Qubit-Efficient Variational Quantum Optimization via Pauli Correlation Encoding: Application to Large-Scale Power Demand Portfolio Optimization

AI summary

Overview

Research area: Quantum computing for combinatorial optimization — specifically variational quantum algorithms, qubit-efficient encodings, and their application to power-systems portfolio problems.

Technical level: Advanced. The paper assumes familiarity with QUBO formulations, variational circuits, Pauli observables, and quantum hardware noise.

Scope (one sentence): The paper analyzes how Pauli correlation encoding (PCE) behaves on dense, fully connected QUBO instances with real-valued nonuniform coefficients, using electric power demand portfolio optimization as the testbed across problem sizes from m = 18 to m = 10,296.

What This Paper Is About

Variational quantum algorithms map binary optimization variables onto qubits, so the qubit count grows linearly with problem size and quickly exceeds available hardware. Pauli correlation encoding instead represents many binary variables through the signs of measured multi-qubit Pauli expectation values, allowing a small register to describe a very large problem. This paper asks whether that compressed representation actually works on realistic dense quadratic problems — and specifically whether improvements in the continuously relaxed objective reliably translate into better decoded binary solutions.

Key Contributions

1. A systematic study of PCE in the dense-QUBO regime. The authors apply PCE to electric power demand portfolio optimization, which produces fully connected QUBO instances with real-valued, nonuniform coefficients — a setting that differs substantially from the structured benchmark problems (such as Max-Cut) studied previously. They report near-optimal performance with normalized cost gaps on the order of 10⁻⁴ relative to solutions with certified optimality from the commercial solver Gurobi (run with a relative optimality gap tolerance of 10⁻⁴).

2. A two-stage hybrid formulation. A time-averaged problem (Model 1, averaging window n_T = 24 starting at T = 1) supplies initialization for a time-resolved optimization (Model 2), where the hourly cost is minimized independently at each time step. The optimized Model 1 parameters θ* are reused to warm-start Model 2 and to evaluate hardware results, without additional random restarts.

3. A mechanistic account of relaxation versus discretization. By examining the distribution of Pauli correlators, the authors show that optimization performance is governed by the interplay between the continuously decreasing relaxed loss and the step-like behavior of the decoded binary cost, and that the "effective resolution" of the correlator representation changes with system size.

4. Hardware validation on a trapped-ion processor. Circuits with parameters optimized in noiseless simulation are executed on a QPU; useful binary solutions are recovered through post-processing despite noise and finite sampling, up to m = 756 in terms of preserving problem-dependent structure.

Main Findings

• Combinatorial qubit efficiency. With correlation order k = n/2 (n even), the number of encodable variables is m = 3·C(n,k). The largest instance, m = 10,296, uses only n = 14 qubits, k = 7, and 595 variational parameters. Instances studied: m = 18 (n = 4, k = 2), 60 (n = 6, k = 3), 210 (n = 8, k = 4), 756 (n = 10, k = 5), 2,772 (n = 12, k = 6), 10,296 (n = 14, k = 7).

• Near-optimal simulated results. The complete PCE workflow (sign decoding followed by greedy post-processing) stays consistently below the greedy-from-random baseline across all problem sizes. At m = 18 the solver reaches the exact optimal cost, giving a zero normalized cost gap (the point is therefore not plotted).

• Three measurement bases regardless of size. All correlators are estimated from the all-X, all-Y, and all-Z bases, so the number of state preparations and measurement configurations is fixed at three and does not scale with m.

• Continuous loss and discrete cost decouple. During optimization the normalized loss gap decreases smoothly and monotonically, whereas the decoded binary cost moves in step-like, non-monotonic fashion because of the nonlinear threshold map x_i = (1/2)[sign(⟨Π_i^(k)⟩) + 1]. For m = 18, the cost reaches zero around iteration ~80.

• Larger systems behave more consistently. Averaging effects and a growing number of near-optimal configurations make the effective landscape smoother as m increases. This is reflected in the baselines: greedy from all-zero shows a size-independent lower bound, while greedy from random initial strings improves with m.

• Regularization preference shifts with size. β > 0 is beneficial for smaller instances, while β = 0 gives more stable performance for larger systems. The scaling coefficient α_sc ranged from 0.1 to 1.5 across the reported instances (α = α_sc·n^⌊k/2⌋, e.g. α = 6.0 at m = 18 and α = 274.4 at m = 10,296).

• Greedy beats local search for post-processing. The authors tested both the local search of Sciorilli et al. and a greedy single-pass flip (variables ranked by individual cost reduction, each flipped at most once) and found greedy consistently better.

• Hardware degrades gracefully, then saturates. QPU results follow the simulation trend up to m = 756, indicating that problem-dependent structure survives in hardware-generated states. For m = 2,772 and m = 10,296 the QPU results approach the mean of the greedy (random) distribution, i.e. the initialization becomes effectively random at the level of post-processed cost. This is attributed to correlators concentrating near zero, where sign decoding is sensitive to finite-shot projection noise scaling as O(1/√N_shot).

• Not reported in the provided text. The specific shot count used on hardware, the identity of the trapped-ion device, and the numerical match-probability values from Sec. IV.3 and the correlator distributions in Appendix B.2 are referenced but not given in the excerpt.

