A quantum computer can store a probability distribution over a million grid points in twenty qubits, because each added qubit doubles the size of the grid. This "logarithmic encoding" is the main resource argument for simulating particles on fine spatial grids. The catch is readout. Each run of the circuit returns only one random grid point, so recovering the whole distribution takes about as many runs ("shots") as there are grid points. That cancels the memory savings. Rare-event physics needs something much smaller, though. A search for neutrinoless double-beta decay or a narrow collider resonance has only a handful of observed events, and the discovery test needs the theory's predicted density only near those few events, and only at the resolution the detector can resolve. This proposal argues that this regime is where the readout wall is lowest, and it turns that argument into an explicit shot budget.
The core claim is concrete. Suppose a log-encoded quantum simulation supplies the signal prediction and the data are $k$ Poisson-distributed events. Then a discovery test can be run by sampling the simulator while ignoring its least-significant qubits, so that each event is compared against a coarse grid cell rather than a single point. The shot cost is then set by how much predicted signal probability lands in the cells around the events. It does not grow with the total grid size $2^n$. In a binary encoding, choosing how many qubits to read is the same choice as choosing a bin width, so the physicist's unbinned-versus-binned question becomes a shots-versus-power question. We will derive the optimal number of qubits to read, build a likelihood that accounts for shot noise and does not manufacture false discoveries, and test whether hardware noise distorts the tail probabilities at the five-standard-deviation ($5\sigma$) discovery threshold.
Why this gap is real
Logarithmic encodings are well developed for simulation and optimization. First-quantized simulation stores the coordinates of $\eta$ particles in $O(\eta \log N)$ qubits for an $N$-point grid (Babbush et al.). For rapidly oscillating wave packets in the semi-classical regime, the qubit count grows only as $\log(1/\varepsilon)$ in the oscillation scale $\varepsilon$ (Jin et al.). Optimization work packs a million variables into twenty qubits (Venkatesh et al.). Zhou et al. and Kossaifi et al. show similar compression for quadratic programs and semidefinite programs. These papers supply the encoding we reuse. They also motivate the problem: Venkatesh et al. state plainly that decoding the full distribution needs shots exponential in the qubit count. None of them asks what a downstream statistical test actually needs to read.
On the physics side, unbinned extended likelihood fits evaluate the model density at each event's exact coordinates, and they are the standard way to get the most out of scarce data (Lee, Bergfeld et al., Achasov et al.). Look-elsewhere corrections can erase apparent excesses: Abbasi et al. found that cosmic-ray hotspots fell below $5\sigma$ once trial factors were included. Low-significance signals often vanish with more data (Bai et al.). These results set the standard any quantum-assisted test must meet: calibrated tail probabilities, not just a good average classification accuracy.
The closest crossover work is Chen, Delgado et al.. They classify rare Higgs events using 10 qubits loaded with 10 principal-component features, one feature per qubit. They report areas under the ROC curve of 0.83 for their quantum variational classifier on a simulator versus 0.80 for a boosted decision tree on $H\to\mu\mu$, using 100 training events. That work uses no logarithmic encoding, computes no discovery significance, and treats the quantum device as a classifier rather than as the source of the physical prediction. The Delgado review likewise treats classification separately from likelihood inference. Srikanth et al. prove fast convergence rates for maximum likelihood over quantum density operators. Their setting is idealized independent tokens with no Poisson counts or circuits, so we use it only as theoretical context.
Core Claim
Let the simulator prepare an $n$-qubit state whose computational-basis probabilities $p_s(x)$ are the predicted signal density on a $2^n$-cell grid. Measure only the top $\ell \le n$ bits of each coordinate. This yields samples from the coarse-grained density $P_\ell(c)$ over $2^\ell$ cells, and the unread qubits can be discarded or reset mid-circuit. The circuit-level reuse of reset qubits follows Wawrzyniak et al..
Now take the $k$ observed events $x_1,\dots,x_k$, with $c_\ell(x_i)$ the coarse cell containing event $i$. Given $S$ shots, the count $m_i$ of shots that land in that cell gives an unbiased estimate $\hat P_\ell(c_\ell(x_i)) = m_i/S$, whose relative variance is about $1/(S\,P_\ell)$. If each event's signal weight should carry relative error at most $\delta$, the required budget is approximately
$$ S^*(\ell) \approx \frac{1}{\delta \,\min_i P_\ell\big(c_\ell(x_i)\big)}, $$
where $P_\ell$ is the predicted signal probability of a coarse cell. This expression depends on how concentrated the signal is at resolution $\ell$, not on $n$. Reading more bits shrinks the cells, which lowers $P_\ell$ and raises $S^*$. But once the cell width falls below the larger of the detector resolution and the signal's intrinsic width, finer cells add almost no statistical power. The predicted optimum is therefore $\ell^* \approx \log_2(\text{range}/\text{resolvable width})$, and any qubits beyond that serve the simulation's accuracy, not the readout. For semi-classical wave packets, $\ell^*$ tracks $\log(1/\varepsilon)$ until detector smearing takes over. This is a direct and testable link between the qubit-scaling result of Jin et al. and the statistics of the search.
