pith. sign in

arxiv: 2605.17623 · v2 · pith:IEVGHJR3new · submitted 2026-05-17 · 🪐 quant-ph · math.OC· q-fin.PM

Where the Quantum Lives in D-Wave Hybrid Portfolio Optimization: An Operational Decomposition Audit

Luis Lozano This is my paper

Pith reviewed 2026-06-30 18:53 UTC · model grok-4.3

classification 🪐 quant-ph math.OCq-fin.PM
keywords hybrid quantum-classicalportfolio optimizationD-Wave LeapHybridtiming auditcardinality constraintmean-variance optimizationstructural theoremquantum contribution
0
0 comments X

The pith

D-Wave hybrid portfolio optimization performance comes from classical pipelines with quantum contributing under 1% of runtime.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper audits D-Wave's hybrid solvers on cardinality-constrained mean-variance portfolio problems from N=10 to 640 using the LeapHybridCQM and LeapHybridBQM interfaces against Gurobi and classical anchors. It establishes that the quantum unit access time averages 0.034 seconds out of a 5-second budget, or 0.68% of nominal time, with the remaining 99% spent on the service's classical decomposition and feasibility-aware reassembly. A matched-wall-clock classical TabuSampler reaches objective values within 0.001 of the hybrid results. A structural theorem proves that the cardinality penalty adds a dense rank-one term that fully connects the encoded graph regardless of input covariance density, which explains observed BQM degradation. Out-of-sample results on Fama-French data show QPU portfolios with lower Sharpe ratios than the 1/N baseline.

Core claim

LeapHybridCQM matches Gurobi's proven optima on all 54 head-to-head instances at N <= 120, but the mean QPU access time is 0.034 seconds out of the 5-second nominal wall-clock budget -- 0.68% of the nominal budget, approximately 0.72% of measured run time -- and the remaining ~99% is the service's classical decomposition and feasibility-aware reassembly. TabuSampler on the penalty-encoded BQM reaches final exact-K objectives within mean absolute delta 0.001 of hybrid CQM on 24 tested instances. The cardinality penalty contributes a dense rank-one term that fully connects the encoded logical graph independent of the input covariance density, an effect proven as a structural theorem; the resul

What carries the argument

The four-metric audit protocol based on SDK timing fields t_run, t_charge, and t_QPU to isolate the quantum hardware contribution from classical overhead in hybrid solvers.

If this is right

  • Reported D-Wave hybrid wins on this problem class are constraint-native classical pipelines, not quantum-sampling wins.
  • Classical heuristics reach the same objective levels as the hybrid service at the same wall-clock budget.
  • The cardinality penalty contributes a dense rank-one term that fully connects the logical graph and degrades BQM performance independent of covariance density.
  • QPU-selected portfolios deliver a mean Sharpe ratio of 1.94 versus 2.22 for the 1/N baseline on Fama-French 49 industry portfolios.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar timing audits on other hybrid quantum-classical solvers could reveal whether classical decomposition dominates reported performance on constrained combinatorial problems.
  • Problems that admit native cardinality handling may show different scaling than those requiring penalty encoding for quantum sampling.
  • Standardized timing isolation protocols across providers would enable clearer separation of quantum versus classical contributions in future benchmarks.
  • Extending the decomposition to instances beyond N=120 could test whether the classical reassembly step creates new bottlenecks as problem size grows.

Load-bearing premise

The SDK timing fields accurately isolate the quantum hardware contribution without attributing classical overhead to the QPU time measurement.

What would settle it

An independent measurement of actual QPU execution time showing it exceeds 5% of total runtime on the same instances, or a classical heuristic that cannot reach objective values within 0.001 of the hybrid service under identical wall-clock limits.

Figures

Figures reproduced from arXiv: 2605.17623 by Luis Lozano.

