Quantum Boosting (QBoost): Sparse QUBO Ensembles

• Quantum Boosting (QBoost) is a quantum annealing-based ensemble method that selects a sparse set of weak learners using a QUBO formulation for binary optimization.
• It transforms ensemble construction into a binary quadratic problem, balancing label correlations with weak learner diversity and ensuring efficient model regularization.
• Practical applications range from defect detection in manufacturing using decision tree classifiers to material phase classification with interpretable feature selection.

Searching arXiv for recent and foundational papers on QBoost and related quantum boosting formulations. I’m checking arXiv for the cited papers and adjacent work on quantum boosting formulations. Quantum Boosting, commonly abbreviated QBoost, most often denotes a quantum-annealing-based ensemble method in which a strong predictor is obtained by selecting a sparse subset of weak learners through a quadratic unconstrained binary optimization problem. In this canonical form, the ensemble weights are binary inclusion variables rather than arbitrary real coefficients, and the optimization is delegated to a quantum annealer. The literature, however, does not use the label uniformly: alongside the annealing-based classifier, there are regression adaptations of QBoost and several distinct PAC-theoretic “quantum boosting” algorithms built from AdaBoost-, SmoothBoost-, and projection-based constructions rather than QUBO subset selection (Guijo et al., 2022, Góes et al., 2021, Arunachalam et al., 2020).

1. Terminology and historical scope

In the annealing literature, QBoost appears as an early quantum-annealer-driven method for supervised binary classification. Its defining idea is to cast weak-learner inclusion as a regularized squared-loss optimization over binary variables and to solve that optimization on hardware natively designed for binary quadratic objectives. An important extension in this lineage is RQBoost, which repeats QBoost over multiple resampled splits in order to improve data usage and produce probability estimates rather than only hard class decisions (III et al., 2016).

The same general phrase, “quantum boosting,” later came to denote several algorithmically different objects. One line studies PAC-theoretic boosting with quantum subroutines for weighted-error estimation or smooth distribution preparation; another develops confidence-rated quantum boosting for non-binary hypotheses; another proves a supervised-learning guarantee for quantum AdaBoost on variational quantum classifiers; and a separate line uses the name QuantumBoost for a projection-based PAC booster with lazy Bregman projections. A boosting method for ensembles of QSVMs is also explicitly presented as distinct from canonical QBoost, because its quantum component is the feature map rather than annealer-based subset selection (Arunachalam et al., 2020, Izdebski et al., 2020, Bera et al., 2021, Wang et al., 2024, Abbas et al., 6 Oct 2025, Rastunkov et al., 2022).

This nonuniformity matters conceptually. In one usage, QBoost is a sparse ensemble-selection model over weak classical learners; in another, it is a weak-to-strong learner conversion theorem in quantum PAC learning. A plausible implication is that “QBoost” now names a family of related quantum-ensemble ideas rather than a single fixed algorithmic object.

2. Canonical annealing-based formulation

The canonical classification form of QBoost is a sparsity-regularized ensemble objective over binary weak-learner selection variables. In the manufacturing study, the optimization problem is written as

$$w^{opt}=arg \min_{w}\Big(\sum_{s}^{S}\Big(\frac{1}{N}\sum_{i}^{N}w_{i}h_{i}(\vec{x}_{s})-y_{s}\Big)^{2}\Big)+\lambda\Vert w\Vert_{0},$$

where $h_i(\vec{x})\in[-1,1]$, $y_s\in\{-1,+1\}$, $w_i$ is binary, and $\|w\|_0$ imposes sparsity. The stated role of the regularization term is to promote weight sparsity, discourage overly complex models, and improve generalization. Because the variables are binary, $w_i^2=w_i$, and the objective can be written in QUBO form $\sum_{i,j}Q_{ij}w_iw_j$, with off-diagonal entries encoding pairwise weak-learner correlations and diagonal terms encoding the linear bias contribution (Guijo et al., 2022).

The same structural idea is restated in the high-entropy-alloy study. There the ensemble prediction rule is

$$H(x)=\mathrm{sign}\left(\sum_{i=1}^{N}\omega_i h_i(x)\right),$$

with $\omega_i\in\{0,1\}$, and training minimizes a squared-loss objective with an $L_0$-type sparsity penalty,

$h_i(\vec{x})\in[-1,1]$0

Its QUBO expansion is given as

$h_i(\vec{x})\in[-1,1]$1

with $h_i(\vec{x})\in[-1,1]$2 and $h_i(\vec{x})\in[-1,1]$3, followed by the Ising mapping $h_i(\vec{x})\in[-1,1]$4 (Hoyos et al., 14 Jul 2025).