Research Program
The first component is the statistical estimator. Plugging the noisy estimates $\hat P_\ell$ directly into the extended likelihood is dangerous. Under the background-only hypothesis, upward shot fluctuations at event locations look exactly like signal, so they inflate the test statistic $q_0 = -2\ln(\mathcal L_0/\mathcal L_{\max})$ and produce false discoveries. This is the same problem as fitting with templates built from finite simulation samples. We will reuse the standard remedy: treat each shot count $m_i$ as an auxiliary Poisson measurement with its own nuisance parameter, and profile over it. We call this the shot-aware likelihood. We will also test the simpler alternative of sample splitting, where independent shot batches are used for fitting and for testing. The background density is assumed to be classically known, which keeps the mixture likelihood finite.
The second component is the analysis of compound noise. We will derive how the effective significance combines Poisson fluctuations in the data with shot noise as a function of $k$, $S$, and $\ell$. The expectation is that each shot-noise variance term enters with a weight proportional to the corresponding event's signal fraction. If so, sparse data needs fewer total shots per unit of significance lost than dense data does. We will then check this by simulation.
The third component is hardware noise. Depolarizing errors and readout errors redistribute probability across cells. Because $P_\ell$ is small at the events, a small fraction of misattributed probability is a large relative bias, and it shifts the null distribution of $q_0$. We will measure this effect with a background-only control circuit, analogous to a sideband, and correct for it there.
Baselines
The naive reference point is a cut-and-count analysis. It counts events in a fixed window around the predicted signal peak and computes the Poisson significance against the known background. This is what one does without any density information, and any method that cannot beat it has no reason to exist.
The strongest competitor is the classical unbinned extended likelihood evaluated with the exact predicted density, in the style of Lee and Bergfeld et al.. On our testbeds the density can be computed classically by state-vector simulation, so this baseline is an oracle and an upper bound on power. The quantum procedure cannot beat it. Success means coming within 5% of the oracle's median significance at a shot budget near the predicted $S^*$, with that budget flat in $n$ once $n > \ell^*$. No reported discovery-significance result was found on Lacuna for any quantum-assisted method on these or comparable benchmarks. The nearest reported numbers are classification scores (ROC area 0.80 for the boosted tree versus 0.83 for the simulated quantum classifier on $H\to\mu\mu$, Chen et al.) and EXO-200's "~25% improvement in sensitivity" from a classical deep network (EXO-200). Neither is directly comparable, and we cite them only as context.
The full-readout baseline spends the same shot budget reconstructing the full $2^n$-cell histogram and then runs a classical binned likelihood. This is the approach implied by Venkatesh et al.'s measurement wall. We expect its required shots to scale as $2^n$, and beating it means matching its significance with orders of magnitude fewer shots at $n \ge 16$.
The classifier-then-count baseline follows Chen et al. It trains a 10-qubit angle-encoded variational classifier, applies a threshold, and counts events. It gets the same training data and the same total circuit executions.
Two ablations isolate the proposal's ingredients. The first replaces the shot-aware likelihood with the plug-in likelihood at an identical budget, to isolate the false-discovery correction. The second fixes $\ell = n$ (read every bit) instead of $\ell^*$, to isolate the benefit of partial readout. All methods share the same event samples, background model, and pseudo-experiment seeds, and every tunable choice, including $\ell$, gets the same tuning budget on a separate validation set.
Why not just…?
Why not just run a classical unbinned likelihood? If the predicted density can be computed classically, that is exactly what one should do, and our oracle baseline does it. The proposal matters only when the prediction comes from a log-encoded quantum simulation whose densities cannot be tabulated classically, as in the first-quantized and semi-classical regimes of Babbush et al. and Jin et al. We deliberately use classically checkable instances so that the gap to the oracle can be measured exactly rather than assumed.
Why not just take enough shots to read out the whole distribution? Covering every cell takes on the order of $2^n$ shots, as Venkatesh et al. note. Experiment 1 measures this scaling directly against partial readout as $n$ grows from 10 to 22 qubits.
Why not use a swap test or Hadamard test to compute the overlap between a data state and the model state? This was our first design, and it fails. Rare events do make the data state sparse and cheap to prepare. But the swap-test fidelity $F = (\sum_i \sqrt{p_s(x_i)})^2/k$ is never larger than $P = \sum_i p_s(x_i)$, by the Cauchy–Schwarz inequality. Distinguishing signal from background needs roughly $1/F$ shots, compared with roughly $1/P$ for direct sampling, so the coherent overlap is no better than sampling and is usually worse. Quantum amplitude estimation could cut the shot cost quadratically, from $1/\epsilon^2$ to $1/\epsilon$, but it requires deep coherent circuits. We treat it only as a fault-tolerant extension and compute the circuit depth at which it would overtake sampling.
