quantum-end-to-end-learning-qel

name

quantum-end-to-end-learning-qel

description

Quantum End-to-End Learning (QEL) methodology for contextual combinatorial optimization. First quantum computing-based end-to-end learning framework leveraging QAOA with context re-uploading phase-separator. Enables joint end-to-end training with stationarity guarantee, avoiding NP-hard optimization solvers. Use when: (1) solving contextual combinatorial optimization problems, (2) implementing quantum ML for decision-making under uncertainty, (3) combining QAOA with end-to-end learning, (4) designing quantum surrogate policies for optimization. Activation: QEL, contextual combinatorial optimization, quantum end-to-end learning, QAOA, context re-uploading, decision-focused quantum learning, quantum surrogate policy, quantum decision-making. Based on: "Quantum End-to-End Learning for Contextual Combinatorial Optimization" (Lee & Kwon, arXiv:2605.20222, May 2026).

Quantum End-to-End Learning (QEL) for Contextual Combinatorial Optimization

arXiv: 2605.20222 | Submitted: 13 May 2026 | Authors: Jaehwan Lee, Changhyun Kwon (KAIST)

Overview

QEL is the first quantum computing-based end-to-end learning framework for Contextual Combinatorial Optimization (CCO). It leverages Quantum Approximate Optimization Algorithms (QAOA) with a novel context re-uploading phase-separator to jointly capture relations among contexts, uncertain coefficients, and optimal solutions.

Key Innovation

Whereas classical end-to-end learning for CCO either requires solving NP-hard optimization problems (PnO/Predict-and-Optimize) or lacks interpretability (DR/Decision Rule), QEL exploits an optimization-aware structure grounded in physical principles — specifically the QAOA ansatz — that classical methods cannot readily leverage.

Core Methodology

1. Problem Formulation (Contextual Combinatorial Optimization)

Given context s (observed data) and uncertain coefficients c, the goal is to find a decision x ∈ X (combinatorial set) minimizing expected cost:

min E_{c|s}[f(x, c)]

2. Context Re-Uploading Phase-Separator

Inspired by data re-uploading in quantum ML (where classical data is encoded at multiple circuit depths), QEL proposes a context re-uploading phase-separator:

• The problem Hamiltonian (cost operator) receives the context s via a contextual encoder g_φ(s)
• The encoded context is mixed with the phase-separator operator at each QAOA layer
• This allows the same circuit to adapt its optimization behavior based on different contexts

U_P(γ, s) = exp(-i γ · g_φ(s) · H_P)

where H_P is the problem Hamiltonian and g_φ(s) encodes context-dependent coefficients.

3. Quantum Surrogate Policy

The quantum surrogate policy π_θ(x|s) is defined as:

1. Prepare initial state |+⟩^⊗ⁿ
2. Apply p layers of QAOA:
   • Context re-uploading phase-separator: exp(-i γ_p · g_φ(s) · H_P)
   • Mixer: exp(-i β_p · H_M)
3. Measure in computational basis → decision x

4. Joint End-to-End Training

• Train the contextual encoder g_φ and QAOA parameters {γ, β} jointly
• Loss function: task loss (actual cost of the decision)
• Backpropagation through the quantum circuit using parameter-shift rules or finite-difference gradients
• Stationarity convergence guarantee: QEL provides a theoretical guarantee that the joint training converges to a stationary point (unlike vanilla heuristic quantum optimization)

5. Solver-Free Inference

At inference time, given a new context s:

1. Encode s through g_φ(s) → modified Hamiltonian
2. Run QAOA with trained parameters
3. Sample measurement outcomes → near-optimal decision

No NP-hard optimization solver calls required at inference time.

Key Advantages

Aspect	Classical PnO	Classical DR	QEL (Ours)
Solver calls during training	NP-hard per iteration	None	None
Train on task loss	Yes (through solver)	Indirect	Direct
Parameter efficiency	High	Low	Very High
Interpretability	Via solver	Black-box	Physical structure
Stationarity guarantee	No	Usually	Yes

Implementation Notes

Circuit Design

• Qubit count: Equal to number of decision variables (typically 4-12 for NISQ era)
• Depth: p = 2-4 layers sufficient for many problems
• Encoder architecture: Classical neural network g_φ(s) producing coefficient vectors

Training Details

• Use stochastic gradient descent with Adam optimizer
• Gradient estimation for quantum parameters: parameter-shift rule (exact for gates of the form exp(-iθP) where P²=I)
• Batch-size: match the number of context samples per iteration
• Initialization: warm-start from random QAOA parameters

Problems Demonstrated

The paper validates QEL on:

1. Contextual knapsack: Resource allocation with uncertain item values
2. Portfolio optimization: Asset allocation under uncertain returns (budget + risk constraints)
3. Shortest path with stochastic costs: Route planning with learned edge costs

When to Use

• You have: A combinatorial optimization problem with contextual features and a quantum computer (or simulator)
• You need: An end-to-end trained policy that avoids calling classical solvers
• You want: Parameter-efficient quantum models with stationarity guarantees
• Do NOT use: When the problem has no combinatorial constraints, or when classical solvers are already extremely fast and available

Related Work

• Classical PnO: Decision-focused learning, SPO+ (Elmachtoub & Grigas), DFL (Wang et al.)
• Classical DR: Learning to optimize, direct policy learning
• QAOA: Farhi et al. (2014), standard variational quantum optimization
• Data re-uploading: Pérez-Salinas et al. (2020), universal quantum classifiers

References

• Lee & Kwon, "Quantum End-to-End Learning for Contextual Combinatorial Optimization", arXiv:2605.20222, 2026.
• Farhi, Goldstone, & Gutmann, "A Quantum Approximate Optimization Algorithm", arXiv:1411.4028, 2014.
• Elmachtoub & Grigas, "Smart 'Predict, then Optimize'", Management Science, 2022.

Keywords: quantum end-to-end learning, contextual combinatorial optimization, QAOA, quantum approximate optimization, context re-uploading, quantum surrogate policy, decision-focused quantum learning, quantum machine learning, quantum decision-making, stationarity guarantee, parameter-shift rule, NISQ optimization