Figure 1
Figure 1. Figure 1: Density-axis collapse under penalty encoding. Three structurally different covariance [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Density-axis collapse under penalty encoding. Three structurally different covariance [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Formulation choice for D-Wave portfolio optimization. The constraint-native CQM [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: Formulation choice for D-Wave portfolio optimization. The constraint-native CQM [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: D-Wave hybrid solver architecture. The solver iterates between classical decomposition [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: D-Wave hybrid solver architecture. The solver iterates between classical decomposition [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Mean chain-break fraction and (b) embedding overhead (physical / logical qubits) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Mean chain-break fraction and (b) embedding overhead (physical / logical qubits) [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Chain-break fraction vs N split by covariance-density family (diagonal, block, dense), on Pegasus (left) and Zephyr (right). The three curves nearly overlap at each N, confirming that penalty encoding collapses the density axis [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: Chain-break fraction vs N split by covariance-density family (diagonal, block, dense), on Pegasus (left) and Zephyr (right). The three curves nearly overlap at each N, confirming that penalty encoding collapses the density axis. 5.2 Hybrid CQM vs BQM: Formulation Comparison [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Mean objective value vs problem size across solver families (averaged over 3 density [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: Mean objective value vs problem size across solver families (averaged over 3 density [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Relative gap to Gurobi optimal vs N. CQM (green) is indistinguishable from the optimal baseline; BQM (red) and SA (orange) diverge as penalty dilution intensifies. 5 Discussion 5.1 Formulation Choice Dominates Solver Choice On the tested instance families, the constraint-native CQM formulation produces lower objective values than the penalty-encoded BQM formulation at every N > 10, across all three density… view at source ↗
Figure 7
Figure 7. Figure 7: Relative gap to Gurobi optimal vs N. CQM (green) is indistinguishable from the optimal baseline; BQM (red) and SA (orange) diverge as penalty dilution intensifies. 10.0 12.5 15.0 17.5 20.0 22.5 25.0 27.5 30.0 Number of industries N 0.0 0.2 0.4 0.6 0.8 1.0 Chain-break fraction Pegasus (Advantage) 10.0 12.5 15.0 17.5 20.0 22.5 25.0 27.5 30.0 Number of industries N Zephyr (Advantage2) FF49 equity instances (p… view at source ↗
Figure 8
Figure 8. Figure 8: Direct QPU on real Fama–French 49 equity data: chain-break fraction vs [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: Direct QPU on real Fama–French 49 equity data: chain-break fraction vs [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Budget response curves for hybrid BQM (dashed) and CQM (solid) across problem [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 9
Figure 9. Figure 9: Budget response curves for hybrid BQM (dashed) and CQM (solid) across problem [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Stochastic validation: box plots of objective values across 10 repeated runs for CQM [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 10
Figure 10. Figure 10: Stochastic validation: box plots of objective values across 10 repeated runs for CQM [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
read the original abstract

We audit the operational decomposition of D-Wave's hybrid quantum-classical portfolio-optimization service on cardinality-constrained mean-variance-turnover instances spanning N=10 to 640, with the constraint-native LeapHybridCQM interface, the penalty-encoded LeapHybridBQM interface, and Gurobi MIQP and simulated-annealing classical anchors. We report all three SDK timing fields (t_run, t_charge, t_QPU) and define a candidate four-metric audit protocol for hybrid quantum-classical solvers. Three findings. First, the LeapHybridCQM service matches Gurobi's proven optimum on all 54 head-to-head instances at N <= 120, but the mean QPU access time is 0.034 seconds out of the 5-second nominal wall-clock budget -- 0.68% of the nominal budget, approximately 0.72% of measured run time -- and the remaining ~99% is the service's classical decomposition and feasibility-aware reassembly. Second, in a CPU-only matched-wall-clock counterfactual, TabuSampler on the penalty-encoded BQM reaches final exact-K objectives within mean absolute delta 0.001 of hybrid CQM on 24 tested instances; this does not ablate the LeapHybridCQM pipeline internals, but it shows that these objective levels are reproducible by a classical heuristic at the same wall-clock budget. Third, the cardinality penalty contributes a dense rank-one term that fully connects the encoded logical graph independent of the input covariance density, an effect we prove as a structural theorem; the resulting density-axis collapse explains the BQM degradation observed in the empirical comparison. Out-of-sample on Fama-French 49 industry portfolios, the QPU-selected portfolios deliver a mean Sharpe ratio of 1.94 versus 2.22 for the 1/N baseline. The practical implication is that reported D-Wave hybrid wins on this problem class are constraint-native classical pipelines, not quantum-sampling wins.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript audits D-Wave hybrid solvers (LeapHybridCQM and LeapHybridBQM) versus Gurobi and TabuSampler on cardinality-constrained mean-variance-turnover portfolio problems (N=10 to 640). It reports that CQM matches Gurobi optima on all 54 instances at N<=120, with mean t_QPU of 0.034 s (0.68% of 5 s nominal budget), the remainder being classical decomposition/reassembly; TabuSampler matches CQM objectives within mean absolute delta 0.001 on 24 instances; a structural theorem proves the cardinality penalty adds a dense rank-one term that collapses the BQM density axis; out-of-sample Fama-French 49 portfolios yield mean Sharpe 1.94 versus 2.22 for 1/N.

