Quantum Bootstrap Sampling (QBS)

• Quantum Bootstrap Sampling (QBS) is a framework that combines consistency constraints, resampling techniques, and local-to-global reconstruction to analyze quantum systems.
• It enables applications from numerical bootstrapping in quantum mechanics to statistical inference with classical shadows and global process reconstruction.
• QBS improves computational efficiency by optimizing over feasible quantum data via SDP feasibility, quantum amplitude estimation, and hardware-accelerated resampling.

Quantum Bootstrap Sampling (QBS) denotes a family of bootstrap-style procedures in quantum science and quantum computing in which admissible states, observables, resamples, or device models are explored through consistency constraints, resampling rules, or quantum superposition. The phrase is not used uniformly in the literature: some works treat it as scanning feasible moment data subject to positive-semidefinite constraints in quantum mechanics, some as nonparametric bootstrap applied to quantum-generated classical data, some as a quantum algorithm that computes the ideal bootstrap by encoding all resamples in superposition, and some as a local-to-global reconstruction principle for processes, controls, embeddings, or codes (Bhattacharya et al., 2021, Ghysels et al., 12 Nov 2025, Chen et al., 1 Apr 2026, Govia et al., 2019, Wiebe et al., 2014, Liu et al., 2023, Li, 29 Jan 2026). This suggests that QBS is best understood as an umbrella label whose precise meaning is fixed by context.

1. Terminological scope

Usage family	Core object	Representative arXiv ids
Constraint-feasibility bootstrap	Moments, correlators, spectra, SDP feasibility regions	(Bhattacharya et al., 2021, Lawrence et al., 9 Dec 2025, Bao et al., 2018)
Statistical bootstrap	Resamples of quantum data or exact bootstrap CDFs	(Ghysels et al., 12 Nov 2025, Chen et al., 1 Apr 2026, Yu et al., 24 Aug 2025)
Structural bootstrapping	Global models inferred from local fragments or pairwise data	(Govia et al., 2019, Wiebe et al., 2014, Liu et al., 2023, Li, 29 Jan 2026)

In numerical bootstrap work on quantum mechanics, the operative object is a feasible set of moments or expectation values. One does not resample experimental shots; instead one scans or optimizes over data vectors constrained by recursion relations, commutators, and positive semidefiniteness of moment matrices. In this sense, “sampling” refers to exploration of a convex or nearly convex consistency region in data space rather than Monte Carlo over raw observations (Bhattacharya et al., 2021, Lawrence et al., 9 Dec 2025).

In statistical settings, the meaning is closer to the classical bootstrap. One either resamples quantum-generated classical data such as classical-shadow snapshots, or uses a quantum computer to represent all bootstrap resamples coherently and estimate the ideal bootstrap functional via amplitude estimation. Here the word “bootstrap” retains its standard inferential meaning, but the implementation is quantum-specific (Ghysels et al., 12 Nov 2025, Chen et al., 1 Apr 2026, Yu et al., 24 Aug 2025).

A third usage is structural. Pairwise tomographic data, fragment solutions, local control characterizations, or lower-dimensional chain complexes are used to bootstrap a larger quantum object subject to global consistency equations. In these works, bootstrap means assembling a global model from local constraints rather than resampling observations (Govia et al., 2019, Wiebe et al., 2014, Liu et al., 2023, Li, 29 Jan 2026).

2. Constraint-based QBS in quantum mechanics

In “Numerical Bootstrap in Quantum Mechanics,” the basic data for an energy eigenstate of $H=p^2+\mathcal V(x)$ are expectation values $\langle x^n\rangle$. Recursion relations derived from $\langle[H,\mathcal O]\rangle=0$ and $$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$ reduce all moments to a finite data vector $\mathscr D$, while positivity of $\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$ implies a Hankel matrix $M_{ij}=\langle x^{i+j}\rangle$ satisfying $M(\mathscr D)\succeq 0$. The resulting procedure scans $\mathscr D$, keeps only PSD points, and identifies “allowed islands” associated with discrete states (Bhattacharya et al., 2021).

