{"id":"9b75025d-1be7-4d9e-8c1f-142c88d96483","arxiv_id":"2501.11576","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper introduces a Riemannian gradient descent method, with a derived gradient formula, to compute lower bounds on the Holevo capacity of general quantum channels, with numerical demonstrations.","lead":"This paper proposes a Riemannian gradient descent algorithm to compute lower bounds on the Holevo capacity of any quantum channel, by optimizing over pure input states and their probabilities on a product manifold. It reports large speedups over existing methods on depolarizing, classical-quantum, and entanglement-breaking channels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's probability-component gradient does not satisfy the simplex tangent condition required by the paper's own metric (4), so the Riemannian descent update and the O(1/ε²) convergence guarantee of Algorithm 1 do not follow as stated.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing flaw: the probability-component gradient in Proposition 1 is not tangent under the metric defined in Eq. (4). I verified the algebra: summing q_j over the d_A² probability components leaves the residual d_A² − 1 − Σ_j D(σ_j‖σ) + Σ_k p_k D(σ_k‖σ), which is generically nonzero. The stated q also cannot be obtained by the Euclidean projection in Eq. (4), since that projection would subtract a constant mean from the partial derivative and would not introduce a p_j(Σ_k p_k D(σ_k‖σ) − 1) term. Because the convergence theorem cited in [32] requires a genuine Riemannian gradient, Algorithm 1's theoretical guarantee is not established. I do not see a separate concern that outweighs this one. The numerical results may still suggest practical utility, and the error is likely repairable by replacing q with its Euclidean projection, but as submitted the central argument for the convergence claim and the resulting lower bound is not supported. The code being unavailable at the referenced URL is a secondary reproducibility issue, not the primary logical defect.","tokens_in":11635,"tokens_out":7199,"duration_ms":78596,"concrete_test":"For the qubit cq channel with p_0=p_1=1/2, ρ_0=|0⟩⟨0|, ρ_1=|+⟩⟨+|, and ρ=(ρ_0+ρ_1)/2, compute D_j=D(ρ_j‖ρ). Corollary 2 gives Σ_j q_j = 1 − (D_0+D_1)/2; direct evaluation yields D_0=D_1≈0.601, so Σ_j q_j≈0.399≠0. This violates the tangency condition Σ_j q_j=0 demanded by T_pΔ. Alternatively, apply the projection Proj_p from Eq. (4) to the Euclidean derivative ∂p f_N and compare entry-by-entry with the formula in Proposition 1; any nonzero difference confirms the stated gradient is not the Riemannian gradient for the paper's metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Algorithm 1 is Riemannian gradient descent on the product manifold M_{d_A^2+1} with the Euclidean product metric (4), and hence inherits the O(1/ε²) convergence guarantee of Boumal--Absil--Cartis [32]. This requires the vector labeled gradf_N(m) to be tangent: in particular, its probability component q must satisfy Σ_j q_j = 0 because T_pΔ = {ṗ : Σ_i ṗ_i = 0}. However, the formula in Proposition 1 gives q_j = 1 − D(σ_j‖σ) + p_j(Σ_k p_k D(σ_k‖σ) − 1), whose sum over j = 0,...,d_A²−1 equals d_A² − 1 − Σ_j D(σ_j‖σ) + Σ_k p_k D(σ_k‖σ), which is not identically zero. More directly, the projection formula in Eq. (4) would subtract the mean of the Euclidean derivative, producing a vector of the form −D(σ_j‖σ) + (1/d_A²)Σ_k D(σ_k‖σ) (up to the sign convention used for ∂p f), not the stated expression with the extra p_j(Σ_k p_k D(σ_k‖σ) − 1) term. Since q is not tangent, the retraction R_p in Eq. (5) is not being applied to a valid tangent vector, the update is not Riemannian gradient descent under the stated metric, and the cited convergence theorem cannot be invoked. The claim that Algorithm 1 computes a first-order critical point within O(1/ε²) steps is therefore unsupported as written. A corrected projection may be fixable, and numerical experiments may still be informative, but the paper's theoretical guarantee and its stated lower-bound argument rest on this gradient formula.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reformulates the computation of the Holevo capacity of a quantum channel as a minimization problem over the product manifold M_{d_A^2+1} = Δ_+^{d_A^2-1} × (S^{d_A-1})^{d_A^2}, where the variables are a probability vector and a collection of pure input states. A Riemannian gradient descent (RGD) algorithm with Armijo backtracking is proposed, together with a stated O(1/ε²) convergence guarantee to first-order critical points. The authors also present numerical experiments for depolarizing, classical-quantum, entanglement-breaking, Pauli, and composed qutrit channels, reporting high accuracy and large speedups over SDP-based methods.","tokens_in":12020,"tokens_out":12326,"duration_ms":133245,"significance":"If the theoretical claims were correct, this would be a valuable contribution: a scalable, geometrically principled heuristic for lower-bounding Holevo capacities of general quantum channels, with numerical evidence of substantial improvements over existing SDP-based