Interferometric Quantum Polynomial Chaos Expansion as a Generative Model for Calorimeter Shower Simulation

AI summary

Overview

Research area: Quantum machine learning for scientific simulation — specifically quantum generative models applied to particle-physics detector simulation (calorimeter showers), bridging quantum computing, polynomial chaos theory, and high-energy physics software.

Technical level: Advanced. The paper assumes familiarity with variational quantum circuits, generative modeling, and statistical distance measures, though the central design ideas are explained geometrically and can be followed with an intermediate background.

Scope (one sentence): The paper introduces the quantum polynomial chaos expansion (QPCE), a generative model in which a single re-uploading interferometric circuit with zero fitted classical parameters is the entire model, trains it on Geant4 calorimeter shower data, verifies its dependence structure with a built-in exact null, and executes the identical circuit on the ibm_fez superconducting processor.

What This Paper Is About

Calorimeter simulation is one of the heaviest computing loads in high-energy physics, and the High-Luminosity LHC is expected to produce on the order of 10^10 events per year, making fast generative surrogates a practical necessity. Existing quantum approaches to this task are mostly hybrid: a quantum device supplies features and a trained classical network reconstructs the image, which means the classical part absorbs most of the fitted capacity and it becomes impossible to say what the quantum resource actually contributed. The paper's goal is to remove that ambiguity by making the circuit itself the generator — producing both the per-cell densities and the correlations between cells from one circuit, with exactly zero fitted classical parameters — and then to hold that construction to account with controls that are settings of the trained model rather than imported baselines.

Key Contributions

1. A fully quantum chaos expansion (QPCE) with no fitted classical parameters. Each of the n calorimeter cells gets one qubit; cell intensity is read from the Pauli-Z expectation via a fixed affine map Y_i = m_i + s_i⟨Z_i⟩ using 2n stored constants computed once from the training range. Draw n private uniform germs plus one shared germ, re-upload them at the head of each of L blocks, and depth L becomes the order of the chaos expansion the circuit realizes — verified to machine precision.

2. A verification protocol that is a parameter setting of the model. Setting all R_ZZ coupler phases to zero and freezing the shared germ at its median makes the generated cells exactly independent for any remaining angles (Proposition 1). The null is therefore a limit point of the trained family, not an external baseline, and attribution of learned correlations to individual gates becomes a measurement. Choosing between two measurement bases shot by shot sharpens this into certification: the trained model violates the Bell bound obeyed by every classical generative model with local response, whatever its size.

3. A hardware-native circuit and an executed deployment. The dense ring-plus-skip-2 coupler graph C_8(1,2) (|E| = 16) embeds in no present device topology, and routing it inflates two-qubit depth by a factor of 6. The authors instead use a seven-edge path native to the heavy-hex lattice with zero SWAPs, edge-coloured to the chromatic-index floor: 42 fractional two-qubit phase pulses in 12 two-qubit layers and about 1 μs of circuit time. Executed unchanged on ibm_fez, it reaches D = 0.873 [0.674, 0.964] and mean |ρ_S| = 0.5246, against 0.5549 in noiseless simulation, with the accuracy loss predicted analytically in advance.

4. A no-go theorem bounding the whole readout class. The paper proves a no-go theorem for the tail dependence of every smooth generator read out through expectation values (Proposition 4), motivated by a measured target tail asymmetry of λ_L − λ_U = 0.097 at q = 0.05 with disjoint bootstrap intervals, and it identifies the circuit primitive that removes this limit.

Main Findings

• Depth is the truncation order, measured not assumed. Encoding the latent noise only once caps every expectation value at trigonometric degree one in each germ at any depth (a single-harmonic fit leaves relative residual 8×10^-16 at L = 2 and 4×10^-15 at L = 6). Re-uploading the germs per block makes ⟨Z_i⟩ an exact degree-L trigonometric chaos expansion; the degree-L fit residual is ≤ 2×10^-15 while the degree-(L−1) residual is O(1) for L = 2, 3, 4. The exponentially many coefficients are not free but coherent functions of O(nL) angles.

• The model carries its own controls, and they separate two dependence channels. With couplers off and the shared wire frozen, the deployed checkpoint reaches max|ρ_S| = 0.0188 against a sampling ceiling of 0.0236 — i.e., independence at the noise floor. The full model reaches mean |ρ_S| = 0.5549 and a dependence score D = 0.895 (one when generated rank structure matches the data, zero at independence). With couplers removed but the shared wire released, the common-factor channel alone carries mean |ρ_S| = 0.112.

