{"id":"5955f92d-a0ea-4625-85c4-eea4d98ddee1","arxiv_id":"1908.08519","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Unfolding is formulated as a QUBO optimization and solved on a D-Wave quantum annealer, with results comparable to classical unfolding in a 5-bin toy model.","lead":"This paper shows how unfolding, the statistical correction of detector smearing in particle physics, can be encoded as an optimization problem that runs on a quantum annealer. The authors demonstrate it on a 5-bin toy model and compare the D-Wave quantum processor with classical algorithms.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The QUBO objective in Eq. 2.3 is an unweighted least-squares loss, not the Poisson likelihood from Eq. 1.6; calling this a Gaussian approximation is incorrect for count data with unequal variances.","rationale":"The reader's verdict is reasonable, but its weakest_assumption misses a more immediate problem. The QUBO algebra is internally consistent for the unweighted L2 objective; the derivation in the appendix checks out. The issue is the mapping from the stated Poisson likelihood to that objective: an unweighted L2 norm is not a Gaussian approximation to Poisson data unless all expected counts are equal, which the toy clearly violates. This is a correctness risk attached directly to the central claim that the implementation is likelihood-based. I still would not reject: the method can be fixed by including per-bin weights, and the paper is explicitly a proof-of-concept. The response-matrix uncertainty and 5-bin scale concerns raised by the reader are real but are scope limitations that the authors acknowledge; the loss-function mismatch is an unacknowledged internal inconsistency. Hence the CONDITIONAL verdict stands, with the additional condition that the paper either implement the weighted objective or weaken the 'likelihood-based' language.","tokens_in":10250,"tokens_out":13119,"duration_ms":129098,"concrete_test":"Re-run the 5-bin peaked toy from Fig. 2 with Poisson-sampled pseudo-data (drawing d_i from Pois((R x_true)_i)) and compare three unfoldings with the same lambda: (i) Eq. 2.3 as published; (ii) the same QUBO after rescaling rows of R and entries of d by 1/sqrt(mu_i) (iterated or using d_i as an estimate); (iii) a direct continuous or brute-force maximization of the Poisson log-likelihood Eq. 1.6. If (i) departs from (iii) by more than the reported statistical uncertainty in any low-count bin while (ii) does not, then Eq. 2.5 does not encode the claimed likelihood and the abstract's agreement claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 1.6 defines the unfolding likelihood as a product of Poisson terms with a Tikhonov prior exp(-beta rho). Section 2 replaces the Poisson factor by ||R x - d||^2 (Eq. 2.3), described as a Gaussian approximation. For Poisson counts, the Gaussian log-likelihood is proportional to -sum_i (d_i - mu_i)^2 / mu_i (or, in a Neyman variant, / d_i), not the unweighted sum used here. Thus the QUBO coefficients in Eqs. 6.11-6.12 encode a different statistical model from the one advertised: bins with mu_i around 2 and bins with mu_i around 14 (as in Fig. 2) receive equal weight. This can bias the unfolded spectrum in low-statistics bins and invalidates the abstract's 'likelihood-based' description. It is not a scope issue like the perfectly known response matrix; it is an internal mismatch between Eq. 1.6 and Eq. 2.3. A variance-weighted version, obtained by premultiplying R and d by the inverse square root of the covariance, would still be QUBO-representable, so the proof-of-concept could be repaired, but as written the central claim is not demonstrated for the stated likelihood.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a QUBO/Ising formulation of regularized unfolding and demonstrates it on a D-Wave quantum annealer. The authors define a Poisson-likelihood unfolding problem with Tikhonov regularization (Eq. 1.6), replace the Poisson term by the squared L2 norm (Eq. 2.3), derive binary encodings with offset and scale parameters, provide explicit QUBO weights in the appendix (Eqs. 6.11-6.12), and extend the formulation to shape-type nuisance parameters (Eq. 3.4). They test the method on a 5-bin toy model with peaked and steeply falling spectra, comparing CPU simulated annealing, two D-Wave QPUs, a hybrid solver, and D'Agostini iterative Bayesian unfolding. The central claim is that likelihood-based regularized unfolding can be implemented on a quantum annealer and achieves very good agreement with the true