For the quartic double well $H=p^2-m^2x^2+gx^4+\mathcal V_0$, the independent data are $\langle x^n\rangle$0. For $\langle x^n\rangle$1, $\langle x^n\rangle$2, $\langle x^n\rangle$3, the bootstrap finds $\langle x^n\rangle$4, $\langle x^n\rangle$5, and $\langle x^n\rangle$6; for $\langle x^n\rangle$7, $\langle x^n\rangle$8, $\langle x^n\rangle$9, it finds $\langle[H,\mathcal O]\rangle=0$0, $\langle[H,\mathcal O]\rangle=0$1, and $\langle[H,\mathcal O]\rangle=0$2. The widths of the allowed islands satisfy $\langle[H,\mathcal O]\rangle=0$3 with $\langle[H,\mathcal O]\rangle=0$4, so the error scales roughly like $\langle[H,\mathcal O]\rangle=0$5 (Bhattacharya et al., 2021).

The same framework extends to supersymmetric partner potentials and to the singlet sector of $\langle[H,\mathcal O]\rangle=0$6 vector quantum mechanics. For $\langle[H,\mathcal O]\rangle=0$7, the independent data become $\langle[H,\mathcal O]\rangle=0$8, and the spectra of $\langle[H,\mathcal O]\rangle=0$9 and $$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$0 agree with the expected SUSY level shift within error bars. In the $$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$1 model, the bootstrap reproduces strong-coupling scaling $$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$2, and the large-$$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$3 saddle gives $$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$4 (Bhattacharya et al., 2021).

“Quantum bootstrap for central potentials” generalizes the moment-matrix program to three-dimensional radial problems, including non-algebraic Yukawa and Gaussian potentials. The generic moment matrix is $$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$5, while for ground states the additional matrix $$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$6 is also PSD. On the radial half-line, boundary terms at $$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$7 induce anomaly variables such as $$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$8 and $$\langle H\mathcal O\rangle=E\langle\mathcal O\rangle$$9, which must be included explicitly in the SDP. For the Cornell potential, the method determines the critical coupling to better than one part in $\mathscr D$0; lower bounds on energies are occasionally precise to greater than one part in $\mathscr D$1 (Lawrence et al., 9 Dec 2025).

A related high-energy-theory variant appears in quantum algorithms for the conformal bootstrap, where crossing equations are mapped from polynomial matrix programs to SDPs. There, “sampling” refers to exploring the space of CFT data consistent with positivity and crossing via quantum SDP solvers based on Gibbs sampling, trace estimation, and block-encoding (Bao et al., 2018).

3. Resampling-based QBS in quantum statistical inference

In the classical-shadow setting, the raw data are i.i.d. shadow snapshots

$\mathscr D$2

and nonparametric bootstrap proceeds by sampling these snapshots with replacement, recomputing the estimator, and using the bootstrap distribution for uncertainty quantification. The paper studies both mean estimators and the median-of-means estimator

$\mathscr D$3

Its central empirical conclusion is that bootstrap distributions are “very different from the Gaussian approximations,” with heavy tails and asymmetry, and that the theoretical shadow bounds are not tight for the circuits studied (Ghysels et al., 12 Nov 2025).

The same work reframes bootstrap output as a risk object. Using bootstrap replicates, one computes EV@R and ES from empirical tail quantiles rather than from a Gaussian surrogate. For a particular observable $\mathscr D$4, the $\mathscr D$5 EV@R is $\mathscr D$6 under the bootstrap distribution and $\mathscr D$7 under the Gaussian approximation, a roughly $\mathscr D$8 discrepancy in the tail estimate. Aggregated over all $\mathscr D$9 observables in the circuit, mean absolute differences $\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$0 for EV@R and ES are around $\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$1 with standard deviation about $\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$2 (Ghysels et al., 12 Nov 2025).

“Quantum Statistical Bootstrap” moves the same inferential target onto quantum hardware. For data $\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$3, the ideal bootstrap CDF is

$\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$4

with $\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$5. QBOOT prepares the uniform superposition over all $\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$6 index-resamples,

$\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$7

computes $\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$8 coherently, and extracts $\langle \mathcal O^\dagger \mathcal O\rangle\ge 0$9 by quantum amplitude estimation. The resulting estimator satisfies $M_{ij}=\langle x^{i+j}\rangle$0, while the work scales as $M_{ij}=\langle x^{i+j}\rangle$1 for target accuracy $M_{ij}=\langle x^{i+j}\rangle$2, compared with $M_{ij}=\langle x^{i+j}\rangle$3 classically (Chen et al., 1 Apr 2026).