• The shared germ wire targets a real, measured structure in the data. The dataset's rank-correlation matrix has one eigenvalue carrying 64.1% of the dependence, with eigenvector (−0.33, −0.40, −0.42, −0.38, −0.16, +0.28, +0.40, +0.38) — a global front/back sign flip interpreted as shower-start depth. A rank-one reconstruction from that mode alone reproduces all 28 off-diagonal correlations at correlation 0.970. The circuit adds 2n single-qubit rotation coefficients and zero two-qubit gates, zero SWAPs and zero two-qubit depth.

• Training is exact and cheap in gradient cost. The loss is the energy distance, equal to twice the squared MMD with the parameter-free distance kernel (confirmed numerically to 1.9×10^-14), minimized by L-BFGS with adjoint gradients costing three state evolutions irrespective of the number of parameters; adjoint gradients agree with the parameter-shift rule to 10^-14 at 28× lower cost. Because the frozen-batch objective keeps falling after the model stops improving, validation uses fresh germs with early stopping and the best-validation iterate.

• No latent collapse in the converged model. The per-cell W_1 is 0.0039, and fitted private amplitudes remain O(0.3) to O(0.6) on seven of the eight cells. The full energy distance, rather than a correlation score, is what neutralizes the collapse attractor.

• The dense logical graph is undeployable, not merely suboptimal. An exhaustive subgraph-embedding search finds zero embeddings of C_8(1,2) on the heavy-hex lattice and zero on the square lattice. Gate-level compilation at L = 4 gives 316 two-qubit pulses in 98 two-qubit layers against a commuting-block depth floor of χ′(C_8(1,2))·L = 16, a routed circuit fidelity of 0.12, and a shot budget for full certification exceeding 10^8.

• The ansatz is not an IQP circuit, and it is classically simulable at the operating point. The conjugated operator H^⊗n U H^⊗n carries off-diagonal weight 0.90 for the deployed form at L = 2, 0.40 even with the walls pinned at π/2, and 0.000 for a true IQP control. At n = 8, L = 6 the causal cone covers the register and the model lives in a 2^8-dimensional statevector, with light-cone cost O(n 2^{4L+1}); the paper makes no hardness claim in either direction and treats simulability as a feature (exact gradients, exact verification, byte-level agreement with the reference).

• The tail asymmetry exposes a class-wide limit. The target's λ_L − λ_U = 0.097 at q = 0.05 has disjoint bootstrap intervals, a quantity every elliptical family fixes at zero identically. The circuit reproduces 0.060 of it at the same quantile, and Proposition 4 characterizes the asymptotic tail behaviour of the entire smooth expectation-value class.

Methodology in Plain English

The authors start from classical polynomial chaos expansion, which writes a random output as a sum of orthogonal polynomials of independent input random variables and then fits the coefficients. They invert it: the input randomness stays, entering the circuit as rotation angles, but nothing is expanded on paper and nothing is solved for afterwards. Each calorimeter cell is one qubit, and the only input to the circuit is freshly drawn uniform noise — no data is ever encoded into the quantum register.

Two design choices do the heavy lifting. First, real calorimeter cell intensities are skewed, and a circuit that injects noise only once can only produce a very coarse (degree-one) response in that noise no matter how deep it is. So the noise is re-injected at the start of every block, which makes the measured expectation a trigonometric polynomial whose degree equals the number of blocks; depth literally buys model expressivity. Second, because the data show one dominant global mode, one extra noise wire is shared by every qubit with a trainable amplitude, so the collective behaviour of the shower is delivered directly instead of being transported cell-to-cell through entanglers.

Training has no classical decoder in the loop. The training showers enter only through a distribution-level loss (twice the squared MMD, with the parameter-free distance kernel), and the optimiser updates gate angles with exact adjoint gradients. At the end of every fit the authors re-run the same circuit twice more at modified settings: once with the entanglers switched off and the shared wire frozen, which by construction produces exactly independent cells and serves as the null; once with entanglers off but the shared wire free, which isolates the shared-latent channel. Finally, the same circuit is compiled onto a specific device — chosen by building the coupler graph out of the graph the hardware already has, so no routing or SWAPs are needed — and run there, with the shot-noise-induced degradation of each correlation predicted analytically before submission and readout calibration measured by separate circuits that never touch the model.