approaches. The numerical validation against King's analytical formula for depolarizing channels and independent SDP upper bounds is a positive feature, as is the availability of the implementation. However, the central Riemannian gradient formula for the probability component is inconsistent with the paper's own metric, and the stated convergence guarantee is therefore not established as written.","major_comments":[{"comment":"The probability component q of the claimed Riemannian gradient is not tangent to the simplex. With T_pΔ = {ṗ : Σ_j ṗ_j = 0} and the Euclidean metric (4), any tangent vector must have zero component sum. However, q_j = 1 − D(σ_j∥σ) + p_j(Σ_k p_k D(σ_k∥σ) − 1) gives Σ_j q_j = d_A^2 − 1 − Σ_j D(σ_j∥σ) + Σ_k p_k D(σ_k∥σ), which is not identically zero. The correct orthogonal projection of the Euclidean derivative e_j = 1 − D(σ_j∥σ) onto T_pΔ is e_j − (1/d_A^2)Σ_k e_k, not the stated formula. Because q is not a tangent vector, the retraction in Eq. (5) is applied to an invalid direction, the update in Algorithm 1 is not Riemannian gradient descent under the metric (4), and the O(1/ε²) convergence guarantee from [32] cannot be invoked. This invalidates the stated convergence claim for Algorithm 1.","section":"Section III, Eq. (4) and Proposition 1"},{"comment":"The smoothing N' = (1−δ)N + δD changes the objective, but the paper does not quantify the effect. The convergence theorem is stated for the Holevo cost f_N, whereas the algorithm in practice minimizes f_{N'}; a first-order critical point of f_{N'} need not be near a critical point of f_N, and a lower bound on χ(N') does not by itself give a lower bound on χ(N). The paper should either restrict the claim to channels for which N(|ψ⟩⟨ψ|) is nonsingular for all pure inputs, or add a perturbation bound showing that the smoothed problem approximates the original problem in the appropriate sense.","section":"Section III, paragraph after Algorithm 1"}],"minor_comments":[{"comment":"The summation in the last term should be Σ_j p_j D(ρ_j∥ρ); as written it uses D(ρ_x∥ρ) inside the sum, which changes the formula.","section":"Corollary 2"},{"comment":"The output line 'm(t) ∈ M_{dA+1}' should read 'm(t) ∈ M_{d_A^2+1}'.","section":"Algorithm 1, output line"},{"comment":"The symbol q is used both for the probability-component tangent vector in Eq. (4) and for the gradient component in Proposition 1; this reuse is confusing and should be avoided.","section":"Section III, Eq. (4)"},{"comment":"The indexing in Proposition 1 is inconsistent: q_j is indexed from 0 to d_A^2−1 while the state components are indexed from 1 to d_A^2; please unify.","section":"Proposition 1"},{"comment":"The analytical value h2((1+ϵ)/2) for binary cq channels uses ϵ, which is never defined in the text.","section":"Table II"},{"comment":"Proposition 3 is stated without proof and its formula is not self-evident; in particular the factors (1−q_j)|ψ_j⟩ and (1−p_i)|ψ_i⟩ mix different components of the product ensemble. A precise statement and derivation are needed before the remark on product input states can be assessed.","section":"Proposition 3"}],"recommendation":"major_revision","confidential_remarks":"The main technical error is a miscalculated projection and is repairable; however, the numerical experiments may need to be rerun with the corrected gradient, and the smoothing analysis added. I see no evidence of circular reasoning, since the validation is against independent analytical formulas and SDP bounds. If the authors can fix the gradient formula and quantify the smoothing error, the paper could become a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's core idea is good: recast the Holevo capacity as an optimization problem on a product manifold of a simplex and complex spheres, then run Riemannian gradient descent. That formulation is new for general quantum channels, and the numerical results look genuinely useful — depolarizing channels up to d=21, cq channels with |X|=100, entanglement-breaking channels, all with better speed and scale than the SDP baselines. The Euclidean partial derivatives are mostly derived correctly, and the intent to invoke Boumal–Absil–Cartis convergence is natural.\n\nThe problem is load-bearing. In Proposition 1, the probability-component q is not a tangent vector under the metric the paper defines. With the Euclidean product metric (4), the tangent space of the simplex is { ṗ : Σ ṗ_j = 0 }, and the projection (4) subtracts the mean of the Euclidean derivative. The correct Riemannian gradient component would be −D(σ_j‖σ) + (1/d_A²)Σ_k D(σ_k‖σ). Instead the paper gives q_j = 1 − D(σ_j‖σ) + p_j(Σ_k p_k D(σ_k‖σ) − 1), and the sum of those components is not identically zero. A non-tangent vector cannot be fed into the retraction (5), and the cited O(1/ε²) convergence theorem does not apply. So the central theoretical guarantee for Algorithm 1 is unsupported as written.