distribution.","tokens_in":10554,"tokens_out":8895,"duration_ms":86140,"significance":"The proof-of-concept is timely and contains genuinely useful elements: an explicit algebraic mapping, open-source code, and validation of the QUBO algebra by simulated annealing and brute-force checks on a small problem. If the formulation is corrected, it could motivate hybrid quantum-classical solvers for unfolding. As it stands, however, the paper's central statistical claim is not established: the objective in Eq. (2.3) is not the Gaussian approximation to the Poisson likelihood in Eq. (1.6), and the QPU results in Fig. 3 do not support the abstract's 'very good agreement' for the falling spectrum. The 5-bin, perfectly known response matrix test also leaves the scaling discussion in Section 6 speculative.","major_comments":[{"comment":"The replacement of the Poisson term by ||Rx-d||^2 is described as taking the Gaussian approximation, but for Poisson bin counts with expectation mu_i the Gaussian log-likelihood is proportional to -sum_i (d_i-mu_i)^2/mu_i (or, in the Neyman variant, /d_i), not the unweighted residual sum of squares. Consequently, the QUBO coefficients in Eqs. (6.11)-(6.12) minimize a different statistical objective from the likelihood advertised in Eq. (1.6); bins with low expected counts and bins with high expected counts in Fig. 2 and Fig. 3 receive equal weight. This is an internal mismatch, not merely a scope limitation, because it changes the optimum when counts vary by orders of magnitude. It is repairable: premultiplying R and d by the inverse square root of the Poisson covariance preserves QUBO form.","section":"Section 2, Eq. (2.3); Section 1, Eq. (1.6)"},{"comment":"The abstract's 'very good agreement' is not supported by the QPU results for the steeply falling spectrum. Section 5 itself states that larger lambda values produce 'flatter and flatter solutions which do not agree well with the truth level,' and Fig. 3 shows the QPU lambda=1 point substantially below the true distribution in every bin. The abstract should be qualified to refer to the peaked-spectrum case with appropriately chosen regularization, or to the CPU/hybrid results only.","section":"Abstract; Section 5, Fig. 3"},{"comment":"The representable solution space is fixed by the user-supplied offset alpha_i and scaling beta_i, and the appendix acknowledges that these are chosen by a heuristic. If the chosen range does not contain the true bin contents, the ground state of the QUBO cannot be the desired unfolding; no criterion or automatic update rule is given. Since the central claim is that this QUBO encodes the unfolding problem, the method's validity is conditional on this external choice. At minimum, the paper should state the ranges used in the toy and explain why they are guaranteed to bracket the truth.","section":"Appendix, Eq. (6.8); Section 5"}],"minor_comments":[{"comment":"As typeset, the smoothness penalty reads (theta_{j+1}+theta_{j-1})^2, which is not the discrete second derivative used in Eq. (6.1). Please correct to (theta_{j+1}-2 theta_j+theta_{j-1})^2 or clarify the intended penalty.","section":"Eq. (1.9)"},{"comment":"The definition of S x-tilde is hard to parse; it should be stated as an explicit matrix S multiplying the extended vector x-tilde. The present typesetting makes the penalty term look like a product of two vectors.","section":"Eq. (3.3)"},{"comment":"The QPU results are reported as averages over 20 runs but no error bars are shown; a table of means and standard deviations, or error bars in the figures, would make the comparison with the truth-level distribution quantitatively assessable.","section":"Figures 3-5"},{"comment":"No quantitative goodness-of-fit metric (e.g., chi2/ndf) is given for any of the comparisons; statements such as 'agreement always in the order of about one standard deviation' in Fig. 2 are therefore not testable.","section":"Section 5"},{"comment":"The paper does not state the D-Wave software release, embedding parameters, or the exact simulated-annealing settings; adding these would improve reproducibility.","section":"Section 5 and appendix"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable proof-of-concept, and the QUBO derivation is the main contribution. The unweighted least-squares issue in Eq. (2.3) is the central problem; if the authors either rewrite the method as least-squares unfolding (with the paper's claims adjusted accordingly) or implement variance weighting and rerun the toy studies, the result could become publishable. The 'very good