The sample-mean demonstration uses $M_{ij}=\langle x^{i+j}\rangle$4, $M_{ij}=\langle x^{i+j}\rangle$5, and $M_{ij}=\langle x^{i+j}\rangle$6, for which the exact ideal bootstrap value is $M_{ij}=\langle x^{i+j}\rangle$7. At matched cost, QBOOT exhibits the expected $M_{ij}=\langle x^{i+j}\rangle$8 error scaling, whereas classical Monte Carlo bootstrap follows $M_{ij}=\langle x^{i+j}\rangle$9; a median-of-$M(\mathscr D)\succeq 0$0 aggregation suppresses QAE outliers while preserving the asymptotic advantage (Chen et al., 1 Apr 2026).

4. Quantum-native resampling and sampling architectures

A more explicit hardware-level interpretation appears in “Exploring Quantum Bootstrap Sampling for AQP Error Assessment: A Pilot Study.” There QBS is a hybrid classical–quantum framework for Approximate Query Processing in which sample tuple results $M(\mathscr D)\succeq 0$1 are bootstrapped on a quantum computer. A Hadamard layer prepares an index superposition, QRAM maps $M(\mathscr D)\succeq 0$2, and a quantum counter aggregates the sampled bits into a bootstrap replication $M(\mathscr D)\succeq 0$3, which is then classically scaled to $M(\mathscr D)\succeq 0$4 (Yu et al., 24 Aug 2025).

In the paper’s abstraction, each measurement of the circuit yields one bootstrap replication with $M(\mathscr D)\succeq 0$5 measurement complexity, so generating $M(\mathscr D)\succeq 0$6 bootstrap replications costs $M(\mathscr D)\succeq 0$7. The pilot implementation, carried out in Qiskit simulation with up to three address qubits, validates both the quantum resampler and the counter and focuses on COUNT queries, while extensions to SUM and AVG are formulated through QRAM value loading and quantum ripple-carry addition (Yu et al., 24 Aug 2025).

An older but conceptually related line is Quibbs, a code generator for quantum Gibbs sampling of Bayesian networks. That work does not define QBS explicitly, but it combines Szegedy operators, adaptive fixed-point Grover, quantum phase estimation, and quantum multiplexors to bootstrap a Gibbs or stationary distribution from an initial state. In that setting, “sampling” refers to coherent preparation of

$M(\mathscr D)\succeq 0$8

rather than nonparametric resampling of observed data (Tucci, 2010).

This contrast is important. In AQP and QBOOT, bootstrap means resampling with replacement from an empirical distribution. In Quibbs and related quantum Gibbs-sampling work, the same vocabulary is attached to state preparation and amplitude amplification around a target distribution. The two uses are operationally different even when both are quantum sampling procedures (Yu et al., 24 Aug 2025, Tucci, 2010).

5. Structural bootstrapping from local quantum data

In process tomography, the pairwise perturbative ansatz bootstraps a multi-qubit process from two-qubit reductions. The full process $M(\mathscr D)\succeq 0$9 is modeled as a sequential composition of all pairwise two-qubit processes, so the number of free parameters becomes $\mathscr D$0, in contrast to $\mathscr D$1 for brute-force $\mathscr D$2-qubit process tomography. The reconstruction enforces consistency of all reduced two-qubit Choi states through a nonlinear least-squares fit under CPTP constraints (Govia et al., 2019).

Operationally, one characterizes each pair $\mathscr D$3 via two-qubit QPT or GST, obtaining $\mathscr D$4, computes the corresponding reduced Choi state $\mathscr D$5 implied by the global ansatz, and solves $\mathscr D$6 for all pairs. In simulations of noisy three-qubit gates, PAPA reconstructions are about one order of magnitude closer than the ideal-gate hypothesis, both for full three-qubit processes and for reduced two-qubit processes. Experimentally, PAPA+GST reconstructs $\mathscr D$7 three-qubit gates from pairwise gate sets on a superconducting device (Govia et al., 2019).

“Quantum Bootstrapping via Compressed Quantum Hamiltonian Learning” uses a smaller trusted simulator to learn a larger untrusted device. The protocol is Bayesian, particle-based, and locality-aware: SMC maintains a posterior over Hamiltonian parameters, while Lieb–Robinson bounds justify truncating the simulation region. Fisher-information analysis shows that short-time evolution is suboptimal, whereas Interactive Quantum Likelihood Evaluation with the particle-guess heuristic chooses times $\mathscr D$8 so that experiments remain informative. Numerically, an $\mathscr D$9-qubit Ising simulator can calibrate and control a $H=p^2-m^2x^2+gx^4+\mathcal V_0$0-qubit Ising simulator using only about $H=p^2-m^2x^2+gx^4+\mathcal V_0$1 kilobits of experimental data (Wiebe et al., 2014).

