Pith. sign in

REVIEW 3 major objections 6 minor 40 references

Freezing a minimal set of obstructing vertices turns partitionability into an enforceable property, so divide-and-conquer QAOA covers dense graphs that previously aborted.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 12:28 UTC pith:HTSVKFZ5

load-bearing objection Real fix for DC-QAOA’s dense-graph abort: min-cut freezing at the partition layer, full coverage, and better quality than QAOA2 at equal settings—large-n numbers rest on a scoped upper-bound driver. the 3 major comments →

arxiv 2607.08138 v1 pith:HTSVKFZ5 submitted 2026-07-09 quant-ph

Adaptive Qubit Freezing Enables Robust Graph Partitioning for Divide-and-Conquer QAOA

Sokea Sang , Leanghok Hour , Dongmin Kim , Youngsun Han This is my paper
classification quant-ph
keywords QAOAdivide-and-conquer QAOAgraph partitioningqubit freezingMaxCutNISQminimum vertex cutIsing bias folding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Standard divide-and-conquer QAOA aborts on dense or highly connected graphs because it needs a small vertex separator that simply does not exist. This paper claims that an adaptive front-end called FrozenLGP can force a valid bipartition by classically freezing the smallest set of obstructing vertices (found via a max-flow minimum-vertex-cut) and folding their energetic contributions into linear bias terms on the remaining active qubits. The result is 100 percent decomposition coverage up to 10,000-vertex graphs across multiple topology families, while approximation quality is preserved on the sparse instances that ordinary divide-and-conquer already solves and is competitive with or better than other full-coverage strategies. Because freezing also strips entangling gates, the same mechanism improves noise robustness on near-term hardware. A sympathetic reader cares because it removes the structural gatekeeper that currently blocks distributed QAOA from the hard graphs that matter most.

Core claim

FrozenLGP converts partitionability from an assumption into an enforceable property: when no small vertex separator exists, it freezes the provably minimum obstructing vertex set obtained as a max-flow minimum vertex cut, folds the removed interactions into Ising linear biases on neighboring active qubits, and thereby recovers a valid bipartition whose sub-circuit Hamiltonians remain rigorous. Across families up to 10,000 vertices this yields 100 percent coverage versus 4.6 percent for the standard baseline on high-connectivity instances, while end-to-end MaxCut quality is statistically indistinguishable from ordinary divide-and-conquer where both succeed and higher than QAOA-in-QAOA at equa

What carries the argument

Adaptive qubit freezing via minimum-vertex-cut (MVC) computed by node-split max-flow: the smallest set F of vertices whose classical spin assignment (+1 or −1) severs residual connectivity is frozen, and each removed coupling Ju,v is folded into the linear bias hu of every surviving neighbor, so the energetic contribution is preserved without any quantum gate.

Load-bearing premise

That the polynomial-time upper-bound partitioner used for all large-scale coverage numbers is tight enough that the reported 100 percent coverage and the exact freeze-budget threshold still hold for the true minimum-vertex-cut algorithm on graphs far larger than the moderate set where the two were checked head-to-head.

What would settle it

Run the exact exhaustive minimum-vertex-cut partitioner and the paper’s upper-bound driver on a common suite of dense random-regular and BA graphs at n = 100–500; if coverage or the predicted freeze-budget threshold Bf = κ − (k − 1) diverges, the headline scalability claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript proposes FrozenLGP, a three-phase decomposition front end for divide-and-conquer QAOA. When standard Large Graph Partitioning fails to find a small vertex separator, Phase 2 freezes a minimum vertex cut of the residual graph (via node-split max-flow) and folds the removed couplings into linear Ising biases on active neighbors; Phase 3 (CPP) is a deterministic midpoint-split fallback with classical cut-edge rescoring that guarantees coverage on graphs no vertex-removal method can bipartition. Partition-level experiments across BA, dense ER, two-cluster, and random d-regular families up to n=10,000 report 100% coverage versus 4.6% for Phase-1-only DC-QAOA on high-connectivity instances, with a sharp recovery threshold at B_f=κ−(k−1). End-to-end MaxCut (mainly n≤20 and vs QAOA2 up to n=100) shows quality preservation on DC-QAOA-solvable instances, recovery of previously unsupported dense graphs, and higher approximation ratios than QAOA-in-QAOA at equal full coverage, with noise simulations indicating reduced entangling-gate exposure.

Significance. If the claims hold, FrozenLGP fills a genuine gap in the DC-QAOA stack: decomposition has been treated as a given, yet it is the gatekeeper, and standard LGP aborts on dense or high-connectivity graphs. Making partitionability enforceable via energy-preserving freezing, while keeping separator-based exact reconstruction when possible, is a clean design point relative to arbitrary-partition methods (QAOA2) and exponential circuit-cutting reconstruction (CutQC). Strengths include a topology-predicted freeze budget that is confirmed as a step function on d-regular graphs, strict phase priority (zero freeze overhead when LGP succeeds), transparent comparison to full-coverage alternatives, bootstrap CIs / TOST equivalence tests on the common solvable set, and resource-footprint arguments via transpilation. These make the work a useful, falsifiable contribution to distributed NISQ optimization rather than a purely heuristic partitioner.

major comments (3)
  1. Sec. 4.1–4.2 and Appendix F.5: the headline 100% coverage to n=10,000, the B_f=κ−(k−1) threshold, and reduction ratios up to ~39% are produced exclusively by the polynomial-time upper-bound driver, not by exact Algorithm 1 (MVC). Exact agreement is shown only on a 48-instance set with n=12–24 (Table 5). Although the paper scopes tightness to δ=κ families, the abstract and Sec. 3 still advertise a “provably minimum” MVC procedure as the source of the large-scale numbers. Either (i) strengthen validation of driver vs exact MVC at substantially larger n (or on more families), or (ii) restate abstract/claims so that large-n coverage and thresholds are explicitly attributed to the upper-bound driver, with exact MVC reserved for the moderate-scale check. Without that, the scalability half of the strongest claim is not fully supported as stated.
  2. Sec. 4.3–4.4 vs Sec. 4.1: end-to-end approximation quality (preservation vs DC-QAOA, recovery of unsupported dense graphs, and superiority to QAOA2) is demonstrated only up to n≤20 (dense unsupported set) and n≤100 (QAOA2 head-to-head), while the 10k-scale regime reports only preprocessing metrics. Table 3 also shows that at n≤20 classical heuristics already reach AR≈0.994–1.000, so that block mainly tests coverage and non-degradation. The manuscript should more sharply separate what is established at each scale and avoid implying that quality preservation at n≤100 independently corroborates the n=10^4 partition results. A limited larger-n quality probe (even with classical leaf solvers or coarser QAOA settings) or an explicit “partition-only beyond n=100” boundary in the abstract would close this gap.
  3. Sec. 3.2 / Algorithm 1 and Appendix C: the quantum overhead is bounded by 2^{m_max} (or 2^{m_max−1} under bit-flip symmetry) per partition, with default m_max=3. On recursive decompositions of dense graphs this multiplies across levels; Appendix D notes that inherited nonzero bias disables the symmetry halving. The paper does not quantify cumulative sub-circuit counts or wall-clock quantum cost on the same n=10^3–10^4 instances used for coverage, only leaf counts and NRL (Fig. 13). Because the operating point B_f∈{2,3} is justified partly by “minimal downstream quantum cost,” a load-bearing cost accounting (total enumerated assignments × leaf count under recursion) should be reported for the high-connectivity families, or the claim should be limited to per-partition overhead.
minor comments (6)
  1. Abstract and Introduction: “provably minimum” / “exactly minimal set” language should be cross-referenced to the Appendix F.5 scoping so readers do not equate Algorithm 1 with the large-n driver before reaching the appendix.
  2. Fig. 1 caption and Sec. 3.1: the five-node toy example is clear, but the shared-separator duplication vs frozen-node exclusion could be labeled more explicitly in panels (b) vs (e) for readers new to DC-QAOA MDR.
  3. Table 2: NRL for DC-QAOA (1.90) is averaged only over the 4.6% successful subset while FrozenLGP’s NRL is over all instances; the caption explains this, but a row restricted to the common solvable set would make the comparison less easy to misread.
  4. Appendix I: the CPP rescoring ablation is underpowered (few CPP-firing instances); stating that non-significance is weak evidence of equivalence (as the text partly does) in the main Sec. 3.3 would help.
  5. Typos / polish: “This paper makes two testable claims” (Sec. 4 opening) reads as a fragment; “ann= 10,000” style missing spaces appear in Appendix F; arXiv IDs and journal formatting of references are uneven.
  6. Data availability: “available from the corresponding author upon reasonable request” is weak for a methods paper whose main claims are empirical; releasing the partition driver and benchmark instance seeds would substantially strengthen reproducibility.

Circularity Check

0 steps flagged

No significant circularity: coverage, AR, and Bf thresholds are measured against external baselines and topology-derived predictions, not fitted or self-defined targets.

full rationale

The paper's load-bearing claims (100% decomposition coverage vs DC-QAOA Phase-1, quality preservation on the common solvable set, outperformance of QAOA2 at equal coverage, and the Bf=κ−(k−1) recovery threshold) are evaluated against external methods and external ground truth (CP-SAT, Goemans–Williamson, local/random search, monolithic QAOA, FrozenQubits, QAOA-in-QAOA). The freeze-budget threshold is derived from standard graph connectivity (vertex connectivity κ vs separator budget k−1) and then tested on held-out topology families; on random d-regular graphs it appears as a binary staircase matching the a priori formula, not as a parameter fitted to invent coverage. Bias folding is an algebraic identity of the Ising Hamiltonian (Ju,v Zu Zv → (Ju,v su) Zv), not a circular definition. Max-flow / Menger foundations are classical. Self-citations ([10], [31]) appear only as related/future work and do not justify the central algorithm or results. The upper-bound driver vs exact MVC gap (Appendix F.5) is a methodological scoping limitation, not a circular reduction of a claimed prediction to its own inputs. No equation or headline metric reduces by construction to a quantity defined by the same fitted target.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central claims rest on standard network-flow and Ising facts plus a small set of engineering knobs (k, B_f, t, p) chosen for NISQ-scale experiments. No new physical entities. Invented pieces are algorithmic (FrozenLGP phases, CPP rescoring). Free parameters are operational budgets, not fits that define the reported AR by construction.

free parameters (4)
  • freeze budget m_max / B_f
    Hand-chosen operating point (default 3; swept 1–10). Caps both classical search and 2^m subcircuit enumeration; coverage knee depends on this choice relative to κ−(k−1).
  • qubit budget k
    Hardware register width used throughout (k=6 default; k=8 in Li-replication). Defines separator budget k−1 and leaf size; all coverage thresholds are relative to this choice.
  • MDR top-t truncation
    t=20 fixed following DC-QAOA; affects reconstruction quality and classical cost but is not derived.
  • QAOA depth p and classical optimizer settings
    p=1 primary (p=1–3 noise/depth sweep); COBYLA ≤300 iterations, 10^4 shots, 3 restarts—standard but free experimental choices that bound reported AR.
axioms (5)
  • standard math Max-Flow Min-Cut theorem and Menger’s theorem: node-split unit-capacity construction yields a minimum-cardinality vertex cut.
    Appendix B; used to claim Phase-2 freezes the fewest obstructing vertices for a given separator.
  • domain assumption Unweighted MaxCut maps to Ising with J_ij = w_ij/2 and initially h_i=0; freezing u replaces J_uv Z_u Z_v by (J_uv s_u) Z_v.
    Sec. 2.1–2.2; energy preservation of frozen edges depends on this linear folding identity.
  • domain assumption Bit-flip symmetry C(z)=C(−z) for unbiased unweighted MaxCut halves the 2^|F| enumeration when inherited external bias is zero.
    Sec. 2.2, Appendix D; reduces quantum overhead but is invalid for weighted MaxCut or nonzero inherited fields (paper notes this).
  • domain assumption DC-QAOA LGP validity requires a separator S with |S|≤k−1 leaving exactly two components; MDR reconstructs by separator-bit agreement.
    Appendix A / Li et al.; FrozenLGP is defined as a repair of this structural prerequisite.
  • ad hoc to paper On studied families, δ=κ so the polynomial upper-bound driver returns the exact freeze threshold B_f=κ−(k−1).
    Appendix F.5; large-scale 100% coverage claims rely on this tightness outside the n=12–24 exact check.
invented entities (2)
  • FrozenLGP three-phase partitioner (LGP → MVC freeze → CPP) no independent evidence
    purpose: Enforce bipartitionability on any input while preferring exact separator reconstruction and bounding freeze overhead.
    Core algorithmic object of the paper; evaluated empirically, not an independent physical object.
  • Connectivity-Preserving Partitioning (CPP) with multiplicative cut-edge rescoring no independent evidence
    purpose: Unconditional 100% coverage floor on graphs no vertex-removal strategy can split (e.g., K_n).
    Graceful-degradation fallback defined in Sec. 3.3; rescoring rule is a design choice (ablation in Appendix I).

pith-pipeline@v1.1.0-grok45 · 37874 in / 3935 out tokens · 50152 ms · 2026-07-10T12:28:03.485504+00:00 · methodology

0 comments
read the original abstract

Divide-and-conquer variants of the Quantum Approximate Optimization Algorithm (QAOA) provide a promising route for executing combinatorial optimization problems beyond the qubit capacity of near-term quantum devices. However, existing approaches rely on the existence of small vertex separators and fail entirely on dense or highly connected graphs where such decompositions do not exist. We introduce Frozen Large Graph Partitioning (FrozenLGP), an adaptive decomposition framework that transforms partitionability from an assumption into an enforceable property. When standard partitioning fails, FrozenLGP identifies the minimum set of obstructing vertices through a minimum-vertex-cut computation based on max-flow and classically freezes their spin assignments. The energetic contributions of the removed interactions are rigorously preserved by folding them into linear bias terms in the Ising Hamiltonian of neighboring active qubits. Across graph sizes up to 10,000 vertices and multiple topology families, FrozenLGP achieves 100\% decomposition coverage, compared with 4.6\% for the standard divide-and-conquer baseline on high-connectivity instances. End-to-end MaxCut experiments demonstrate that FrozenLGP preserves approximation quality on instances already solvable by conventional divide-and-conquer QAOA while extending applicability to previously unsupported graphs, and outperforming alternative full-coverage decomposition strategies. Noise simulations further show improved robustness arising from reduced entangling-gate requirements. These results establish FrozenLGP as a topology-robust front end for distributed QAOA on near-term quantum hardware.

Figures

Figures reproduced from arXiv: 2607.08138 by Dongmin Kim, Leanghok Hour, Sokea Sang, Youngsun Han.

Figure 1
Figure 1. Figure 1: Overview of the FrozenLGP Partitioning Mechanism. (a) A standard graph exceeding [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic of a 1-layer (p = 1) QAOA circuit with linear bias terms. The circuit initializes an equal superposition via Hadamard gates. The problem unitary is then applied, explicitly showing the local linear bias rotations Rz(hiγ1) followed by the pairwise edge coupling rotations Rz(Ji,jγ1) implemented via CNOT cascades. A mixing layer of Rx(β1) rotations concludes the quantum evo￾lution. The measured outp… view at source ↗
Figure 3
Figure 3. Figure 3: Standard QAOA circuit for a fully active 4-node network. Node [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Circuit simplification via Qubit Freezing. By classically freezing [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overall architecture and workflow of the FrozenLGP pipeline. For simplicity and visu [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (a) Fraction of instances decomposed by freezing alone (CPP excluded) versus freeze budget [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Fraction of random d-regular instances decomposed by freezing alone (CPP excluded), as a function of degree d and freeze budget Bf , pooled over all sizes and seeds. Validity recovery follows the predicted staircase Bf ≥ d−(k−1) exactly, with a binary 0%→100% transition at every degree. The single off-pattern cell at d=9 (≈90%) is the n=10 instance group, for which the 9-regular graph on ten vertices is th… view at source ↗
Figure 8
Figure 8. Figure 8: Approximation ratio vs. n on the sparse scaling arm: FrozenLGP’s separator-preserving reconstruction leads QAOA2 at every size and the gap widens with scale, at equal (100%) coverage. Markers/bands are means and bootstrap 95% CIs over three seeds and two sparse families. worthwhile. It is therefore essential to compare against a method that already achieves full coverage. The leading such method is QAOA-in… view at source ↗
Figure 9
Figure 9. Figure 9: Node-splitting transformation underlying the minimum-vertex-cut computation. (a) An [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Classical reconstruction cost (log10 scale) versus graph size for CutQC gate cutting (4 K, solid) and FrozenLGP (∼n/k sub-circuits × 2 m, dashed) on two dense families. CutQC grows exponentially in n; FrozenLGP grows linearly. For comparison, CutQC’s circuit-cutting framework requires O(4K) evaluations for K cut gates [8] on dense graph families, while standard DC-QAOA returns no solution at all on the sa… view at source ↗
Figure 11
Figure 11. Figure 11: Full-decomposition runtime versus graph size for the four benchmark families (log–log; [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Budget dependence on the high-connectivity families ( [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Partition quality versus graph size for FrozenLGP ( [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Budget dependence on random d-regular graphs (n ≤ 1000, successful runs). (a) Graph reduction ratio ρ and (b) node redundancy level, per degree. Beyond the threshold B∗ f = d − (k − 1) both quantities are budget-independent; the reduction grows with the connectivity surplus, from ≈1–3% at d=7 to ≈25% at d=10, while NRL peaks near d=8 (≈4.3) and decreases again at higher degrees as freezing removes a large… view at source ↗
Figure 15
Figure 15. Figure 15: QAOA depth sweep (p ∈ {1, 2, 3}) on 40 dense instances (ER p=0.8 and BA d=3, n ∈ {8–16}, noiseless). (a) Coverage (%): DC-QAOA is flat at 70% at every depth; FrozenLGP (Bf=3) is flat at 100%. (b) Approximation ratio with bootstrap 95 % CI shading. the driver returns the exact threshold on these instances. To confirm the two tracks coincide, we ran the exact exhaustive MVC head-to-head against the upper-bo… view at source ↗
Figure 16
Figure 16. Figure 16: End-to-end budget sweep on 30 hard instances (ER [PITH_FULL_IMAGE:figures/full_fig_p026_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Transpiled circuit resources of monolithic QAOA (dashed) versus FrozenLGP sub-circuits [PITH_FULL_IMAGE:figures/full_fig_p027_17.png] view at source ↗

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

40 extracted references · 40 canonical work pages · 4 internal anchors

  1. [1]

    Preskill J 2018 Quantum Computing in the NISQ era and beyondQuantum279 ISSN 2521-327X URLhttps://doi.org/10.22331/q-2018-08-06-79

    work page internal anchor Pith review doi:10.22331/q-2018-08-06-79 2018

  2. [2]

    Farhi E, Goldstone J and Gutmann S 2014 A quantum approximate optimization algorithm arXiv preprint arXiv:1411.4028 27 IOP PublishingJournalvv(yyyy) aaaaaa Authoret al

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  3. [3]

    Lotshaw P C, Humble T S, Herrman R, Ostrowski J and Siopsis G 2021 Empirical performance bounds for quantum approximate optimizationQuantum Information Processing20403

    work page 2021

  4. [4]

    Wurtz J and Lykov D 2021 Fixed-angle conjectures for the quantum approximate optimization algorithm on regular maxcut graphsPhys. Rev. A104(5) 052419 URL https://link.aps.org/doi/10.1103/PhysRevA.104.052419

    work page doi:10.1103/physreva.104.052419 2021

  5. [5]

    Galda A, Liu X, Lykov D, Alexeev Y and Safro I 2021 Transferability of optimal qaoa parameters between random graphs2021 IEEE International Conference on Quantum Computing and Engineering (QCE)pp 171–180

    work page 2021

  6. [6]

    Wang Z, Hadfield S, Jiang Z and Rieffel E G 2018 Quantum approximate optimization algorithm for maxcut: A fermionic viewPhys. Rev. A97(2) 022304 URL https://link.aps.org/doi/10.1103/PhysRevA.97.022304

    work page doi:10.1103/physreva.97.022304 2018

  7. [7]

    Cuomo D, Caleffi M and Cacciapuoti A S 2020 Towards a distributed quantum computing ecosystemIET Quantum Communication13–8

    work page 2020

  8. [8]

    Tang W, Tomesh T, Suchara M, Larson J and Martonosi M 2021 Cutqc: using small quantum computers for large quantum circuit evaluationsProceedings of the 26th ACM International Conference on Architectural Support for Programming Languages and Operating Systems ASPLOS ’21 (New York, NY, USA: Association for Computing Machinery) p 473–486 ISBN 9781450383172 U...

    work page doi:10.1145/3445814.3446758 2021

  9. [9]

    Peng T, Harrow A W, Ozols M and Wu X 2020 Simulating large quantum circuits on a small quantum computerPhys. Rev. Lett.125(15) 150504 URL https://link.aps.org/doi/10.1103/PhysRevLett.125.150504

    work page doi:10.1103/physrevlett.125.150504 2020

  10. [10]

    Sang S, Hour L and Han Y 2026 Learning-optimized qubit mapping and reuse to minimize intercore communication in modular quantum architecturesPhys. Rev. Appl.25(5) 054023 URLhttps://link.aps.org/doi/10.1103/rfn7-c5n6

    work page doi:10.1103/rfn7-c5n6 2026

  11. [11]

    Li J, Alam M and Ghosh S 2023 Large-scale quantum approximate optimization via divide-and-conquerIEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems421852–1860

    work page 2023

  12. [12]

    Zhou Z, Du Y, Tian X and Tao D 2023 Qaoa-in-qaoa: Solving large-scale maxcut problems on small quantum machinesPhys. Rev. Appl.19(2) 024027 URL https://link.aps.org/doi/10.1103/PhysRevApplied.19.024027

    work page doi:10.1103/physrevapplied.19.024027 2023

  13. [13]

    Lu Y, Tian G and Sun X 2023 Qaoa with fewer qubits: a coupling framework to solve larger-scale max-cut problemarXiv preprint arXiv:2307.15260

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  14. [14]

    Huang P H, Li X R, Chuang C, Tu C H and Hung S H 2026 Paraqaoa: Efficient parallel divide-and-conquer qaoa for large-scale max-cut problems beyond 10,000 verticesarXiv preprint arXiv:2603.26232

    work page arXiv 2026

  15. [15]

    Gilbert E N 1959 Random graphsThe Annals of Mathematical Statistics301141–1144

    work page 1959

  16. [16]

    Barabási A L and Albert R 1999 Emergence of scaling in random networksScience286 509–512

    work page 1999

  17. [17]

    Goldberg A V 1985 A new max-flow algorithm URL https://hdl.handle.net/1721.1/149100

    work page 1985

  18. [18]

    Dantzig G and Fulkerson D R 2003 On the max flow min cut theorem of networksLinear inequalities and related systems38225–231

    work page 2003

  19. [19]

    Dittmann C 2017 Menger’s theoremArchive of Formal Proofs

    work page 2017

  20. [20]

    Bondy J A, Murty U S Ret al1976Graph theory with applicationsvol 290 (Macmillan London)

    work page

  21. [21]

    Lucas A 2014 Ising formulations of many np problemsFrontiers in PhysicsVolume 2 - 2014 ISSN 2296-424X URL https://www.frontiersin.org/journals/physics/articles/10.3389/fphy.2014.00005 28 IOP PublishingJournalvv(yyyy) aaaaaa Authoret al

    work page doi:10.3389/fphy.2014.00005 2014

  22. [22]

    Hadfield S, Wang Z, O’Gorman B, Rieffel E G, Venturelli D and Biswas R 2019 From the quantum approximate optimization algorithm to a quantum alternating operator ansatz Algorithms12ISSN 1999-4893 URLhttps://www.mdpi.com/1999-4893/12/2/34

    work page 2019

  23. [23]

    Nielsen M A and Chuang I L 2010Quantum computation and quantum information (Cambridge university press)

    work page

  24. [24]

    Peruzzo A, McClean J, Shadbolt P, Yung M H, Zhou X Q, Love P J, Aspuru-Guzik A and O’brien J L 2014 A variational eigenvalue solver on a photonic quantum processorNature communications54213

    work page 2014

  25. [25]

    Kandala A, Mezzacapo A, Temme K, Takita M, Brink M, Chow J M and Gambetta J M 2017 Hardware-efficient variational quantum eigensolver for small molecules and quantum magnets nature549242–246

    work page 2017

  26. [26]

    Li G, Ding Y and Xie Y 2019 Tackling the qubit mapping problem for nisq-era quantum devicesProceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating SystemsASPLOS ’19 (New York, NY, USA: Association for Computing Machinery) p 1001–1014 ISBN 9781450362405 URL https://doi.org/10.1145/3297858.3304023

    work page doi:10.1145/3297858.3304023 2019

  27. [27]