Two geometric features recur across these formulations. First, weak learners that are strongly correlated with the labels reduce the linear term through the label-correlation quantity. Second, pairwise weak-learner correlations contribute quadratic penalties, so highly correlated learners tend not to be selected together. This reproduces the standard boosting intuition that ensemble diversity supports generalization (Guijo et al., 2022).

3. Weak learners, optimization workflow, and hardware mapping

Canonical QBoost does not prescribe a single weak-learner family. In the manufacturing application, the weak classifiers are decision tree classifiers, often described as “small decision trees” or DTCs, and hyperparameter studies vary the initial number of weak classifiers, the decision-tree depth, and the regularization $h_i(\vec{x})\in[-1,1]$5. The weak-learner generation process is described in boosting language: classifiers are trained sequentially while penalizing previously misclassified examples, with distribution initialization $h_i(\vec{x})\in[-1,1]$6, weighted error computation for each weak learner, and the standard exponential reweighting update

$h_i(\vec{x})\in[-1,1]$7

Final classification is by thresholding the selected weak-learner aggregate,

$h_i(\vec{x})\in[-1,1]$8

An important practical detail is that the DTC pool is trained classically once and then reused while sweeping $h_i(\vec{x})\in[-1,1]$9, because $y_s\in\{-1,+1\}$0 affects only the QUBO stage. The paper states that this makes tuning much faster than retraining an entire classical boosting model for each regularization value, and reports use of D-Wave’s Hybrid Sampler for at least the regularization exploration (Guijo et al., 2022).

In the HEA phase-classification study, the weak learners are even more tightly constrained: each weak classifier is a decision stump, trained on a single unscaled feature, with one stump per descriptor and thus 31 weak classifiers total. This design makes QBoost simultaneously a classifier and an interpretable feature selector, because each selected stump corresponds directly to a selected physical descriptor (Hoyos et al., 14 Jul 2025).

The hardware consequences of the QUBO form were already visible in early work. Because the squared-loss expansion couples weak learners pairwise, the induced QUBOs are often dense. On a hardware platform with 512 nominal qubits, 476 functional qubits, sparse connectivity, and chain-based embeddings, the authors report that only about 25 fully connected binary variables could be embedded. QBoost therefore proceeded greedily over batches of weak learners rather than by global optimization over the full pool. The same study emphasizes several practical constraints: heuristic embedding, chain-strength tuning, a quadratic qubit penalty for fully connected problems, and coefficient precision of about $y_s\in\{-1,+1\}$1 of full range within one standard deviation due to intrinsic control error (III et al., 2016).

Later application papers expose a more mature but still hardware-specific workflow. In the HEA pipeline, QBoost is run on D-Wave Advantage with Pegasus topology through the D-Wave Ocean SDK; minor embeddings are generated using the default heuristic algorithm; chain strength is $y_s\in\{-1,+1\}$2; annealing time is $y_s\in\{-1,+1\}$3; the number of reads is $y_s\in\{-1,+1\}$4; and the retained solution is the lowest-energy sample (Hoyos et al., 14 Jul 2025).

4. Industrial and scientific applications

The most detailed industrial benchmark uses QBoost for defect detection in manufacturing from X-ray images of automotive cast parts. The dataset is the “Castings” subset of GDXray+, containing 2727 images labeled as “with defect” or “without defect” according to defect bounding-box annotations. Images are reshaped to $y_s\in\{-1,+1\}$5, flattened, standardized, and normalized. Unlike the gate-based QSVM in the same study, QBoost does not require PCA for feasibility, and one of the paper’s main conclusions is that QBoost remains usable even without PCA because the annealer optimization scales with the number of weak learners rather than directly with the raw number of image pixels. On the no-PCA task, exhaustive QBoost gives $y_s\in\{-1,+1\}$6 and D-Wave QBoost gives $y_s\in\{-1,+1\}$7, while AdaBoost with 10 trees gives $y_s\in\{-1,+1\}$8. Runtime on the same raw-image problem is reported as 6.0 min for QBoost on D-Wave, 6.2 min for AdaBoost with 10 trees, and 113.0 min for exhaustive QBoost. For deployment, the study emphasizes that once training is complete, QBoost is only a sparse ensemble of selected decision trees and can run classically; inference on 546 test images averages about 2.94 ms per image with 10 initial weak learners (Guijo et al., 2022).