Significance. If the SDK timing fields isolate quantum sampling time without classical overhead, the audit would show that reported hybrid wins on this class are classical pipelines, not quantum-sampling advantages. The direct reporting of all three SDK fields (t_run, t_charge, t_QPU), the parameter-free structural theorem, and the matched-wall-clock TabuSampler counterfactual are concrete strengths that would make the decomposition protocol useful for future hybrid-solver evaluations.

major comments (2)
  1. [Abstract] Abstract: the central claim that 'reported D-Wave hybrid wins ... are constraint-native classical pipelines, not quantum-sampling wins' rests on t_QPU faithfully measuring only quantum sampling time. No cross-check (raw API qpu_access_time logs or quantum-disabled run) is described, so classical pre/post-processing could be folded into the 0.034 s figure and undermine the 0.68% fraction.
  2. [Abstract] Abstract: exact matching is asserted on all 54 head-to-head instances at N<=120, yet no instance-selection criteria, data-exclusion rules, or per-instance error analysis are supplied; without these the 'matches Gurobi's proven optimum' statement cannot be fully evaluated.
minor comments (2)
  1. [Abstract] The abstract introduces a 'candidate four-metric audit protocol' but does not enumerate the four metrics, making the protocol hard to reproduce from the given text.
  2. [Abstract] The out-of-sample Sharpe-ratio comparison (1.94 vs 2.22) is reported without standard errors or the number of Fama-French periods used, limiting assessment of statistical significance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive review. We address each major comment point by point below, indicating where revisions will be made.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'reported D-Wave hybrid wins ... are constraint-native classical pipelines, not quantum-sampling wins' rests on t_QPU faithfully measuring only quantum sampling time. No cross-check (raw API qpu_access_time logs or quantum-disabled run) is described, so classical pre/post-processing could be folded into the 0.034 s figure and undermine the 0.68% fraction.

    Authors: We agree that the manuscript reports the SDK t_QPU field without independent cross-validation such as raw API qpu_access_time logs or quantum-disabled runs. The D-Wave documentation defines t_QPU as the time spent on the QPU (distinct from t_run and t_charge), and this is the standard reporting convention, but the absence of the suggested checks means the isolation cannot be externally confirmed from the data presented. In the revised version we will explicitly discuss the SDK timing definitions, acknowledge this limitation of the current audit, retain the reported 0.68% fraction as derived from the documented fields, and recommend that future hybrid-solver audits incorporate such verification steps. revision: partial

  2. Referee: [Abstract] Abstract: exact matching is asserted on all 54 head-to-head instances at N<=120, yet no instance-selection criteria, data-exclusion rules, or per-instance error analysis are supplied; without these the 'matches Gurobi's proven optimum' statement cannot be fully evaluated.

    Authors: The exact-match claim is based on objective-value comparisons to Gurobi optima across the 54 instances. We acknowledge that the manuscript does not supply instance-selection criteria, data-exclusion rules, or per-instance error details. We will revise the methods and results sections to document the instance-generation procedure, selection criteria (including N ranges and random seeds), confirmation that all 54 comparisons produced identical objective values (zero deviation), and a summary table or statement of per-instance results. This will make the claim fully evaluable. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on direct measurements and independent proof

full rationale

The paper's central claims rest on SDK timing fields reported as raw measurements (t_run, t_charge, t_QPU) and a structural theorem proved mathematically for the rank-one penalty term. These steps do not reduce to fitted parameters renamed as predictions, self-citations, or definitional equivalence. The audit protocol and counterfactual comparisons are presented as independent checks without load-bearing self-reference. No equations or derivations in the provided text exhibit the enumerated circular patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper relies on the domain assumption that SDK timing fields accurately separate quantum and classical times; no free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption SDK timing fields (t_run, t_charge, t_QPU) accurately isolate the quantum hardware contribution
    Used to calculate the 0.68% QPU share in the first finding.

