Modular quantum computers wire many small processor chips ("cores") together with quantum links that share entanglement between chips. The compilers that decide which qubit lives on which core usually minimize how often the links are used, and they assume every link is equally good. Real links differ, and they will differ more as systems grow. We propose to borrow a tool from genomics. Biologists face the same situation when a bulk tissue measurement averages over millions of cells and hides the rare cells that matter. Their response is deconvolution: they infer each component's contribution from mixed measurements, using a known recipe of how the components combine.
In a modular quantum computer that recipe is known exactly. The compiler records how many times each circuit uses each link and each core. Every circuit run therefore acts as a "bulk sample" of the hardware. From the pass/fail counts of ordinary workloads we can estimate a per-link error rate, flag the few bad links that averages hide, and feed those estimates back into the mapping compiler. In two sentences: aggregate fidelity numbers in distributed quantum hardware mask a minority of bad links, much as bulk sequencing masks rare cell populations. Because the compiler knows the "mixing matrix", reference-based deconvolution of routine circuit outcomes can recover per-link quality cheaply and make heterogeneity-aware mapping possible.
Core Claim
Treat circuit $c$ as succeeding only if every link and core operation it uses succeeds, independently. Then its success probability $P_c$ satisfies
$$ -\log P_c \;=\; \sum_{\ell} W_{c\ell}\,\theta_\ell, \qquad \theta_\ell = -\log(1-e_\ell) \ge 0 . $$
Here $W_{c\ell}$ is the number of times circuit $c$ uses component $\ell$, which is read from the compiled schedule. $e_\ell$ is the unknown per-use error rate of that component, and $\theta_\ell$ is the same rate rewritten in log form. This is the bulk-deconvolution model $Y = W H$ from genomics, with three changes. The "signature matrix" $W$ is exact rather than taken from a reference atlas. The unknowns $\theta$ play the role of cell-type proportions. The observations are binomial shot counts rather than expression values.
We claim three things. First, per-link heterogeneity large enough to change compiler decisions can be recovered from production-style workloads, without dedicated per-link benchmarking. Second, the hard part is the same one biologists hit: links that are always co-used cannot be told apart. This mirrors the "purity trap" in tumor deconvolution, where cancer cells that resemble their tissue of origin are mistaken for it. Third, making the mapper aware of the estimated $\theta$ improves end-to-end success on distributed circuits. This matters most for designs that assume ideal links.
Where This Sits in Prior Work
Several papers motivate the gap. A Scalable Distributed Quantum Optimization Framework via Factor Graph Paradigm splits optimization problems across worker processors joined by shared entangled pairs. Its authors state that their proofs assume ideal, deterministic entanglement, while real inter-chip entanglement is probabilistic and lossy. Assessing the Role of Communication in Modular Multi-Core Quantum Systems builds the qcomm simulator and shows that halving teleportation counts with TeleSABRE still leaves classical communication at up to 40% of runtime in 10×10 core arrays. That work counts communications and does not weight them by link quality.
Three compiler papers supply the mapping machinery we modify, and all of them optimize link usage under homogeneous links:
- Mapping quantum circuits to modular architectures with QUBO casts core allocation as a binary quadratic problem with a graph-Laplacian penalty on cross-core gates.
- Circuit Partitioning for Multi-Core Quantum Architectures with Deep Reinforcement Learning partitions circuits with reinforcement learning.
- A Genetic Approach to Minimising Gate and Qubit Teleportations for Multi-Processor Quantum Circuit Distribution minimizes teleportations with genetic search.
We replace their uniform link weights with estimated weights.
Qubit noise deconvolution inverts a known single-qubit noise channel to correct measured observables on a Rigetti Aspen-9 chip. It shares a word with our work but solves the opposite problem. It needs the noise model and corrects outcomes, whereas we estimate the noise model from outcomes. Our estimates could supply its inputs.
The cellular-heterogeneity literature supplies the method and warns about how it fails. DECONbench shows that simple factorization methods are hard to beat: NMF-based r_WNM reached a mean absolute error of 0.024. It also shows that elaborate fusion added little, which tells us to test simple linear baselines seriously. Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures documents two failure modes we expect to meet. One is systematic mis-attribution between similar components as purity rises. The other is the trade-off in which hspe has the fewest false positives but many false negatives, which makes it blind to rare populations. For us, the rare population is the handful of bad links. Count Bridges enable Modeling and Deconvolving Transcriptomic Data argues that integer counts should be modeled on their own terms rather than with Gaussian surrogates, and it deconvolves with expectation-maximization under the Energy Score. We adopt the count-likelihood principle in a far simpler form, a binomial likelihood. We do not need its generative bridge because our mixing matrix is known.
Research Program
We build a simulated heterogeneous modular machine. Link and core error rates are drawn from a heavy-tailed distribution in which most links sit near a median and a small fraction are several times worse. Link physics is grounded in the cavity-mediated transfer simulator of Accurate and Scalable Simulation of Cavity-Based Networks in Modular Quantum Architectures, and scheduling and communication timing come from qcomm. Rates also drift slowly over time.
The estimator has three parts:
- Likelihood. Binomial on success counts, with the log-linear link above.
- Prior. Sparsity on each link's deviation from the global median $\theta$, which is the ingredient aimed at catching rare bad links.
- Identifiability check. Before estimating, we compute which combinations of $\theta$ the workload's $W$ can resolve. Unresolvable groups are reported as groups rather than guessed.
