HyQ-LC: Improving Light-Curve Classification with Hybrid Quantum--Classical Learning

AI summary

Overview

Research area: Quantum machine learning applied to astronomy — specifically hybrid quantum–classical classification of stellar light curves from time-domain surveys.

Technical level: Advanced. The paper assumes familiarity with variational quantum circuits, amplitude encoding, data re-uploading, Pauli-expectation readouts, class-weighted cross-entropy, and bootstrap-based uncertainty estimation.

Scope: A controlled benchmark study asking whether a nine-qubit quantum readout fused into a classical window-MLP backbone improves ten-class stellar light-curve classification, evaluated against a parameter-matched pure MLP, a classical readout replacement, fixed score fusion, and LightGBM references on a shared 20,000-object test set.

What This Paper Is About

Large time-domain sky surveys produce enormous numbers of light curves — sequences recording how a star's brightness changes over time — and these must be sorted into variable-star classes at scale. The authors build HyQ-LC, a hybrid model that feeds 359 handcrafted light-curve features into a classical neural network plus a small simulated quantum circuit, and ask whether the quantum part actually helps. Crucially, the paper is framed as a careful evaluation study: it separates "does the whole hybrid model beat a control?" from "does the quantum branch specifically contribute?" and reports that the first question gets a modest yes while the second remains unresolved.

Key Contributions

1. A parameter-matched hybrid architecture with a residual quantum head. HyQ-LC combines a 23,170-parameter window MLP backbone, a 540-parameter nine-qubit readout (amplitude encoding plus grouped angle re-uploading across twelve layers), and a 20,528-parameter residual fusion head, totalling 44,238 parameters — exactly the count of the pure-MLP control it is compared against.

2. A controlled readout-replacement experiment. Beyond comparing the hybrid to a separate MLP, the authors swap the quantum readout for a 540-parameter classical MLP readout inside the identical residual architecture, isolating the quantum branch more specifically than a whole-model comparison can.

3. Separate fixed-fusion and tree-based references. A token-wise quantum classifier (32,410 parameters) is fused with the window MLP at a 55,581-scalar budget, and compared against a classical fusion control (54,631 scalars) and an approximately budget-matched LightGBM with 55,500 leaf values, plus a full-budget LightGBM reference with 112,220 leaf values.

4. A released, recomputable evaluation package. Shared test features, labels, per-object predictions, result tables, checkpoint and source-prediction hashes, and a portable metric-recomputation script are made public, along with a feature manifest recording each of the 359 feature indices, names, and group assignments.

Main Findings

• The hybrid beats the parameter-matched pure MLP on point estimates. On 20,000 shared test objects, the residual hybrid gained +0.8276 percentage points Macro-F1 on seed 42 (hybrid 92.2913 vs. MLP 91.4637) and +1.2393 pp on seed 44 (92.3005 vs. 91.0611). The seed-42 F1 interval, [-0.0650, +1.6945], included zero; the seed-44 interval, [+0.5153, +2.0053], did not.

• Accuracy gains were positive on both seeds. ΔACC was +0.6000 pp with interval [+0.3450, +0.8550] on seed 42, and +0.8450 pp with interval [+0.5900, +1.1000] on seed 44.

• The quantum-specific contribution is not resolved. Replacing the nine-qubit readout with a matched 540-parameter classical readout inside the same residual architecture gave differences of only +0.0205 pp (seed 42, interval [-0.3299, +0.3726]) and +0.2965 pp (seed 44, interval [-0.0333, +0.6982]). Both intervals include zero, so the quantum branch's own contribution cannot be separated from architecture, pre-training, or the residual head.

• Fixed quantum fusion slightly edges classical fusion. The fixed-fusion configuration reached 92.3371% Macro-F1 and 94.7450% accuracy, versus 92.1510/94.7500 for classical fusion: ΔF1 +0.1860 pp (interval [-0.2869, +0.6740]) with essentially unchanged accuracy (ΔACC -0.0050 pp). Different mixing rules (logit averaging for the quantum fusion, probability averaging at 0.6/0.4 for the classical one) prevent attributing this to the quantum component alone.