pith-pipeline@v0.9.1-grok · 5891 in / 1239 out tokens · 39781 ms · 2026-06-30T18:53:01.543708+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 18 canonical work pages · 1 internal anchor

  1. [1]

    Andrist, Grant Salton, Martin J

    Atithi Acharya, Romina Yalovetzky, Pierre Minssen, Shouvanik Chakrabarti, Ruslan Shaydulin, Rudy Raymond, Yue Sun, Dylan Herman, Ruben S. Andrist, Grant Salton, Martin J. A. Schuetz, Helmut G. Katzgraber, and Marco Pistoia. Decomposition pipeline for large-scale portfolio optimization with applications to near-term quantum computing. Physical Review Resea...

    work page doi:10.1103/physrevresearch.7.023142 2025

  2. [2]

    Bailey and Marcos L\'opez de Prado

    David H. Bailey and Marcos L\'opez de Prado. The S harpe ratio efficient frontier. Journal of Risk, 15 0 (2): 0 3--44, 2012

    2012

  3. [3]

    Bailey and Marcos L\'opez de Prado

    David H. Bailey and Marcos L\'opez de Prado. The deflated S harpe ratio: Correcting for selection bias, backtest overfitting, and non-normality. Journal of Portfolio Management, 40 0 (5): 0 94--107, 2014. doi:10.3905/jpm.2014.40.5.094

    work page doi:10.3905/jpm.2014.40.5.094 2014

  4. [4]

    Algorithm for cardinality-constrained quadratic optimization

    Dimitris Bertsimas and Romy Shioda. Algorithm for cardinality-constrained quadratic optimization. Computational Optimization and Applications, 43 0 (1): 0 1--22, 2009. doi:10.1007/s10589-007-9126-9

    work page doi:10.1007/s10589-007-9126-9 2009

  5. [5]

    Next-generation topology of D-Wave quantum processors

    Kelly Boothby, Paul Bunyk, Jack Raymond, and Aidan Roy. Next-generation topology of D-Wave quantum processors. 2020. doi:10.48550/arXiv.2003.00133. Pegasus and Zephyr topology design and embedding constants

    work page doi:10.48550/arxiv.2003.00133 2020

  6. [6]

    Best practices for portfolio optimization by quantum computing, experimented on real quantum devices

    Giuseppe Buonaiuto, Francesco Gargiulo, Giuseppe De Pietro, Massimo Esposito, and Marco Pota. Best practices for portfolio optimization by quantum computing, experimented on real quantum devices. Scientific Reports, 13: 0 19434, 2023. doi:10.1038/s41598-023-45392-w

    work page doi:10.1038/s41598-023-45392-w 2023

  7. [7]

    Hybrid solver service, 2025 a

    D-Wave Systems . Hybrid solver service, 2025 a . URL https://docs.dwavequantum.com/en/latest/quantum_research/hybrid.html. Accessed April 7, 2026

    2025

  8. [8]

    Hybrid solver service: Documentation and timing fields, 2025 b

    D-Wave Systems . Hybrid solver service: Documentation and timing fields, 2025 b . URL https://docs.dwavequantum.com/en/latest/quantum_research/hybrid.html. Defines run\_time , charge\_time , and qpu\_access\_time for the LeapHybridCQM and LeapHybridBQM samplers; accessed 2026-05-31

    2025

  9. [9]

    Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? Review of Financial Studies, 22 0 (5): 0 1915--1953, 2009

    Victor DeMiguel, Lorenzo Garlappi, and Raman Uppal. Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? Review of Financial Studies, 22 0 (5): 0 1915--1953, 2009. doi:10.1093/rfs/hhm075

    work page doi:10.1093/rfs/hhm075 1915

  10. [10]

    Kenneth R. French. 49 industry portfolios, 2026. URL https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/Data_Library/det_49_ind_port.html. Accessed March 31, 2026

    2026

  11. [11]

    Quantum computing for finance

    Dylan Herman, Cody Googin, Xiaoyuan Liu, Yue Sun, Alexey Galda, Ilya Safro, Marco Pistoia, and Yuri Alexeev. Quantum computing for finance. Nature Reviews Physics, 5 0 (8): 0 450--465, 2023. doi:10.1038/s42254-023-00603-1

    work page doi:10.1038/s42254-023-00603-1 2023

  12. [12]