Methodology in Plain English

The authors take a real portfolio-selection task from power aggregator operations: choose a subset of consumers so that the total procured demand reduction lands close to a target, where the target is set to one half of the total expected available reduction. Hourly consumption data come from the Energy Management System Open Data platform — 61 days from April 1 to May 30, covering 2,143 consumer-level time series, with available reduction defined as 10% of actual consumption. For portfolios of up to 2,143 consumers, m are drawn at random from the data; for larger portfolios, synthetic profiles are created by mixup augmentation, forming convex combinations of pairs of independent consumer time series.

The cost for a given hour is the expected squared deviation between procured power and the hour's target, which expands into a covariance-weighted quadratic term plus a squared mean term — a dense QUBO. Model 1 averages this cost over a 24-hour window so a single portfolio is chosen for the whole window; Model 2 minimizes each hour separately, which is the operationally realistic version.

On the quantum side, a shallow layered circuit (R_y on all qubits, R_z on all qubits, then fully connected R_zz between every pair, repeated for depth l = 5) prepares a state. The binary variables are not qubits; instead, each is read off from the expectation value of a k-body Pauli operator, passed through a sigmoid with tunable sharpness to give a continuous variable in [0,1]. Optimization runs on these continuous variables with BFGS in the Qulacs state-vector simulator, using exact backpropagated gradients, with five random restarts per Model 1 instance and selection of the run giving the lowest decoded cost. Finally, the continuous values are thresholded at zero to recover bits, and a one-pass greedy flip refines the result. The "complete PCE workflow" means all three stages.

Why This Matters

Impact on research. The paper moves PCE beyond structured benchmark instances into dense, real-valued QUBO territory and makes the relaxation-to-discretization gap the object of study rather than an afterthought. The finding that the continuous loss can keep improving while the decoded binary solution is unchanged is a caution for the whole variational-optimization literature, and the correlator-distribution analysis offers a concrete diagnostic. The qutrit-like efficiency result — over ten thousand binary variables on fourteen qubits — reframes what "problem size" means for near-term devices.

Real-world applications (implied by the framework):

• Demand-response procurement by power aggregators selecting consumer portfolios under stochastic availability.
• Portfolio and asset selection in energy trading, where objectives are dense and coefficients are real-valued and nonuniform.
• General binary quadratic programs in logistics and scheduling where the variable count dwarfs available qubit budgets.
• Risk-aware procurement planning where the mean–variance structure of Eq. (8) is the natural objective.

Industry relevance. The two-stage warm-start design (cheap averaged model first, time-resolved refinement second) maps onto how operators actually plan — a strategic portfolio, then hourly dispatch — and the greedy post-processing step is a cheap classical add-on that practitioners can deploy immediately. The honest reporting that hardware advantage fades beyond m = 756 sets realistic expectations for near-term procurement.

Future Directions

• Ansatz expressivity and depth scaling. The depth is fixed at l = 5 by fiat, justified by the O(l·n²) entangling-gate cost; a systematic study of expressibility versus depth is explicitly left open.
• Mitigating the near-zero correlator bottleneck. Because decoding errors arise when |⟨Π_i^(k)⟩| ≲ 1/√N_shot, better shot allocation, bias correction, or alternative decoding rules for large instances are natural next steps.
• The β and α_sc schedules. Hyperparameters were selected by grid search and their preferred values flip between small and large systems; a principled size-dependent schedule, rather than per-instance tuning, remains unresolved.
• Scaling hardware demonstrations. Since problem-dependent structure survived only up to m = 756 on the trapped-ion processor, testing error mitigation or hardware-native layouts on larger instances is the obvious next experiment.

Target Audience

Quantum optimization researchers working on qubit-efficient encodings and variational algorithms; power-systems and energy-market researchers interested in quantum-assisted portfolio selection; and industry quantum teams evaluating whether compressed correlator representations are ready for dense, real-valued combinatorial problems rather than curated benchmarks. Readers need a working knowledge of QUBO formulations and variational circuits; the paper is not an introductory treatment of either.

Authors’ abstract

Variational quantum algorithms offer a promising route to combinatorial optimization, but their applicability is limited by the challenge of encoding large-scale problems within restricted qubit resources. In this work, we introduce a qubit-efficient variational framework based on Pauli correlation encoding (PCE) and apply it to electric power demand portfolio optimization. Binary variables are represented through expectation values of Pauli correlation operators, which encode multi-body correlations of the quantum state and provide a continuous relaxation enabling compact representations with few qubits. We further propose a two-stage hybrid formulation, in which a time-averaged problem provides initialization for a time-resolved optimization. Numerical simulations demonstrate near-optimal performance across problem sizes ranging from $m=18$ to $10{,}296$, with normalized cost gaps on the order of $10^{-4}$ relative to solutions with certified optimality. We show that the performance is governed by the interplay between continuous relaxation and discretization: the effective resolution of the correlator representation determines how reliably improvements in the continuous loss translate into better discrete solutions, with larger systems exhibiting more consistent behavior. Finally, we demonstrate robustness on a trapped-ion quantum processor, where high-quality solutions are obtained despite noise and finite sampling. These results establish PCE as a physically motivated and qubit-efficient framework for large-scale combinatorial optimization.