A second substantial application uses QBoost for interpretable phase classification of high-entropy alloys. The pipeline begins from 31 physics-informed descriptors and uses QBoost both as a standalone sparse classifier and, when its standalone performance reaches a ceiling, as a feature selector feeding a second-stage QSVM. The learning problem is decomposed into one-vs-rest binary tasks for FCC, BCC, Sigma, Laves, Heusler, and Al-X-Y B2 phases. Evaluation uses 5-fold cross-validation and an independent held-out experimental test set of 86 HEAs synthesized in the lab. For imbalanced tasks such as FCC, BCC, Sigma, and Laves, the paper uses 30 independent under-sampling rounds. QA-based QBoost generally matches or exceeds SA-based QBoost; for example, in BCC classification the reported test accuracy and test F1 are $y_s\in\{-1,+1\}$9 and $w_i$0 for both QA and SA, while in Laves classification QA reaches test accuracy $w_i$1 and test F1 $w_i$2, compared with $w_i$3 and $w_i$4 for SA. The paper also reports that each QBoost QUBO solved by QA took approximately $w_i$5–$w_i$6 seconds per instance, versus $w_i$7–$w_i$8 seconds per instance for simulated annealing on the cited cluster hardware. Interpretability is central: selected descriptors such as VEC, PFP_A1, PFP_A2, $w_i$9, and $\|w\|_0$0 are analyzed as physically meaningful indicators of phase stability (Hoyos et al., 14 Jul 2025).

These two studies emphasize different virtues of the same formalism. In manufacturing, the stress is on operation without drastic dimensionality reduction and on classical deployability after quantum training. In materials science, the stress is on sparse feature selection and physically readable decision rules. This suggests that the most mature application niche for annealing-based QBoost is not generic “quantum boosting,” but sparse subset selection in regimes where interpretability and hardware-compatible binary structure matter as much as predictive accuracy.

5. Regression generalization and PDE solving

A major extension of QBoost adapts it from classification to regression in order to combine several weak regressors into a stronger regressor for solving partial differential equations. The target example is the 1D viscous Burgers’ equation. The ensemble is

$\|w\|_0$1

or, in the PDE setting,

$\|w\|_0$2

where the $\|w\|_0$3 are pretrained neural-network weak learners. The regression loss is the mean squared error

$\|w\|_0$4

and the regularized objective is

$\|w\|_0$5

A further adaptation imposes the affine-mixture constraint $\|w\|_0$6 (Góes et al., 2021).

Because D-Wave hardware optimizes binary variables, the real coefficients $\|w\|_0$7 are represented through an $\|w\|_0$8-bit expansion,

$\|w\|_0$9

followed by

$w_i^2=w_i$0

In the reported experiments, $w_i^2=w_i$1, so $w_i^2=w_i$2. Substituting this encoding into the constrained quadratic objective yields a QUBO over the binary precision variables $w_i^2=w_i$3, with coefficients assembled from weak-learner output correlations, target values, finite-precision scale factors, and $w_i^2=w_i$4 (Góes et al., 2021).

The weak learners in this regression-QBoost study are classical neural-network regressors trained with a physics-informed residual objective, not labels alone. Diversity comes primarily from different network architectures. For the QBoost stage, the ensemble is trained on $w_i^2=w_i$5 samples and tested on $w_i^2=w_i$6 samples drawn from analytic Burgers solutions at distinct time slices. The optimization backends are Exact Solver, Simulated Annealing, and Quantum Annealing on the D-Wave 2000Q QPU, with quantum annealing time fixed at $w_i^2=w_i$7. Empirically, the weak learners have test losses $w_i^2=w_i$8, $w_i^2=w_i$9, $\sum_{i,j}Q_{ij}w_iw_j$0, and $\sum_{i,j}Q_{ij}w_iw_j$1, while the QBoost ensembles achieve test losses around $\sum_{i,j}Q_{ij}w_iw_j$2; the best reported quantum-annealing result is $\sum_{i,j}Q_{ij}w_iw_j$3 at $\sum_{i,j}Q_{ij}w_iw_j$4. The exact solver on a 16 GB RAM personal computer is limited to $\sum_{i,j}Q_{ij}w_iw_j$5, whereas simulated annealing requires roughly $\sum_{i,j}Q_{ij}w_iw_j$6 to $\sum_{i,j}Q_{ij}w_iw_j$7 seconds over the tested range. The paper’s explicit lesson is that QBoost can be generalized from binary classifier selection to finite-precision real-valued ensemble regression (Góes et al., 2021).