We also study probe augmentation, the analog of designing mixtures in biology. Given a production workload, we choose a small number of short extra circuits that break collinearity in $W$, and we measure how many extra shots this saves compared with benchmarking each link directly.
Finally, we modify the QUBO mapper so that its cross-core penalty is weighted by the estimated $\theta_\ell$ instead of being uniform. We then check whether this beats both the uniform mapper and a mapper given the true $\theta$ (the oracle).
The independence assumption behind the log-linear model will be wrong in places, for example through crosstalk and correlated drift. We quantify this by simulating correlated errors and measuring how far the estimates degrade.
Baselines
The naive reference is the uniform-link assumption, which every cited mapper uses implicitly: all links get the average fidelity estimated from aggregate success. Detection of bad links is at chance, and mapping is the unmodified QUBO mapper. For mapping quality, the strongest reported number from that line is the one in Mapping quantum circuits to modular architectures with QUBO: an 87% first-attempt mapping success rate over 407 circuits of 50–100 qubits, with as few as 0.78 inter-core communications per two-qubit gate. That number measures communication count, not end-to-end fidelity. Beating it here means a higher simulated circuit success probability at no more than a small increase in communications per gate, reported across the same 407 instances and the same annealer budget.
The strongest estimation competitor is direct per-link benchmarking, which runs dedicated entanglement-fidelity experiments on every link. It is the gold standard for accuracy. We give it exactly the same total shot budget as our method, and our method must match its estimation error or its bad-link recall while using fewer dedicated shots. No reported result on Lacuna was found for per-link fidelity estimation from aggregate circuit outcomes. No reported per-link benchmarking cost for modular systems was found either, so both arms are measured in our own simulator.
The key ablation is non-negative least squares on log success rates. This is the plain linear deconvolution analog of DECONbench's NMF baselines, run with the same $W$, the same data, and hyperparameters tuned on the same held-out drift windows. It isolates whether the count likelihood and the sparse rare-link prior add anything beyond linear unmixing. A second ablation removes probe augmentation. A third gives the mapper the true $\theta$ (oracle), which bounds the achievable gain.
Why not just…?
Why not just benchmark every link directly? For small machines this may be the right answer, and we will say so if the results show it. The case against it is scaling and drift. Links grow with core count, and each dedicated benchmark takes machine time away from useful work. Aggregate workload outcomes, by contrast, arrive for free and continuously. The shot-matched comparison in the baselines decides this. If direct benchmarking wins at every machine size up to 10×10 cores, the proposal reduces to a monitoring tool for detecting drift between calibrations, and we will report it that way.
Why not just use plain linear unmixing, since DECONbench found simple methods hard to beat? This is a real threat. DECONbench found that elaborate combination gave only a 2.12% error reduction. If least squares on log rates recovers bad links as well as our estimator, then the contribution lies in the observation and the identifiability analysis, not the estimator, and we will present it that way. We expect the gap to show up specifically in bad-link recall at low shot counts. That is where Gaussian surrogates blur rare signals, just as hspe's low-false-positive behavior came with poor recall of rare populations in the breast cancer benchmark.
Why not just make mappers robust by minimizing total communication, which already avoids most links? Minimizing link usage and avoiding bad links are different objectives. Assessing the Role of Communication in Modular Multi-Core Quantum Systems shows that even after TeleSABRE halves teleportations, communication remains a large share of execution. A single bad link on a route that is used unavoidably can dominate failure. The oracle ablation settles this. If giving the mapper the true $\theta$ barely beats the uniform mapper, heterogeneity does not matter for mapping and the downstream claim fails.
Proposed Experiments and Decisive Checks
We compile the 407-circuit suite from Mapping quantum circuits to modular architectures with QUBO and MQTBench circuits used with qcomm onto simulated 2×2 through 10×10 core arrays with heavy-tailed link noise. We then run shot budgets ranging from a few hundred to tens of thousands per circuit. Estimation is scored by relative error in $\theta$ and by recall and precision on the worst 5% of links. Mapping is scored by simulated end-to-end success probability, inter-core communications per two-qubit gate, and classical solve time.
The first decisive check is the masking claim itself: under realistic tails, does the uniform-link assumption mis-rank mappings often enough to matter? We measure this as the fraction of instances where the oracle mapper beats the uniform mapper by a meaningful success margin. The second is identifiability. We predict that estimation error concentrates on link groups flagged as collinear in $W$, and that a handful of probe circuits removes most of it. If error does not follow our identifiability diagnosis, the model is wrong. The third is robustness to correlated errors: we inject crosstalk-like correlations and report how far bad-link recall degrades. The fourth tests the factor-graph setting of A Scalable Distributed Quantum Optimization Framework via Factor Graph Paradigm. We rerun its coordinator–worker entanglement pattern on the heterogeneous links and ask how much of its oracle-query advantage survives once link quality is estimated rather than assumed ideal.
References
- A Scalable Distributed Quantum Optimization Framework via Factor Graph Paradigm
- Assessing the Role of Communication in Modular Multi-Core Quantum Systems
- Mapping Quantum Circuits to Modular Architectures with QUBO
- Circuit Partitioning for Multi-Core Quantum Architectures with Deep Reinforcement Learning
- A Genetic Approach to Minimising Gate and Qubit Teleportations for Multi-Processor Quantum Circuit Distribution
- Qubit noise deconvolution
- DECONbench: a benchmarking platform dedicated to deconvolution methods for tumor heterogeneity quantification
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Accurate and Scalable Simulation of Cavity-Based Networks in Modular Quantum Architectures