agreement' wording should also be made quantitative. I recommend major revision rather than rejection because the fix is local to the objective and the experimental methodology can be rerun on the same toy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a legitimate proof-of-concept. The authors are the first to carry a Tikhonov-regularized unfolding problem into QUBO form and run it on a D-Wave annealer, and they release the code. The appendix derivation is algebraically correct for what it minimizes, and the CPU simulated-annealing results confirm that the QUBO objective recovers the truth in their toy model when the regularization is tuned. The extension to nuisance parameters is a nice touch, and the hybrid classical-quantum discussion is sensible. The citation pattern is also reasonable, building on ToQ.jl and standard HEP unfolding tools.\n\nThe soft spot is real and central: the paper calls this 'likelihood-based' and writes a Poisson likelihood in Eq. 1.6, but the objective actually minimized in Eq. 2.3 is an unweighted L2 loss, ||R x - d||^2 + lambda ||D x||^2. That is not the Gaussian approximation to the Poisson log-likelihood for count data with unequal variances. For bins with small expectations, unweighted least squares overweights low-statistics bins and biases the unfolded spectrum. The fix is straightforward—premultiply by the inverse square root of the covariance, which stays QUBO-compatible—but as written, the statistical model advertised is not the one implemented.\n\nThe abstract also overstates the agreement. In the peaked spectrum, the unfolded points are within about one sigma. For the steeply falling spectrum, the unregularized solution agrees well, but Fig. 3 shows that increasing lambda flattens the result and the agreement degrades visibly. The QPU runs, by their own admission, are less accurate due to hardware limitations. So 'very good agreement' is true at best for a narrow setting.\n\nOther limitations are minor given the scope: the response matrix is assumed perfectly known (standard for a first test but not for LHC measurements), the binary encoding offsets are heuristic, and the scaling argument for hundreds of nuisance parameters is untested. These are addressable in follow-up work.\n\nWho gets value: anyone working on NISQ applications in HEP, or on QUBO mappings for inverse problems. It is not a demonstration of practical quantum advantage, and the statistical flaw needs to be corrected. I would send it to peer review—it deserves referee time—but the authors need to replace 'likelihood-based' with 'regularized least squares' or variance-weight the objective, and temper the abstract.","headline":"First QUBO map of regularized unfolding on a quantum annealer, but the 'likelihood-based' claim is wrong and the agreement is oversold.","tokens_in":11076,"tokens_out":3539,"would_cite":false,"duration_ms":37051,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Regularized unfolding of detector-smeared spectra can be encoded as a QUBO problem and solved on a quantum annealer, with a five-bin toy model recovering the true distribution and matching the standard iterative Bayesian method.","keywords":["unfolding","quantum annealing","QUBO","Ising model","Tikhonov regularization","inverse problems","nuisance parameters","high-energy physics"],"falsifier":"Rerun the five-bin toy with each response-matrix entry perturbed by a few percent (within plausible simulation uncertainties) and compare the QUBO-unfolded spectrum to the truth; if the result moves outside the quoted one-standard-deviation band, the known-response assumption is the limiting factor. Alternatively, enumerate all binary assignments of the same QUBO and check whether the annealer's output is the global minimum for each $\\lambda$ and $\\gamma$ choice.","tokens_in":10055,"feed_emoji":"⚛️","tokens_out":10062,"duration_ms":87672,"temperature":0.7,"pith_summary":"The paper claims that the routine high-energy-physics task of unfolding — correcting measured distributions for detector smearing and inefficiency — can be cast as a quadratic unconstrained binary optimization (QUBO) problem, which is exactly the form a programmable quantum annealer solves. The authors translate the regularized least-squares unfolding objective into a binary quadratic cost and test it on a five-bin toy model with sizeable bin-to-bin migrations. Both a classical simulated-annealing solver and the quantum annealing hardware recover the truth distribution within about one standard deviation, matching the iterative Bayesian