• Against budget-matched LightGBM, the ranking depends on the metric. The hybrid had +0.1166 pp Macro-F1 (interval [-0.8111, +1.0763]) but -0.1200 pp accuracy, corresponding to 24 fewer correct classifications.

• The full-budget LightGBM reference wins on both metrics. It scored 92.9681% Macro-F1 and 95.1100% accuracy, versus 92.3371/94.7450 for the fixed hybrid — a gap of -0.6311 pp F1 and -0.3650 pp accuracy.

• Class-level gains are concentrated in three classes. Against the budget-matched tree, the largest positive F1 differences were CEP (+1.0040 pp), RRC (+1.1429 pp), and ROT (+0.5128 pp). CEP and RRC contain only 45 and 357 test objects respectively, so equal class weighting lets these small gains lift Macro-F1 even while total correct predictions fall.

• Two classes dominate the test cohort. EW and SR together account for 61.89% of the 20,000 test objects, while CEP has only 45 objects — a 140-fold range that shapes the discrepancy between Macro-F1 and accuracy.

• The test cohort is not an untouched confirmation set. The authors state the cohort was previously examined during exploratory development and that no separate confirmation cohort is evaluated.

• Everything is classically simulated. All circuits use classical state-vector simulation with automatic differentiation for gradients; no quantum-hardware runtime or quantum-advantage claim is made.

Methodology in Plain English

The data. Each light curve is summarised by 359 precomputed numbers, organised into seven groups: 81 base statistics (mean, standard deviation, variability at 16 time scales from 0.02 to 655.36 days), 17 global descriptors (skewness, kurtosis, trends, observation counts, measurement errors), 32 histogram summaries, 49 pair statistics at seven observation-index lags (1, 2, 4, 8, 16, 32, 64), 63 three-point pattern features, 64 differences across 16 physical-time bins, and 53 cached periodicity descriptors covering candidate periods and harmonic amplitudes. These groups sum to exactly 359. The training pool has 633,034 objects and the test set 20,000. Preprocessing uses frozen medians and scales from a 50,000-object training subset, replaces non-finite values with medians, and clips values to [-10, 10].

The classical half. Inputs are squashed through an arctangent, padded with one zero, and split into 120 three-component tokens. Local window mixing over these tokens yields a 120-dimensional representation and a first set of ten class scores.

The quantum half. All 359 features are loaded into a nine-qubit register two ways: as 512 normalised amplitudes (with zeros padded on and a fallback to the all-zero state if the norm is tiny), and as nine "grouped angles" formed by averaging contiguous blocks of 40 padded entries. Twelve layers then apply trainable rotations, a ring of entangling CNOT gates, and the re-uploaded angles. Rather than reading out a single bitstring, the model measures nine Pauli expectations along X, Y, and Z, giving 27 numbers. The circuit itself has 540 trainable parameters.

Putting them together. The 120 classical features and 27 quantum features are normalised separately, concatenated into 147 dimensions, passed through a small 128-unit residual network, and added to the backbone's original scores. Because the residual layer starts at zero, training begins from exactly the backbone's predictions and the branch learns a correction.

Three kinds of comparison. (1) The full hybrid vs. a pure MLP with the identical 44,238-parameter budget — does more architecture help? (2) The hybrid with the quantum readout swapped for a 540-parameter classical readout — does the quantum part specifically help? (3) A separate fixed-fusion model averaged with the window MLP, compared against classical fusion and LightGBM sized by cumulative leaf values (10 classes × 370 rounds × 15 leaves = 55,500 for the budget-matched model; 10 × 362 × 31 = 112,220 for the full-budget one).

Training and scoring. Class imbalance is handled with inverse-frequency weights; at inference, logits are shifted by a correction strength α chosen from {0, 0.25, 0.5, 0.75, 1}. Checkpoints and α are picked by validation Macro-F1. The pure MLP trains for 80 epochs from scratch; the hybrid uses 40 pre-training epochs plus 40 branch-training epochs, with the backbone frozen for the first ten. Two seeds are used, 42 and 44, sharing a pre-trained backbone. Uncertainty is estimated from 10,000 paired-object bootstrap resamples that reuse the same object indices for both models.