This regression adaptation changes the character of QBoost. The canonical classifier chooses which weak learners are on or off; the regression version solves for a finite-precision real-valued weight vector under a normalization constraint. The shared ingredient is not binary class voting but QUBO compatibility.

6. Broader quantum boosting landscape, limitations, and debates

The empirical record for canonical annealing-based QBoost is mixed. Early work found that QBoost and RQBoost were outperformed by traditional techniques on the evaluated NLP and seizure-prediction tasks, and also documented a critical negative result on a linearly separable synthetic dataset: despite the existence of a perfect separating ensemble among the weak classifiers, QBoost on the commercial annealer repeatedly selected an imperfect “bait” feature and made 90 mistakes out of 1000. That paper interprets the failure as partly algorithmic rather than purely hardware-induced, pointing to binary inclusion weights, the squared-loss surrogate, restricted score normalization, greedy subset-wise construction, embedding overhead, and coefficient-precision limits (III et al., 2016).

Later application studies are more favorable but more cautious in their claims. The manufacturing paper reports strong empirical performance and practical runtime, yet explicitly does not claim a formal quantum speedup; it presents an industrial benchmark with favorable metrics and deployability instead. It also notes sensitivity to hyperparameter tuning and instability under Random Under Sampling, which appears to increase correlations among weak learners and interacts unfavorably with the sparsity penalty. The HEA study likewise distinguishes practical runtime reduction from proof of asymptotic quantum advantage and notes that annealer outputs are near-optimal and can be affected by noise, chain breaks, and embedding quality (Guijo et al., 2022, Hoyos et al., 14 Jul 2025).

Alongside these annealing-based methods, a separate theoretical literature develops quantum boosters with very different guarantees. A 2020 quantum AdaBoost-style booster achieves a quadratic improvement over classical AdaBoost in terms of $\sum_{i,j}Q_{ij}w_iw_j$8, with complexity scaling as $\sum_{i,j}Q_{ij}w_iw_j$9 rather than linearly in $$H(x)=\mathrm{sign}\left(\sum_{i=1}^{N}\omega_i h_i(x)\right),$$0 (Arunachalam et al., 2020). “Improved Quantum Boosting” replaces AdaBoost by SmoothBoost and gives, for constant target error and confidence, complexity $$H(x)=\mathrm{sign}\left(\sum_{i=1}^{N}\omega_i h_i(x)\right),$$1 (Izdebski et al., 2020). QRealBoost extends quantum boosting to non-binary or real-valued weak hypotheses through domain-partitioning hypotheses and reports query complexity $$H(x)=\mathrm{sign}\left(\sum_{i=1}^{N}\omega_i h_i(x)\right),$$2 in the summary table (Bera et al., 2021). A 2024 quantum AdaBoost paper proves a supervised-learning guarantee for binary classification and empirically shows that boosted noisy QCNNs can surpass an unboosted noiseless primitive classifier after a few rounds (Wang et al., 2024). A 2025 projection-based QuantumBoost algorithm with lazy Bregman projections gives runtime $$H(x)=\mathrm{sign}\left(\sum_{i=1}^{N}\omega_i h_i(x)\right),$$3 under its access-model assumptions (Abbas et al., 6 Oct 2025).

Taken together, these results delimit the contemporary meaning of QBoost. In its strictest sense, it is the QUBO-based sparse ensemble method optimized by quantum annealing. In a broader sense, it sits inside a larger quantum-boosting landscape that includes PAC-theoretic, agnostic, confidence-rated, and NISQ-oriented boosters with different objectives, different hardware assumptions, and different notions of advantage. The literature therefore supports two simultaneous conclusions: QBoost is a concrete annealing-compatible sparse ensemble model, and “quantum boosting” is a wider research program whose most rigorous theoretical guarantees do not coincide with the canonical annealing-based classifier.