unfolding method commonly used at the LHC. The formulation is then extended to include nuisance parameters that represent systematic uncertainties, with a penalty on their deviation from zero. If correct, this establishes a proof of principle that a mainstream precision-measurement procedure can be executed on quantum annealing hardware.","feed_headline":"Quantum annealer recovers smeared particle spectra","feed_subtitle":"A QUBO encoding of regularized unfolding matches the LHC baseline on a toy spectrum.","key_machinery":"The load-bearing object is the QUBO matrix built from the regularized unfolding objective. The real-valued truth-level spectrum $x$ is replaced by a fixed-point binary expansion with per-bin offsets and scalings, $x_i = \\alpha_i + \\beta_i \\sum_{j=0}^{n-1} 2^j q_{n i + j}$, so that the continuous objective $\\Vert Rx-d\\Vert^2 + \\lambda \\Vert Dx\\Vert^2$ collapses into a quadratic form $y = q^T C q$ over binary variables. The matrix elements follow from expanding the objective in index notation: for $W = R^T R + \\lambda D^T D$, the off-diagonal QUBO weights are $c_{ab} = 2 W_{jk} \\beta_{ja} \\beta_{kb}$ and the diagonal weights absorb the offset terms, giving $c_{aa} = 2 W_{jk} \\alpha_k \\beta_{ja} + W_{jk} \\beta_{ja}\\beta_{ka} - 2 R_{ij} d_i \\beta_{ja}$. In the systematic-uncertainty extension, $W$ is replaced by $R^T R + \\lambda D^T D + \\gamma S^T S$ over an extended vector that interleaves the spectrum bins and the nuisance-parameter strengths. This mapping is what turns a hardware Ising annealer into an unfolding engine, with the classical side managing the encoding range.","core_discovery":"The central claim is that likelihood-based regularized unfolding can be rewritten as a quadratic unconstrained binary optimization in a way that preserves the statistical content of the problem, and that the resulting QUBO can be solved on a quantum annealer to recover the underlying spectrum. Concretely, the objective $\\Vert Rx-d\\Vert^2 + \\lambda \\Vert Dx\\Vert^2$ — where $R$ is the detector response matrix, $d$ the observed counts, $D$ the discrete Laplacian used as a Tikhonov smoothing operator, and $\\lambda$ the regularization strength — is converted into a quadratic form over binary variables by expanding each truth-level bin as $x_i = \\alpha_i + \\beta_i \\sum_{j=0}^{n-1} 2^j q_{n i + j}$. The linear terms of the resulting QUBO come from the data agreement, the quadratic terms from the smoothness prior, and the offsets and scalings $\\alpha_i, \\beta_i$ are chosen on the classical side. On a five-bin toy spectrum, solutions from simulated annealing and from the quantum processor agree with the truth within one standard deviation and are comparable to the iterative Bayesian benchmark; the unregularized version also handles a steeply falling spectrum, while too-large $\\lambda$ flattens it. The authors further show that appending the strength of a shape systematic as an extra binary-encoded variable, penalized by a term $\\gamma \\Vert S \\tilde{x}\\Vert^2$, allows the systematic shift to be estimated in the same optimization, and they argue that a hybrid classical-quantum search is the route to larger problems.","pith_inferences":["A natural next test, not performed here, is to relax the perfect-response-matrix assumption: perturbing $R$ within its simulated uncertainties and rerunning the five-bin toy would show how response errors propagate through the QUBO encoding, which the current error bars do not cover.","Because the QUBO form is agnostic to the physics, the same construction could be applied to other inverse problems with a response matrix and a smoothness prior, such as image deconvolution or astrophysical source reconstruction.","The paper notes that the quantum processor occasionally misses the brute-force ground state; a systematic comparison of annealer outputs against brute-force enumeration on the same five-bin instances would separate hardware and embedding errors from encoding errors.","The offsets and scalings $\\alpha_i, \\beta_i$ are chosen heuristically and bound the representable spectrum; an adaptive procedure that tests whether the optimum saturates the encoding range would make the method more robust."],"forward_implications":["Unfolding, a frequent step in LHC precision measurements, can in principle be delegated to a quantum annealer without changing the statistical formulation beyond the Gaussian approximation of the Poisson likelihood.","The same QUBO encoding handles nuisance parameters natively: systematic strengths become