    Strategic portfolio optimization using simulated, digital, and quantum annealing

    Jonas Lang, Sebastian Zielinski, and Sebastian Feld. Strategic portfolio optimization using simulated, digital, and quantum annealing. Applied Sciences, 12 0 (23): 0 12288, 2022. doi:10.3390/app122312288

    work page doi:10.3390/app122312288 2022

  13. [13]

    A Penalty-Free Pipeline for Direct Quantum-Annealer Portfolio Optimization

    Luis Lozano. A penalty-free pipeline for direct quantum-annealer portfolio optimization, 2026. URL https://arxiv.org/abs/2605.17628

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  14. [14]

    Ising formulations of many NP problems

    Andrew Lucas. Ising formulations of many NP problems. Frontiers in Physics, 2: 0 5, 2014. doi:10.3389/fphy.2014.00005

    work page doi:10.3389/fphy.2014.00005 2014

  15. [15]

    Portfolio selection

    Harry Markowitz. Portfolio selection. The Journal of Finance, 7 0 (1): 0 77--91, 1952. doi:10.2307/2975974

    work page doi:10.2307/2975974 1952

  16. [16]

    End-to-end portfolio optimization with hybrid quantum annealing

    Sai Nandan Morapakula, Sangram Deshpande, Rakesh Yata, Rushikesh Ubale, Uday Wad, and Kazuki Ikeda. End-to-end portfolio optimization with hybrid quantum annealing. Advanced Quantum Technologies, 2025. doi:10.1002/qute.202500753

    work page doi:10.1002/qute.202500753 2025

  17. [17]

    Dynamic portfolio optimization with real datasets using quantum processors and quantum-inspired tensor networks

    Samuel Mugel, Carlos Kuchkovsky, Escol\'astico S\'anchez, Samuel Fern\'andez-Lorenzo, Jorge Luis-Hita, Enrique Lizaso, and Rom\'an Or\'us. Dynamic portfolio optimization with real datasets using quantum processors and quantum-inspired tensor networks. Physical Review Research, 4: 0 013006, 2022. doi:10.1103/PhysRevResearch.4.013006

    work page doi:10.1103/physrevresearch.4.013006 2022

  18. [18]

    Quantum computing for finance: Overview and prospects

    Rom\'an Or\'us, Samuel Mugel, and Enrique Lizaso. Quantum computing for finance: Overview and prospects. Reviews in Physics, 4: 0 100028, 2019. doi:10.1016/j.revip.2019.100028

    work page doi:10.1016/j.revip.2019.100028 2019

  19. [19]

    Quantum Science and Technology10(2), 025025 (2025) https://doi.org/10.1088/2058-9565/adb029

    Elijah Pelofske. Comparing three generations of D-Wave quantum annealers for minor embedded combinatorial optimization problems. Quantum Science and Technology, 10 0 (2): 0 025025, 2025. doi:10.1088/2058-9565/adb029

    work page doi:10.1088/2058-9565/adb029 2025

  20. [20]

    Oberreuter, Riccardo Aiolfi, Luca Asproni, Branislav Roman, and J \"u rgen Schiefer

    Wolfgang Sakuler, Johannes M. Oberreuter, Riccardo Aiolfi, Luca Asproni, Branislav Roman, and J \"u rgen Schiefer. A real-world test of portfolio optimization with quantum annealing. Quantum Machine Intelligence, 7: 0 43, 2025. doi:10.1007/s42484-025-00268-2

    work page doi:10.1007/s42484-025-00268-2 2025

  21. [21]

    Quantum Machine Intelligence1(1), 17–30 (2019) https: //doi.org/10.1007/s42484-019-00001-w

    Davide Venturelli and Alexei Kondratyev. Reverse quantum annealing approach to portfolio optimization problems. Quantum Machine Intelligence, 1: 0 17--30, 2019. doi:10.1007/s42484-019-00001-w

    work page doi:10.1007/s42484-019-00001-w 2019

  22. [22]

    Penalty and partitioning techniques to improve performance of QUBO solvers

    Amit Verma and Mark Lewis. Penalty and partitioning techniques to improve performance of QUBO solvers. Discrete Optimization, 44: 0 100594, 2022. doi:10.1016/j.disopt.2020.100594

    work page doi:10.1016/j.disopt.2020.100594 2022