Bootstrap embedding provides a fragment-based analogue. Overlapping fragment Hamiltonians are solved on quantum hardware, and consistency is imposed by matching boundary density matrices through a quadratic penalty

$H=p^2-m^2x^2+gx^4+\mathcal V_0$2

SWAP tests estimate the required overlaps, and amplitude amplification reduces sampling complexity from $H=p^2-m^2x^2+gx^4+\mathcal V_0$3 to $H=p^2-m^2x^2+gx^4+\mathcal V_0$4. A distinctive feature is that full density matrices, not only $H=p^2-m^2x^2+gx^4+\mathcal V_0$5-RDMs, can be matched at fragment boundaries at little additional computational cost (Liu et al., 2023).

6. Bootstrap equations, fork complexes, and recurring limitations

In quantum coding theory, the “quantum bootstrap product” defines a CSS code by fixing $H=p^2-m^2x^2+gx^4+\mathcal V_0$6 and $H=p^2-m^2x^2+gx^4+\mathcal V_0$7 from a tensor segment of classical input codes and solving a bootstrap equation for the $H=p^2-m^2x^2+gx^4+\mathcal V_0$8-check map $H=p^2-m^2x^2+gx^4+\mathcal V_0$9: $\langle x^n\rangle$00 The solutions generally produce multiple branches $\langle x^n\rangle$01, called fork complexes. This framework unifies general hypergraph product codes of arbitrary dimensions and fracton codes typically represented by the X-cube code, generates self-correcting quantum codes from input codes with constant energy barriers, and surpasses the code-rate upper bounds inherent to HGP codes (Li, 29 Jan 2026).

Across these literatures, the recurring motif is not a single algorithm but a common logic of constrained construction. A feasible-set bootstrap samples or optimizes over moments subject to PSD and dynamical constraints; a statistical bootstrap resamples quantum data or encodes all resamples coherently; a structural bootstrap infers a global process, control model, embedding, or code from local constituents. This suggests that the stable content of QBS lies in the conjunction of consistency constraints, iterative refinement, and a sampling or search mechanism over admissible quantum objects (Bhattacharya et al., 2021, Ghysels et al., 12 Nov 2025, Govia et al., 2019, Li, 29 Jan 2026).

The same diversity explains the main misconceptions. QBS is not always a quantum-hardware algorithm: the numerical bootstrap in quantum mechanics is entirely classical SDP feasibility on quantum observables (Bhattacharya et al., 2021, Lawrence et al., 9 Dec 2025). Nor does bootstrap always mean resampling with replacement: in process tomography, compressed Hamiltonian learning, embedding, and code design it means local-to-global reconstruction under a physically motivated ansatz or consistency equation (Govia et al., 2019, Wiebe et al., 2014, Liu et al., 2023, Li, 29 Jan 2026). Conversely, when bootstrap does mean resampling, the dominant caveats are classical bootstrap caveats translated into a quantum setting, including i.i.d. assumptions, non-Gaussian tails, and limited validity in extreme tails (Ghysels et al., 12 Nov 2025).

The principal limitations are accordingly heterogeneous. Constraint-based methods face basis truncation, brute-force scanning, anomaly handling on singular domains, and the cost of large SDPs (Bhattacharya et al., 2021, Lawrence et al., 9 Dec 2025, Bao et al., 2018). Quantum resampling algorithms depend on efficient statistic oracles, QRAM-like data access, and deep amplitude-estimation circuits (Yu et al., 24 Aug 2025, Chen et al., 1 Apr 2026). Structural bootstrapping can fail when the modeling ansatz is violated, as with genuinely $\langle x^n\rangle$02-body or non-Markovian errors in PAPA, weak locality in compressed Hamiltonian learning, or current qubit and noise limitations in fragment embedding (Govia et al., 2019, Wiebe et al., 2014, Liu et al., 2023). Within these constraints, however, the term QBS consistently denotes an attempt to replace direct exhaustive computation by a bootstrap mechanism that is either constraint-driven, resampling-driven, or local-to-global.