additional binary variables with a tunable penalty, so unfolding and systematic-shift estimation happen in a single optimization.","On the toy model, the quantum-annealer result agrees with the truth within one standard deviation, and a hybrid classical-quantum run reproduces the simulated-annealing solution, indicating that hybrid search is the practical mode for larger problems.","Because the final object is a standard QUBO, any current or future quantum or classical QUBO solver can substitute for the hardware without reformulating the physics.","The method inherits the known problem-dependence of Tikhonov regularization: the regularization strength $\\lambda$ must be tuned per measurement, since large values over-smooth steeply falling spectra."],"supporting_citations":[{"why":"supplies the approach for rewriting the continuous unfolding objective as a QUBO via binary expansion, which the paper follows directly.","marker":"[21]"},{"why":"provides the regularized unfolding algorithm (Tikhonov-style smoothing) whose objective the QUBO formulation mirrors.","marker":"[18]"},{"why":"supplies the iterative Bayesian unfolding method used as the baseline comparison.","marker":"[19]"},{"why":"provides the unfolding package whose covariance-matrix implementation computes the benchmark uncertainties.","marker":"[30]"},{"why":"the simulated-annealing sampler used as the classical CPU solver for the QUBO problems.","marker":"[28]"},{"why":"reports the quantum annealing hardware with manufactured spins on which the quantum executions ran.","marker":"[6]"}],"fun_headline_variants":["Quantum annealer solves regularized unfolding inverse problem","Unfolding as QUBO: quantum annealing recovers smeared spectra","Quantum annealing matches LHC unfolding baseline on toy data","Rewriting unfolding as QUBO enables quantum annealing solution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the detector response matrix is perfectly known: the paper applies the same matrix to generate both the pseudo-data and the reference truth, so it never tests how errors in the response matrix propagate into the unfolded spectrum.","fun_headline_variants_meta":{"raw":{"variants":["Quantum annealer solves regularized unfolding inverse problem","Unfolding as QUBO: quantum annealing recovers smeared spectra","Quantum annealing matches LHC unfolding baseline on toy data","Rewriting unfolding as QUBO enables quantum annealing solution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1482,"prompt_tokens":1009,"completion_tokens":473,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":405}},"tokens_in":625,"tokens_out":473,"duration_ms":5262,"temperature":1.0,"reasoning_tokens":405,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:36:51.017619+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the five-bin toy with each response-matrix entry perturbed by a few percent (within plausible simulation uncertainties) and compare the QUBO-unfolded spectrum to the truth; if the result moves outside the quoted one-standard-deviation band, the known-response assumption is the limiting factor. Alternatively, enumerate all binary assignments of the same QUBO and check whether the annealer's output is the global minimum for each $\\lambda$ and $\\gamma$ choice.","supporting_citations":[{"cited_title":"ToQ.jl: A high-level programming language for D-Wave machines based on Julia,","cited_arxiv_id":null,"evidence_quote":"supplies the approach for rewriting the continuous unfolding objective as a QUBO via binary expansion, which the paper follows directly."},{"cited_title":"TUnfold, an algorithm for correcting migration effects in high energy physics,","cited_arxiv_id":null,"evidence_quote":"provides the regularized unfolding algorithm (Tikhonov-style smoothing) whose objective the QUBO formulation mirrors."},{"cited_title":"A multidimensional unfolding method based on Bayes’ theorem,","cited_arxiv_id":null,"evidence_quote":"supplies the iterative Bayesian unfolding method used as the baseline comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the unfolding package whose covariance-matrix implementation computes the benchmark uncertainties."},{"cited_title":"Simulated annealing sampler","cited_arxiv_id":null,"evidence_quote":"the simulated-annealing sampler used as the classical CPU solver for the QUBO problems."},{"cited_title":"Quantum annealing with manufactured spins,","cited_arxiv_id":null,"evidence_quote":"reports the quantum annealing hardware with manufactured spins on which the quantum executions ran."}],"review_version":1}