Why not treat this as the familiar finite-simulation-statistics problem? That is largely what it is, and we reuse that machinery rather than reinvent it. Three things are new: the bin width is set by which qubits are measured, which gives a resolution-dependent budget $S^*(\ell)$ with no classical analogue; the cost is tied to the qubit scaling of the simulation; and the template error comes from correlated hardware noise rather than independent statistical noise. If the shot-aware likelihood and sample splitting perform identically, we will report the simpler method.
Why not use a quantum classifier as Chen et al. do? Thresholding a classifier output throws away the per-event density information that unbinned fits rely on. It also yields a ROC area rather than a calibrated significance. The classifier-then-count baseline tests this directly.
Proposed Experiments
Experiment 1 is the semi-classical testbed. We simulate the time-dependent Schrödinger equation with state vectors on 1D and 2D grids of $n = 10$ to $22$ qubits, scanning the oscillation scale $\varepsilon$ in the regime of Jin et al.. The final density is the signal prediction, and its interference fringes give fine-scale structure. We convolve it with a Gaussian detector smearing of width $\sigma_{\text{det}}$, inject $k \in \{3, 5, 10, 30, 100\}$ signal events over a smooth Poisson background, and sweep $\ell$ and $S$. The decisive check has three parts. Median significance should plateau at $\ell^* \approx \log_2(\text{range}/\max(\varepsilon, \sigma_{\text{det}}))$. The required budget $S^*$ should match the formula within a factor of two and stay flat in $n$. Full readout should require shots growing as $2^n$.
Experiment 2 calibrates the null distribution. We run $10^6$ background-only pseudo-experiments, regenerating shot noise for each from the estimated coarse density, and compare the false-positive rate at $3\sigma$ and $4\sigma$ for the plug-in and shot-aware likelihoods. We then extrapolate to $5\sigma$ using the asymptotic distribution of $q_0$, checked with importance-sampled pseudo-experiments. Success means the shot-aware likelihood stays within statistical error of the nominal rate while the plug-in version measurably exceeds it.
Experiment 3 stress-tests realistic kinematics using the R&D dataset from the LHC Olympics 2020. We map dijet mass against a substructure variable onto a $2^7 \times 2^7$ grid and load the signal histogram with generic state preparation, in simulation. This tests the bit-selection trade-off and calibration on real shapes, not any quantum advantage. Background-only injections in the style of the challenge's Black Box 2 check for spurious discoveries.
Experiment 4 measures hardware noise. On 6- to 10-qubit instances, run on hardware where available and otherwise on noisy simulators with measured readout-error models, we quantify the shift in the null distribution of $q_0$ and test whether correcting from the background-only control circuit restores nominal false-positive rates.
The proposal is falsified if, in Experiment 1, partial readout cannot reach within 5% of the oracle's significance without shot budgets that grow with $n$, or if, in Experiment 2, no likelihood correction removes the excess false discoveries.
References
- Quantum simulation of exact electron dynamics can be more efficient than classical mean-field methods
- Quantum simulation in the semi-classical regime
- Qubit-Efficient Variational Quantum Algorithms for Image Segmentation
- A hybrid algorithm for quadratically constrained quadratic optimization problems
- Quantum Goemans-Williamson Algorithm with the Hadamard Test and Approximate Amplitude Constraints
- Update of the search for the neutrinoless decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mrow><mml:mover><mml:mrow><mml:mi>τ</mml:mi></mml:mrow><mml:mrow><mml:mo>→</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:mrow><mml:mi>μ</mml:mi><mml:mi>γ</mml:mi></mml:math>
- Observation of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">B</mml:mi></mml:mrow><mml:mrow><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:mrow><mml:mo>→</mml:mo><mml:mrow><mml:msup><mml:mrow><mml:mi mathvariant="italic">ω</mml:mi><mml:mi mathvariant="italic">K</mml:mi></mml:mrow><mml:mrow><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:mrow></mml:math>and Search for Related<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi mathvariant="italic">B</mml:mi></mml:math>Decays Modes
- Observation of a Neutral Structure near the D(D)over-bar* Mass Threshold in e(+)e(-) -> (D(D)over-bar*)(0)pi(0) at root s=4.226 and 4.257 GeV
- INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT
- Branching fractions for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>ψ</mml:mi><mml:mn/><mml:mo>(</mml:mo><mml:mn>2</mml:mn><mml:mi>S</mml:mi><mml:mo>)</mml:mo><mml:mn/><mml:mo>→</mml:mo><mml:mi>γ</mml:mi><mml:mrow><mml:msup><mml:mrow><mml:mi>η</mml:mi></mml:mrow><mml:mrow><mml:mo>′</mml:mo></mml:mrow></mml:msup></mml:mrow></mml:math>and γη
- Application of Quantum Machine Learning to High Energy Physics Analysis at LHC using IBM Quantum Computer Simulators and IBM Quantum Computer Hardware
- Quantum computing for data analysis in high energy physics
- Quantum Maximum Likelihood Prediction via Hilbert Space Embeddings
- Dynamic Circuits for the Quantum Lattice-Boltzmann Method
- Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset
- The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics