{"id":"4a9b8af6-6d9c-43e3-b5e7-7fd216375b78","arxiv_id":"2412.00601","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A proof-of-concept that maps circle packing for chromatography to a maximum independent set problem and runs QAOA on 18 qubits, with resource estimates for harder sphere packing variants.","lead":"The authors map protein chromatography column packing to sphere packing and solve the simplest 2D version on a 20-qubit quantum computer, while sketching harder 3D versions and estimating quantum resources. The work is a proof-of-concept roadmap, not a demonstration of quantum advantage.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section IV counts only per-layer qubits and CNOTs; total QAOA cost is p times those counts, and no bound or scaling evidence for p is given. The polynomial-resource claim is therefore unproven.","rationale":"The reader's CONDITIONAL verdict is appropriate. The paper contains a genuine small-scale QAOA demonstration on IQM Garnet, a reproducible resource-counting framework, and an interesting parameter-concentration observation; those parts stand. What would have to be true for the central quantum-advantage claim is that the total QAOA circuit cost grows polynomially with problem size. The text supplies qubit and per-layer CNOT polynomials, but the total cost also includes the number of layers p. Since QAOA only guarantees optimality in the p→∞ limit, the paper must either bound p or present scaling evidence; it does neither. The only empirical p data are for one 18-node instance and p≤5, far too little to infer asymptotic scaling. I also note that the classical exponential claim rests on worst-case NP-hardness and the geometric instances may be easier (unit-disk MIS has PTAS), but this does not change the primary concern. A targeted noiseless depth-scaling study, or a known lower-bound argument for QAOA on MIS, would settle whether the polynomial-resource claim survives. Until then, the correct verdict remains CONDITIONAL, matching the reader.","tokens_in":11990,"tokens_out":5489,"duration_ms":60829,"concrete_test":"Use the paper's noiseless simulator to measure, on the same grid family with N=18, 32, 50, 72 qubits, the minimal p at which the optimized QAOA probability of the true MIS exceeds a fixed threshold; if p(N) grows superpolynomially (or cannot be found below the noise-limited depth), the polynomial resource claim is unsupported. This directly supplies the missing p(n) factor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV's resource counts are per QAOA layer: Eq. (15) bounds CNOTs for one e^{-iHCγ} layer and Eqs. (14),(17) bound qubit counts, but the total quantum cost is p times the per-layer count. The paper never bounds p(n), the number of layers needed to reach a good packing; it cites only the p→∞ convergence guarantee in Sec. I.B and reports p=1,3,5 on a single 18-node graph (Figs. 11–12), with parameter concentration results (Figs. 13–14) about transferring optimized parameter values, not about the required depth on larger instances. Thus the 'quantum resources only scale polynomially' statement in Sec. IV is missing the most problem-size-dependent factor. The accompanying classical claim is also overbroad: NP-hardness is a worst-case statement, and the grid graphs constructed in Sec. II are unit-disk-like, for which classical polynomial-time approximation schemes exist, so 'exponential for this problem' would need instance-specific evidence. The decisive gap, however, is p(n): without a polynomial bound on p, total gate count cannot be claimed polynomial.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper models column chromatography as bounded sphere packing and treats three increasing levels of complexity: homogeneous circle packing, heterogeneous circle packing, and heterogeneous sphere packing. Homogeneous circle packing is recast as a maximum independent set problem on a grid graph and solved with QAOA on IQM's Garnet device (18 qubits), with hyperparameter selection, compilation, noisy simulation, hardware execution, and a parameter concentration study. Heterogeneous packing is formulated through two Hamiltonian constructions (\"first quantization\" and \"second quantization\"), and Section IV gives resource estimates intended to show that the quantum resources scale polynomially with problem size.","tokens_in":12241,"tokens_out":5941,"duration_ms":57339,"significance":"If the polynomial-scaling claim were established, the paper would be a valuable end-to-end demonstration of a concrete industrial optimization use case for QAOA, combining a falsifiable hardware prediction with transferable techniques such as noise-aware depth selection and subgraph parameter transfer. The hardware experiment is the strongest part: the noisy simulation correctly predicted p=3 as the best depth, and the device returned the optimal independent set as the most probable state for p=1, 3, and 5. The resource-scaling analysis, however, is incomplete in a way that directly affects the paper's central quantum-advantage claim, because it never bounds the QAOA layer count p as a function of problem size.","major_comments":[{"comment":"The resource counts in Eqs. (14)-(19) are per QAOA layer, but the total quantum cost is p times these counts. The paper never bounds p(n), the number of layers needed to obtain a good packing. The p to infinity convergence guarantee in Sec. I.B and the hardware results for p=1,3,5 on a single 18-node graph (Figs. 11-12) do not establish any scaling for p on larger instances, and the parameter concentration results (Figs. 13-14) concern parameter reuse, not required circuit depth. The statement in Sec. IV that \"the quantum resources for this problem only scale polynomially with increased problem size\" is therefore unproven, and this is the load-bearing claim of the paper.","section":"Section IV, Eqs. (14)-(19)"},{"comment":"The classical-complexity comparison is overbroad. The paper argues that classical computation is exponential because MIS is NP-hard in the worst case, but the graphs constructed in Sec. II are grid graphs with unit-disk-like edges, for which polynomial-time approximation schemes exist. The claim of exponential classical difficulty for this particular instance family and for the heterogeneous packing formulations needs instance-specific evidence (or a more careful statement of the complexity assumption) rather than an appeal to worst-case NP-hardness.","section":"Sections II.B and IV"},{"comment":"The displayed resource bounds contain algebraic errors. In Eq. (15), the chain should give |R|^2 q^d (r_m q / R_b)^d, not |R| q^d (r_m q / R_b)^d; a factor of |R| is missing. In Eq. (19), 2^{2 ceil(log2(|R|+1))} is of order |R|^2, so the coefficient is of order |R|^4, not |R|^3. These errors do not change the polynomial degree in q, but they must be corrected because the resource estimates are the quantitative basis of the paper's scaling claims.","section":"Eqs. (15) and (19)"}],"minor_comments":[{"comment":"The identity operator is defined as I = (|0><0|+|1><1|)/2, but the identity should be |0><0|+|1><1|; as written, the relation x_v = (I-Z_v)/2 is inconsistent.","section":"Eq. (3)"},{"comment":"The text says \"we train the parameters of both the subgraph QAOA circuit (Csub) and the full graph QAOA circuit (Csub)\"; the second circuit name should be Cfull.","section":"Section II.C.7"},{"comment":"The y-axis labels appear as 10^3, 10^2, 10^1, but the probabilities are manifestly below 1; the labels should be 10^{-3}, 10^{-2}, 10^{-1}.","section":"Figure 11"},{"comment":"The comparison between the two formulations says the second-quantization formulation trades an increase in CNOT scaling for an exponential decrease in qubits, but the qubit counts in Eqs. (14) and (17) show that second quantization uses more qubits (|R| q^d) than first quantization (log|R| q^d), not fewer; the sentence should be rephrased.","section":"Section IV, final paragraph"}],"recommendation":"major_revision","confidential_remarks":"The hardware results are solid and the paper is a credible application-driven study, but the headline quantum-advantage claim rests on an unproven assumption about QAOA layer-count scaling. I would ask the authors to either supply a bound or clearly qualify the polynomial-resource claim, and to correct the resource-counting errors, before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe hardware experiment is real and the formulations are reasonable, but the central scaling claim—polynomial quantum resources vs classical exponential—skips the QAOA depth p and is unsupported as written.\n\nWhat is actually new: an 18-qubit QAOA run for a homogeneous circle-packing MIS instance on IQM Garnet, with careful compilation (edge coloring, depth-4), hyperparameter tuning for λ, and a noisy simulation based on device calibration that correctly picked p=3 as the best depth. The hardware results show the optimal state as the most probable for p=1,3,5, and the parameter-concentration comparison is a legitimate empirical data point. The two heterogeneous formulations ('First' and 'Second Quantization') are modest but clearly explained extensions, and the per-layer resource counts in Section IV are roughly correct.\n\nThe soft spot is Section IV's conclusion. They count qubits and CNOTs per QAOA layer, but the total cost is p times those counts. Nowhere do they bound how p grows with problem size; the only evidence is one 18-node graph at p=1,3,5. So 'quantum resources only scale polynomially' is missing exactly the most problem-size-dependent factor. The classical side is also overstated: NP-hardness is worst-case, and the grid graphs are unit-disk-like, where polynomial-time approximation schemes exist. The exponential-vs-polynomial contrast is not established.\n\nThe chromatography motivation is thin—packing density is never connected to separation performance—but that is a framing gap, not a technical flaw. The paper is honest about NISQ noise and does not oversell the hardware result. One smaller issue: the First Quantization bound in Eq. (19) is loose and the |R|^3 step is not derived cleanly.\n\nWho this is for: applied QAOA groups and biopharma process engineers looking for near-term quantum roadmaps. It deserves a serious referee; the correct outcome is major revision, not desk rejection. The authors should either produce evidence or analysis on p scaling, or explicitly reframe the polynomial claim as conditional on an unproven depth assumption. I'd send it to review and push hard on that point.","headline":"A real hardware demo and reasonable formulations, but the scaling claim skips QAOA depth p and overstates the classical comparison.","tokens_in":12749,"tokens_out":5744,"would_cite":false,"duration_ms":68981,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["52C17","90C27","81P68"],"pacs":[],"model":"deepseek-v4-flash","headline":"Bounded sphere packing for chromatography can be cast as a maximum independent set and solved with QAOA on quantum hardware.","keywords":["Quantum Approximate Optimization Algorithm","maximum independent set","sphere packing","column chromatography","heterogeneous packing","resource estimation","parameter concentration","biopharmaceutical manufacturing"],"falsifier":"Run exact or near-exact simulations of the QAOA circuit for the homogeneous circle-packing graphs at increasing grid sizes (for example 18, 40, 80, and 160 nodes) and measure the number of QAOA layers $p$ needed to reach a fixed success probability; if $p$ grows faster than any polynomial, the resource-scaling claim is falsified.","tokens_in":1697,"feed_emoji":"🧪","tokens_out":2320,"duration_ms":105215,"temperature":0.7,"pith_summary":"Column chromatography, the workhorse of biopharmaceutical protein separation, is usually modeled with abstracted particle-level kinetics. This paper proposes replacing that abstraction with a literal sphere-packing model of the resin: circles in a circle for the two-dimensional cross-section, spheres in a cylinder for three dimensions. The homogeneous circle-packing version is discretized onto a grid and cast as a maximum independent set, then solved with the Quantum Approximate Optimization Algorithm on a 20-qubit quantum processor. Heterogeneous circle packing is formulated as a constrained graph optimization, solved classically, and extended to heterogeneous sphere packing; resource counts suggest quantum resources grow polynomially with problem size while classical effort for these NP-hard packings grows exponentially. The value, if the scaling claim holds, is a mechanistic model that captures particle-level adsorption phenomena and a concrete near-term use for quantum optimization.","feed_headline":"Packing chromatography columns: quantum resources scale polynomially","feed_subtitle":"A sphere-packing model of protein separation is recast as QAOA and solved on a 20-qubit chip.","key_machinery":"The carrying mechanism is a graph encoding of continuous packing: a grid of candidate centers inside the boundary circle becomes the vertex set $V$, edges connect pairs whose disks would overlap, and a valid packing is exactly a maximum independent set. The objective, minimizing empty nodes, and the non-overlap constraints are combined into a single QAOA cost Hamiltonian $H_C = \\frac{1}{2}\\sum_{v\\in V} Z_v + \\frac{\\lambda}{4}\\sum_{v,w\\in E}(Z_v Z_w - Z_v - Z_w)$, with the Lagrange multiplier $\\lambda$ treated as a tunable hyperparameter. For heterogeneous radii, the 'first quantization' formulation encodes each node's radius choice in $\\lceil \\log_2(|R|+1)\\rceil$ qubits while the 'second quantization' formulation assigns one qubit per radius-location pair, and the same edge criterion $|v-w| \\le r+s$ builds the constraint Hamiltonian.","core_discovery":"The central discovery is a reduction: bounded heterogeneous sphere packing, the geometry of a packed chromatography column, can be discretized and expressed as a graph optimization whose quantum cost Hamiltonian is compiled into QAOA circuits. For the homogeneous case, the reduction to maximum independent set is exact on the grid, and the resulting 18-node instance produced the optimal 12-circle packing on real hardware, reaching a packing density of 0.68 versus the proven optimum of 0.74. The heterogeneous cases are written in two complementary quantum formulations, one with logarithmic qubit compression and one with a qubit per radius-location pair, and both are bounded by low-degree polynomial resource counts in the discretization points per side and the number of allowed radii. The paper therefore claims a pathway to quantum advantage for a real biopharmaceutical problem: classical solvers scale exponentially because the underlying maximum independent set is NP-hard, while the quantum resources for each QAOA layer scale polynomially.","pith_inferences":["The resource estimate counts qubits and CNOT gates per QAOA layer but never bounds the number of layers $p$ needed to reach a good packing; if $p$ must grow faster than polynomially, the claimed polynomial total resource scaling would fail, and testing $p$-scaling on grid graphs of increasing size would settle this.","The physical model assumes a regular grid discretization of space, whereas real resin beds are random packed and typically achieve densities around 0.64 for monodisperse spheres; a direct comparison to experimental packed-bed porosity data would be a natural testable extension.","The same graph and maximum-independent-set reduction applies to other bounded geometric packing problems such as bin packing or container loading, so a quantum advantage here would carry over to any such NP-hard instance.","The heterogeneous second-quantization Hamiltonian is built from pairwise overlap penalties, so it could also be attacked by classical annealing or other variational circuits; benchmarking QAOA against classical local search at matched instance sizes would clarify where the quantum advantage claim actually stands."],"forward_implications":["If the polynomial resource scaling holds, a sufficiently large chromatography column modeled with this discretization can be optimized on a quantum computer with a qubit and gate count that grows polynomially in grid resolution and the number of radii, whereas classical approximate solvers face an NP-hard search.","Parameter concentration evidence means QAOA parameters trained on a small subgraph of a packing instance can be transferred to the full instance, so large untrainable problems remain accessible by scaling up a small trained circuit.","The same integer program extends from circles in a circle to spheres in a cylinder by stacking two-dimensional slices and adding inter-slice edges, so the homogeneous hardware result is already a template for the three-dimensional model.","The first- versus second-quantization tradeoff offers a resource tunable: use logarithmic qubits with higher-order terms, or more qubits with lower CNOT-degree, depending on which noise source dominates a given device."],"supporting_citations":[{"why":"Supplies the QAOA ansatz and the guarantee that the final state approaches the optimal state as the number of layers grows.","marker":"[17]"},{"why":"Provides the 20-qubit processor's coupling map and noise characteristics, which determine the 18-node hardware graph and the noisy-simulation parameters.","marker":"[18]"},{"why":"Gives the proven optimal density for 12 congruent circles in a circle, the benchmark against which the hardware result (density 0.68) is compared.","marker":"[19]"},{"why":"Frames the noisy intermediate-scale regime that motivates the depth-versus-noise tradeoff and the noisy simulations used to pick p=3.","marker":"[20]"},{"why":"Supplies the parameter-concentration principle used to argue that parameters trained on a small subgraph transfer to the full problem.","marker":"[22]"},{"why":"Provides the QUBO solver used to produce the classical heterogeneous circle packing simulation.","marker":"[23]"}],"fun_headline_variants":["QAOA packs chromatography spheres with polynomial resources","Sphere packing reduced to QAOA: polynomial resource scaling","Quantum sphere packing: chromatography's polynomial shortcut","Polynomial quantum path to efficient chromatography packing"],"cache_read_input_tokens":14976,"weakest_assumption_plain":"The claim that quantum resources stay polynomial assumes the number of QAOA rounds needed to find a good packing grows only as a polynomial of the problem size; the paper never proves this.","fun_headline_variants_meta":{"raw":{"variants":["QAOA packs chromatography spheres with polynomial resources","Sphere packing reduced to QAOA: polynomial resource scaling","Quantum sphere packing: chromatography's polynomial shortcut","Polynomial quantum path to efficient chromatography packing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001225,"raw_usage":{"total_tokens":5016,"prompt_tokens":909,"completion_tokens":4107,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":4050}},"tokens_in":525,"tokens_out":4107,"duration_ms":138851,"temperature":1.0,"reasoning_tokens":4050,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:11:50.849850+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run exact or near-exact simulations of the QAOA circuit for the homogeneous circle-packing graphs at increasing grid sizes (for example 18, 40, 80, and 160 nodes) and measure the number of QAOA layers $p$ needed to reach a fixed success probability; if $p$ grows faster than any polynomial, the resource-scaling claim is falsified.","supporting_citations":[{"cited_title":"Espinoza, S","cited_arxiv_id":null,"evidence_quote":"Supplies the QAOA ansatz and the guarantee that the final state approaches the optimal state as the number of layers grows."},{"cited_title":"Hales, M","cited_arxiv_id":null,"evidence_quote":"Frames the noisy intermediate-scale regime that motivates the depth-versus-noise tradeoff and the noisy simulations used to pick p=3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the parameter-concentration principle used to argue that parameters trained on a small subgraph transfer to the full problem."}],"review_version":1}