | 9/10 |
| 4 | 3/5 ALL | 1/5 ALL |
| 8 | mixed | avg 6/8 |
Table 2. Detection rate (ALL bonds correct) for the two bond scores. Both achieve 20/20 single-bond and 9/10 two-bond. They fail on different trials at higher bond counts, making them complementary.
Both methods support an iterative calibration strategy: declare the top- broken bonds, recalibrate the hardware, then re-run the scan. In each pass the already-fixed bonds no longer contaminate the signal, so the remaining faults surface more clearly. The process stops when everywhere: a physical criterion that requires no additional measurement overhead.
Tensor Network Simulation as Quantum Circuit Execution
The Trotter simulation underlying our method is not merely an approximation of a quantum circuit; it is the quantum circuit, represented in the Pauli operator basis rather than the amplitude basis. The identification is exact:
Each lattice site = one qubit
Each bond gate = one two-qubit gate
Each color class (non-overlapping bonds applied in parallel) = one gate layer
Each belief propagation synchronization after a full Trotter cycle = one barrier (per-color-class BP, as in the Julia L-MPF implementation [3], provides finer sub-cycle resolution matching the quantum circuit barrier structure)
Each PAULI NL output = a mid-circuit Pauli weight measurement
Both the graph coloring (determining which bonds share qubits) and the quantum circuit parallelism constraint (which gates can be applied simultaneously) derive from the lattice topology. The color class ordering is the gate layer ordering.
This identification reveals two distinct timescales. Physical time counts integer Trotter cycles, ; the observables and at cycle granularity are physical, independent of color class ordering to and convergent to true quantum dynamics. Circuit time has sub-cycle, per-color-class granularity, a property of the circuit implementation rather than the Hamiltonian. Two Trotter implementations of the same Hamiltonian with different color orderings agree on to first order in but differ by Trotter errors; their tomography fingerprints differ already at the first sub-cycle. The BP synchronization granularity determines which timescale is accessible: per-cycle BP (as used here) loses sub-step information; per-color BP (Julia L-MPF [3]) respects the causal barrier structure of the quantum circuit.
A perturbed bond (coupling ) is a miscalibrated gate: the two-qubit rotation angle differs from the nominal value. The Pauli transfer matrix for has diagonal entries , which for any gives . Since for all (unitary gate), any coupling deviation attenuates the wave relative to baseline, guaranteeing at the shadow endpoint , so (red) regardless of the sign of . The detection method is robust: the red site is always the shadow endpoint, never inverted. The gradient field of § 5 reveals the complementary effect: while the bond attenuates the wave on the shadow side, it simultaneously causes accumulation on the wall-face side and, by subsequent cycles, amplification along rerouted paths that would otherwise receive less weight.
Belief propagation is exact on trees: messages from disjoint subtrees are independent and the fixed point is unique. The seismic initialization fires multiple centers simultaneously, creating wavefronts that meet and interfere inside the graph. This interference violates the tree assumption structurally: circular message dependencies prevent convergence regardless of iteration count. It is a property of the multi-center initialization, not of the solver.
The detection observable is a ratio of two runs that share the same centers, the same graph, and the same BP schedule. Whatever interference pattern the multi-center initialization creates is therefore structurally identical in both runs. It is not an approximation error to be corrected; it is a shared background that cancels exactly in the ratio. Only the bond-specific perturbation survives. This was verified empirically: running with up to 30,000 damped BP iterations (which achieve convergence for small center sets) gives PAULI_NL values identical to standard 100-iteration BP. The residual BP non-convergence leaves unchanged.
A complete theoretical characterization of multi-center BP on loopy graphs remains open and will ultimately be resolved by hardware experiments where is measured directly, with no BP involved.
This frames our tomography as quantum ATPG (Automatic Test Pattern Generation) [4], the classical circuit testing methodology generalized to quantum gates. In classical ATPG: fault activation (excite the stuck-at gate), path sensitization (find an observable path from fault to output), backward implication (infer the faulty gate). The quantum analog is: perturbation (change ), wavefront propagation (operator spreading from center to shadow site), and cross-entropy inversion (red site + center position bond identity).
Causal Structure, Hardware Path, and Detection Algorithm
The observable is accessible on real quantum hardware via local Pauli measurements after ZNE gate-folding circuits [1]. The cross-entropy requires only the ratio of two measurements, with no absolute calibration. The center placement is determined by the graph topology alone, independent of the Hamiltonian.