Why This Matters

For research. The paper models a discipline that quantum machine learning needs more of: it refuses to let a whole-model win stand in for evidence about the quantum component. Its readout-replacement control shows a hybrid improving over a dense network while the quantum-specific increment stays statistically indistinguishable from zero — a distinction that many benchmark papers blur. The class-balanced scoring analysis also demonstrates concretely how Macro-F1 can rise while total correct predictions fall, which matters for anyone reporting on imbalanced astronomical catalogues.

Real-world applications:

• Survey classification pipelines — automated sorting of light curves from large time-domain surveys into variable-star types for follow-up prioritisation.
• Rare-class discovery — the class-balanced framing is directly relevant to finding sparse categories like CEP and RRC, though the paper is explicit that its gains are class-specific rather than uniform.
• Feature-based tabular astronomy — the 359-feature representation and released manifest give a reusable benchmark substrate for comparing classifiers without touching raw photometry.
• Reproducible ML benchmarking — per-object predictions, hashes, and a recomputation script make the reported statistics independently checkable, useful for groups building audit practices.

Industry relevance. The honest-control template transfers directly to industrial quantum-ML pilots: when a vendor claims a quantum layer improves a model, the informative question is whether a matched classical replacement inside the same network does just as well. The paper's budget-matching discussion — showing that equal parameter counts or equal leaf counts do not imply equivalent function classes, optimisation effort, or inference cost — is practical guidance for anyone evaluating a hybrid proposal.

Future Directions

1. Evaluate on genuinely untouched data. The authors state that the 20,000-object cohort was examined during exploratory development and that no separate confirmation cohort is evaluated, so testing on previously unexamined objects is the most direct next step.

2. Distinguish pre-training from fine-tuning effects. The hybrid and the pure MLP differ in initialisation and training trajectory: the hybrid starts from a 40-epoch pre-trained backbone, and the selected checkpoints land at branch epochs two and four, meaning the backbone is still at its pre-trained values. Independently pre-trained backbones would separate these factors.

3. Match the fusion rules. The quantum and classical fusion configurations use different mixtures (logit averaging at 1/2–1/2 versus probability averaging at 0.6/0.4) and different constituent sizes, so a matched-rule comparison is needed before the ensemble difference can be attributed to the quantum constituent.

4. Test whether the quantum readout matters at larger scale or on hardware. The present readout contributes 540 parameters and 27 measured features, all classically simulated; whether it confers any advantage at larger circuit size, different encodings, or on real quantum processors remains open. The paper explicitly notes that neither a hardware speed advantage nor a general quantum advantage follows from its results.

Target Audience

Researchers working on hybrid quantum–classical machine learning who need a well-controlled negative-leaning result to calibrate expectations; astronomers and survey scientists building light-curve classification pipelines who want to know whether quantum branches are worth the integration cost; and benchmarking or methodology specialists interested in how to design readout-replacement and budget-matched controls. It is not an introductory paper — readers should already be comfortable with variational circuits, Pauli observables, and bootstrap confidence intervals. Practitioners evaluating vendor quantum-ML claims will find the control structure more useful than the headline numbers.

Authors’ abstract

Light curves reveal stellar variability and support studies of stellar evolution and cosmic distances. Reliable classification is therefore important for large time-domain surveys. We present HyQ-LC for ten-class stellar light-curve classification using 359 handcrafted features and 633,034 training objects. The main model combines a window multilayer perceptron (MLP) backbone with a nine-qubit readout using amplitude encoding and angle re-uploading. Both this hybrid and its pure-MLP control have 44,238 parameters. On a shared 20,000-object test set, the hybrid has macro-averaged F1 (Macro-F1) point gains of 0.8276 and 1.2393 percentage points over the pure MLP on two fine-tuning seeds. A separate fixed-fusion configuration achieves slightly higher Macro-F1 but lower accuracy than approximately budget-matched LightGBM. Its positive class-level differences are concentrated in CEP, RRC, and ROT, while the full-budget LightGBM reference scores higher on both metrics. All circuits are simulated classically. The comparisons retain existing validation-selected checkpoints and use the same test objects, with features, predictions, and recomputation code publicly available. The results identify point-estimate improvements over the tested controls while leaving the specific contribution of the quantum branch unresolved.