Why This Matters

The work addresses a credibility problem as much as a performance one. Among papers claiming to outperform a classical method, only 4% report a negative result, and removing entanglement from quantum models often leaves performance unchanged. Because a hybrid model cannot supply a "quantum resource switched off" control — switching off the circuit leaves a trained network that still fits the data — the field has lacked a clean way to attribute results. QPCE makes the control an exact parameter setting and then escalates attribution into Bell-inequality certification, which is a methodological template independent of calorimeters.

Real-world applications:

• Fast calorimeter surrogates for the High-Luminosity LHC, where O(10^10) events per year make full detector simulation a bottleneck.
• Benchmarking and certification tooling for quantum generative models, usable wherever claims of quantum-generated structure need an internal null rather than an external baseline.
• Stochastic field simulation in other domains (transport and detector response modeling, imaging detectors, material and sensor simulation) where skewed marginals and strong global correlations coexist.
• Hardware-aware circuit design for near-term devices, since the embedding analysis gives a concrete rule (build the coupler graph from the device graph) and quantifies what violating it costs.

Industry relevance: the paper quantifies routing overhead, fidelity collapse, and shot budgets for a realistic device — the kind of pre-submission error accounting that industrial quantum deployments require. The finding that a dense logical graph gives 0.12 circuit fidelity and a shot budget above 10^8 is directly actionable for anyone planning production workloads on superconducting hardware.

Future Directions

• Scaling beyond the exactly simulable regime. At n = 8, L = 6 the model is classically tractable, and the light-cone cost O(n 2^{4L+1}) is polynomial in n but exponential in L. Whether larger n or deeper L yields anything a classical surrogate cannot match is left explicitly open, and the paper makes no hardness claim in either direction.
• Removing the tail-dependence limit. Proposition 4 rules out correct asymptotic tail behaviour for every smooth expectation-value generator, and the authors identify the circuit primitive that removes it. Developing that primitive into a trainable, hardware-compatible component is the obvious next construction step.
• Unconditional attribution. Attribution currently holds conditional on the shared germ, because the cheap alternative — an ancilla entangled with the register and traced out — is impractical on heavy-hex, where every eight-qubit region is a forest and reaching all data qubits from one ancilla costs on the order of a hundred two-qubit layers. Finding a cheaper unconditional design is open.
• From 8 cells to realistic calorimeter granularity. The largest existing hybrid surrogate maps a latent of 2048 binary units to a 6480-voxel shower, while the most developed fully quantum generator has been shown on an eight-pixel image from twenty-three trainable angles. Closing that gap with a circuit-only generator, and checking whether the built-in null survives at that scale, is the central open engineering question.

Target Audience

Researchers working on quantum machine learning and quantum generative models, particularly those interested in rigorous benchmarking rather than headline scores; particle physicists and detector-simulation developers evaluating fast surrogate generators for HL-LHC computing; and quantum-hardware practitioners concerned with compiling logical ansätze onto real device topologies without SWAP overhead. It is also valuable for statisticians and applied mathematicians interested in polynomial chaos, uncertainty quantification, and the limits of expectation-value readout.

Authors’ abstract

We present the quantum polynomial chaos expansion, a generative algorithm in which a single circuit is the entire model, and we use it to learn calorimeter images. In a classical chaos expansion the randomness is the input and the coefficients are fitted. Here the randomness is still the only input, entering the circuit as rotation angles and re-uploaded at every block, so that each measured observable is a chaos expansion of the latent variables whose order equals the circuit depth, and what is fitted are the gate angles themselves. Expressivity therefore grows with depth rather than with classical coefficients, correlations between outputs arise only from entangling gates, and a single latent wire read by all qubits carries the collective mode of the data. Nothing fitted stands between the circuit and the sample, so switching the entanglers off is a setting of the model itself and provably yields independent outputs, and attribution of the learned correlations to individual gates becomes a measurement. Choosing between two measurement bases shot by shot sharpens attribution into certification, and the trained model violates the Bell bound obeyed by every classical generative model with local response, whatever its size. We train the model on Geant4 shower data, execute the identical circuit on a superconducting processor with its accuracy loss predicted in advance, prove a no-go theorem for the tail dependence of every smooth generator read out through expectation values, and identify the circuit primitive that removes this limit.