Our method exploits the Lieb-Robinson bound: information cannot propagate faster than hops/cycle. A broken bond creates a causal anomaly; the wavefront on one side cannot inform the other. The first-arrival time is observed directly, without knowing in advance.
The seismic 3-scan cost is fixed at 6 simulations regardless of how many bonds are broken. On the lattice (, 5 cycles, AMD MI100 32 GB GPU), each scan pair runs in 272 s using 11% VRAM with 2 GPUs in parallel.
When two adjacent broken bonds share an endpoint, say bonds and with shared site , the cross-entropy field shows a three-site pattern: site appears red (shadow of bond ), site appears blue (wall face of bond ), and the shared site appears white. At , the red contribution from one broken bond and the blue contribution from the other cancel, producing an interference node. The white fringe is not a detection failure; it is an interference pattern arising from opposite-side contributions canceling at the shared node. The current algorithm detects such pairs via a triplet scorer , skipping the white shared endpoint and reading the outer sites directly.
The seismic and aperture methods are naturally complementary and suggest a two-phase adaptive protocol. In Phase 1 (seismic), all centers fire simultaneously: one quantum experiment, minimum cost, producing a map of RED sites that identifies the region of interest. In Phase 2 (aperture), the RED sites from Phase 1 guide the selection of targeted observation centers geometrically positioned to discriminate between candidate bonds adjacent to those sites. Only centers need to be run, each independently.
This mirrors standard practice in other imaging domains: a wide-field scout scan (seismic) followed by targeted high-resolution acquisition (aperture). On quantum hardware, Phase 1 maps directly to a single quantum experiment; Phase 2 selects the minimum number of additional experiments to confirm the bond identity.
The seismic phase requires only one circuit (all centers in a single initialization) and one set of hardware shots for the perturbed device. The baseline is provided by simulation and reused across all candidate perturbations at no additional hardware cost. The aperture phase then tests the hypothesis with directional projections; perturbation signals that survive across independent projections with geometry-consistent timing are physical, while chi-truncation artifacts and shot noise are uncorrelated across runs and average out.
The following describes wavefront triangulation across scans. When multiple independent scans are available, the temporal order in which each scan detects the broken bond carries geometric information beyond the sign flip alone. This enables a triangulation strategy analogous to GPS: each scan provides a time-of-arrival constraint that localizes the bond.
Formally, scan with nearest center detects the broken bond at cycle , the first cycle where for some bond endpoint . This constrains the bond to lie within hops of : a ball on the graph. Across scans, the bond lies in the intersection:
where is the set of bonds within crossing distance of center .
The sign of the earliest detection (which endpoint is RED vs BLUE) further constrains the direction: the center is on the BLUE (wall-face) side of the broken bond. Subsequent scans confirm or deny: a scan detecting the opposite sign at the same endpoints confirms (center is on the other side of the same bond); a scan detecting different endpoints shifts the candidate set. Each confirmation tightens the intersection; each denial eliminates candidates.
In practice on the lattice with 3 scans (corners+center, quadrants, mid-edges): scan 2 detects bond 9 (11↔12) at cycle 1 (center 13, 1 hop away), scan 3 confirms with opposite sign at cycle 2 (center 4, 2 hops away), scan 1 provides weak confirmation at cycle 2 from the far corners. The intersection of the three balls gives a small candidate set containing the true bond.
BFS-guided fault detection algorithm
The detection method reduces to a graph-theoretic algorithm that operates entirely on the observable , without reference to the Hamiltonian, coupling values, or simulation parameters. Let be the coupling graph, the observation centers, and the detection threshold.
For each center and each site , record the baseline arrival time and perturbed arrival time . These are the actual BFS frontier crossings, not predicted hop distances; may exceed if a broken bond delays the perturbed wave.
Both the simulation and the quantum circuit assume an execution order among independent bonds, determined by Trotter color class. We materialize this order for causal determination: assign each site the pair where is its BFS hop depth and is the color class of the bond used to reach it. The total order gives exact sub-cycle causality: means site is strictly upstream of .
For a bond with upstream of , compute the bond transfer coefficient:
When , both waves share the same source amplitude at the BFS frontier and : the signed cross-entropy directly measures the relative change in bond transfer. A broken bond reduces , giving (red shadow) and (blue wall face).
For each site with , examine neighbors with . Declare bond as a candidate if and have opposite signs. Score each candidate by where if , else 1, and .
Repeat steps 1–4 for each scan independently. Rank all candidate bonds by total score across scans and declare the top- as broken.
From our experiments we have identified three recurrent interference signatures, which we name here as a working vocabulary based on observed cases. Each is a consequence of the BFS frontier geometry and the Trotter color class ordering, with no quantum-specific reasoning required. We expect further patterns to emerge as the geometry of faults and center placements varies; the vocabulary below reflects our current observations, not an exhaustive classification. RED-BLUE is the clean single-bond signature: center on one side, sign flip across the bond. BLUE-WHITE-BLUE arises when a bond is sandwiched between two equidistant centers: the shared site is white (interference node) and the outer sites are blue on both sides, making both adjacent bonds identifiable. BLUE-BLUE-RED occurs when two orthogonal broken bonds share an endpoint: the shared site accumulates from two directions, the outer RED identifies one bond, and the adjacent BLUE s identify the second.
Measurability and Detection Thresholds
The detection method requires to be measurable both from the simulation fidelity standpoint and from the hardware shot budget. Two independent conditions must hold simultaneously.
Recall that is a probability: the fraction of Pauli basis operators that have non-identity support at site after Trotter cycles (§ 3). On real hardware the circuit is executed times. A barrier after each Trotter cycle allows a mid-circuit local Pauli weight measurement at site ; each shot returns either the identity at or a non-identity Pauli (, , or ). The three non-identity outcomes are not distinguished: only the binary question “identity or not” is asked, and the non-identity count accumulates as a success count. This binary nature of each shot makes the Binomial model natural: , where the physical probability is the Bernoulli success probability for each of the independent shots. The empirical estimate is then an unbiased estimator of the true .
The estimator has variance , a direct consequence of the Binomial distribution. At the frontier, : the total Pauli weight is conserved () and spreads over a growing shell of sites, so each individual decreases with distance from the center. In this small- regime and the variance simplifies to , giving a shot noise floor . Note also that only sites where contribute to the detection (§ 3), so the relevant regime is . The baseline is provided by simulation and is exact; only the perturbed run is measured on hardware. The condition is that is a reliable estimate of the true , so that the comparison with the theoretical is meaningful. Reliability requires the standard deviation of the estimator to be much smaller than the quantity being estimated: . As grows, shrinks and the estimate converges to the true value. The minimum observable is set by the detection threshold . For to be a reliable estimate at we require , i.e., , which gives:
For , at least shots are required. Table 3 shows the statistical floor as a function of shot budget.
| shots | at | Status |
| noise exceeds threshold | ||
| noise matches threshold (minimum) | ||
| reliable estimate | ||
| well into reliable regime |
Table 3. Estimator noise at the detection threshold . Reliable detection requires .
The tensor network simulation introduces two distinct noise sources that set a floor on trustworthy values. First, truncation: bond singular values below the cutoff are set to zero, so any arising from truncated components is numerical noise, not physical operator spreading. Second, belief propagation is not strictly causal: messages propagate globally in each sweep, producing tiny non-zero at sites outside the Lieb–Robinson lightcone. In a real quantum circuit these sites would never be activated (only gates within the lightcone fire), so this leakage is a simulation artifact with no hardware counterpart. Both effects are below , which we therefore adopt as the simulation floor.
Other simulators with tighter cutoffs could in principle trust smaller values and use a lower . However, a smaller threshold demands proportionally more shots at the frontier: to estimate reliably one needs shots. A simulation floor of would require shots merely to reach reliable estimates at the threshold, a cost that must be weighed against the physical benefit of the finer resolution.
The two conditions converge at the same threshold : the simulation floor sets the minimum trustworthy signal, and from Table 3 reliable hardware estimates at that level require shots. Both conditions are satisfied simultaneously by:
A tighter simulation cutoff enables a smaller and finer resolution, but demands proportionally more shots at the frontier to maintain reliable estimates.
From § 8, the bond gate PTM has diagonal entries . A coupling deviation shifts the rotation angle and attenuates the diagonal entry by to first order. At this means the perturbed weight at site is simply the baseline attenuated:
so the signal is proportional to the baseline weight and scales linearly with the coupling deviation. The attenuation factor depends only on the Trotter step and the perturbation magnitude, not on the site.
Note that this is a stronger requirement than simply estimating reliably: we need the signal to be measurably larger than the estimator noise . Setting gives the minimum detectable coupling deviation. We emphasize that the present work performs detection: identifying the location and existence of a broken bond from wavemap patterns, not estimation of from the numerical values of . Quantitative estimation of the coupling deviation from the measured is a harder and separate problem, left for future work.
For and (genuine wavefront at first arrival), the detection threshold scales as : at shots , at it drops to . To detect the experimental coupling degradation used throughout this work requires shots. This is independent of the minimum shot budget for reliable estimation (): the two conditions address different questions: can we measure at all, versus can we resolve the specific coupling deviation of interest.
Conclusion
A broken bond leaves a causal signature in the operator wavefront: a red-blue sign flip in at its two endpoints, identifiable from the first wavefront crossing and robust across all center placements. Its spatio-temporal gradient reveals the same bond as a physical singularity: a gradient discontinuity with a characteristic diffraction pattern whose orientation encodes the bond direction. Three seismic scans, each firing a different set of centers simultaneously, cover all 84 bonds of the lattice in 6 simulations. Both detection methods achieve 20/20 single-bond and 9/10 two-bond identification, with failures on complementary trials. The per-bond detection rate remains above 90% for up to 8 simultaneous faults in a single pass. An iterative protocol (declare, recalibrate, re-scan) progressively resolves multi-bond cases, with as the natural stopping criterion.
The measurability analysis shows that shots suffice for reliable estimation of at the detection threshold; resolving the specific coupling deviation requires shots, a stronger condition that governs hardware deployment.
The method is a direct sequitur to the Lindblad MPF framework. Both use the same observable , the same circuits, and the same hardware protocol; MPF infers noise from a known Hamiltonian, while this work infers coupling geometry from a known noise baseline. Two inferences from one measurement, in opposite directions.
Aperture Synthesis with Voronoi Center Placement
An alternative to the seismic approach runs each center independently, in direct analogy with aperture synthesis in radio interferometry. Each center at a different graph position plays the role of one antenna at a different baseline; encodes which side of the broken bond site is on (sign) and the signal strength (magnitude).
Centers are placed with the constraint that no two share the same row or column in the lattice. This guarantees every bond has a strictly unique nearest center (no Voronoi ties): the perpendicular bisector between any two centers never passes through a bond, so one center always dominates the signal at every bond. For on the lattice (one center per row, one per column, a permutation placement), exhaustive search over permutations minimizing gives centers with zero ties across all 84 bonds, maximum cycles, and balanced Voronoi regions of 10–16 bonds each.
Ensemble summation fails for bonds sandwiched between centers on opposite sides, where RED and BLUE contributions partially cancel. A winner-takes-all rule avoids this: for each center compute the signed contrast across every bond,
and declare the globally winning center-bond pair . The winning center is necessarily on the BLUE side of the broken bond; each center is a fully independent observer and the one with the strongest signal wins. On 10 trials (), the adaptive seismic-then-aperture protocol achieves 10/10 correct identifications: 7 bonds identified by seismic alone (2 simulations), 3 by aperture (14 simulations), at an average cost of simulations per bond.
The seismic 3-scan approach in § 6 supersedes this method for the multi-bond case: it achieves the same single-bond accuracy at lower cost, and extends naturally to multiple simultaneous faults without requiring the no-shared-row-or-column placement constraint.
References
- [1]Tudor Giurgica-Tiron, Yousef Hindy, Ryan LaRose, Andrea Mari, and William J. Zeng. Digital zero noise extrapolation for quantum error mitigation. 2020 IEEE International Conference on Quantum Computing and Engineering (QCE), pages 306–316, 2020.
- [2]Jia-Wei Hong and H. T. Kung. I/O complexity: The red-blue pebble game. In Proceedings of the Thirteenth Annual ACM Symposium on Theory of Computing (STOC ’81), pages 326–333, 1981.DOI
- [3]Andrew Robertson et al. Lindblad multiproduct formulas, 2026. Manuscript.arxiv.org/abs/2609.05024
- [4]J. P. Roth. Diagnosis of automata failures: A calculus and a method. IBM Journal of Research and Development, 10(4):278–291, 1966.DOI