    Liu J, Li P and Zhou H 2022 Not all swaps have the same cost: A case for optimization-aware qubit routing2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA)pp 709–725

    work page 2022

  28. [28]

    Niu S, Suau A, Staffelbach G and Todri-Sanial A 2020 A hardware-aware heuristic for the qubit mapping problem in the nisq eraIEEE Transactions on Quantum Engineering11–14

    work page 2020

  29. [29]

    Jurcevic P, Javadi-Abhari A, Bishop L S, Lauer I, Bogorin D F, Brink M, Capelluto L, Günlük O, Itoko T, Kanazawa N, Kandala A, Keefe G A, Krsulich K, Landers W, Lewandowski E P, McClure D T, Nannicini G, Narasgond A, Nayfeh H M, Pritchett E, Rothwell M B, Srinivasan S, Sundaresan N, Wang C, Wei K X, Wood C J, Yau J B, Zhang E J, Dial O E, Chow J M and Gam...

    work page doi:10.1088/2058-9565/abe519 2021

  30. [30]

    Ayanzadeh R, Alavisamani N, Das P and Qureshi M 2023 Frozenqubits: Boosting fidelity of qaoa by skipping hotspot nodesProceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2ASPLOS 2023 (New York, NY, USA: Association for Computing Machinery) p 311–324 ISBN 9781450399166 URLhtt...

    work page doi:10.1145/3575693.3575741 2023

  31. [31]

    Sang S, Hour L, Lee S, Patra A, Park H C, Park M J and Han Y 2026 Landscape-similarity-guided optimization in divide-and-conquer qaoaarXiv preprint arXiv:2602.21689

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  32. [32]

    West D Bet al2001Introduction to graph theoryvol 2 (Prentice hall Upper Saddle River)

    work page

  33. [33]

    Bollobás B 1980 A probabilistic proof of an asymptotic formula for the number of labelled regular graphsEuropean Journal of Combinatorics1311–316 ISSN 0195-6698 URL https://www.sciencedirect.com/science/article/pii/S0195669880800308

    work page 1980

  34. [34]

    ACM421115–1145 ISSN 0004-5411 URLhttps://doi.org/10.1145/227683.227684

    Goemans M X and Williamson D P 1995 Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programmingJ. ACM421115–1145 ISSN 0004-5411 URLhttps://doi.org/10.1145/227683.227684

    work page doi:10.1145/227683.227684 1995

  35. [35]

    Hesterberg T 2011 BootstrapWIREs Computational Statistics3497–526 (PreprintarXiv: https://wires.onlinelibrary.wiley.com/doi/pdf/10.1002/wics.182) URL https://wires.onlinelibrary.wiley.com/doi/abs/10.1002/wics.182

    work page doi:10.1002/wics.182 2011

  36. [36]

    Wilcoxon F 1992Individual Comparisons by Ranking Methods(New York, NY: Springer New York) pp 196–202 ISBN 978-1-4612-4380-9 URL https://doi.org/10.1007/978-1-4612-4380-9_16 29 IOP PublishingJournalvv(yyyy) aaaaaa Authoret al

    work page doi:10.1007/978-1-4612-4380-9_16

  37. [37]

    Powell M J 1994 A direct search optimization method that models the objective and constraint functions by linear interpolationAdvances in optimization and numerical analysis (Springer) pp 51–67

    work page 1994

  38. [38]

    Wille R, Van Meter R and Naveh Y 2019 Ibm’s qiskit tool chain: Working with and developing for real quantum computers2019 Design, Automation, Test in Europe Conference and Exhibition (DATE)pp 1234–1240

    work page 2019

  39. [39]

    Schuirmann D J 1987 A comparison of the two one-sided tests procedure and the power approach for assessing the equivalence of average bioavailabilityJournal of Pharmacokinetics and Biopharmaceutics15657–680 URLhttps://doi.org/10.1007/BF01068419

    work page doi:10.1007/bf01068419 1987

  40. [40]

    Karypis G and Kumar V 1998 Multilevelk-way partitioning scheme for irregular graphs Journal of Parallel and Distributed Computing4896–129 ISSN 0743-7315 URL https://www.sciencedirect.com/science/article/pii/S0743731597914040 30

    work page 1998