\n\nThis is likely fixable — the projection mistake looks isolated, and with the correct projection the argument might go through. But that is exactly what the paper must establish. Two smaller soft spots: the δ-smoothing is not analyzed (f_N is not smooth, and 'sufficiently small' is doing a lot of work), and the code link says 'under polishing', so there is no artifact to inspect. These are minor compared with the gradient error.\n\nWho is this for? People computing channel capacities numerically. The approach, once corrected, could be a useful complement to SDP upper bounds. I would not cite it in its present form, but the underlying idea is not silly. A good referee can verify the tangent-space issue in a few minutes, and the authors deserve the chance to fix it. Send it to peer review, but expect the gradient to be corrected and the convergence claim re-proven.","headline":"Original Riemannian approach to Holevo capacity, but the probability-component gradient is not tangent, so the headline convergence guarantee does not follow.","tokens_in":12509,"tokens_out":4203,"would_cite":false,"duration_ms":43149,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P45","81P68","90C26","65K10"],"pacs":["03.67.-a","03.67.Hk"],"model":"deepseek-v4-flash","headline":"The paper claims that the Holevo capacity of any quantum channel can be lower-bounded efficiently by a Riemannian gradient descent on a product of a simplex and spheres.","keywords":["Holevo capacity","quantum channel","Riemannian optimization","gradient descent","classical capacity","lower bound","nonconvex optimization","quantum information theory"],"falsifier":"Compute $\\sum_j q_j$ from Proposition 1 at a random ensemble for a specific channel, say a 2-qubit depolarizing channel, and check whether the sum vanishes; if it does not, the gradient is not tangent, so the retraction step and the cited $O(1/\\epsilon^2)$ convergence do not apply.","tokens_in":11453,"feed_emoji":"⚛️","tokens_out":9016,"duration_ms":76053,"temperature":0.7,"pith_summary":"The paper tries to establish that the Holevo capacity of an arbitrary quantum channel can be usefully lower-bounded by solving a nonconvex optimization problem on a product of a probability simplex and complex unit spheres. It proposes a Riemannian gradient descent algorithm with an Armijo line search that, by the paper's claim, reaches a first-order critical point in $O(1/\\epsilon^2)$ iterations and thereby yields a valid lower bound on the classical capacity. The authors argue that the geometric formulation makes the method scale to much larger channels than existing semidefinite-programming approaches, and they support this with numerical comparisons on depolarizing, classical-quantum, entanglement-breaking, Pauli, and qutrit channels. A sympathetic reader would care because exact capacity computation is intractable in general, so a fast and provably converging lower bound is the practical alternative.","feed_headline":"Riemannian descent computes Holevo capacity lower bounds","feed_subtitle":"A gradient method on spheres and a simplex runs in seconds and scales where SDP methods fail.","key_machinery":"The central object is the product manifold $M_{d_A^2+1}$ of one probability simplex (for the ensemble weights) and $d_A^2$ complex unit spheres (for the pure input states), equipped with the Euclidean product metric. The mechanism is the Riemannian gradient formula of Proposition 1, which projects the Euclidean partial derivatives onto the tangent spaces of the simplex and spheres; this gradient is then used with a retraction (the sphere retraction via normalization, and a second-order retraction on the simplex) in a standard Riemannian gradient descent with Armijo backtracking. The convergence guarantee is imported from the general $O(1/\\epsilon^2)$ global rate theorem for nonconvex Riemannian optimization.","core_discovery":"The central discovery is a closed-form Riemannian gradient for the Holevo cost function on the product manifold $M_{d_A^2+1} = \\Delta_{d_A^2-1}^+ \\times (S^{d_A-1})^{\\times d_A^2}$. For a channel $N$ and an ensemble $\\{p_i, |\\psi_i\\rangle\\}$, with $\\sigma_i = N(|\\psi_i\\rangle\\langle\\psi_i|)$ and $\\sigma = \\sum_i p_i \\sigma_i$, the gradient components are $q_j = 1 - D(\\sigma_j\\|\\sigma) + p_j(\\sum_k p_k D(\\sigma_k\\|\\sigma) - 1)$ for the probabilities and $2 p_i [N^\\dagger(\\log \\sigma - \\log \\sigma_i) + D(\\sigma_i\\|\\sigma)] |\\psi_i\\rangle$ for the pure states. Using this gradient, the paper's Riemannian gradient descent (Algorithm 1) is claimed to converge to a first-order critical point, giving a lower bound on the Holevo capacity $\\chi(N)$ for arbitrary quantum channels.","pith_inferences":["A cheap sanity check inside the loop could assert that the probability-gradient components sum to zero; if they do not, the computed step is not a true Riemannian step and the convergence certificate does not apply.","The method could be adapted to estimate the regularized capacity $\\chi(N^{\\otimes n})/n$ directly by running the same algorithm on the tensor-product channel, though the dimension grows exponentially and the product-state gradient formula suggests that non-product random initializations are needed.","Testing on additional analytically solvable channels, such as Werner or Holevo-Werner channels, would show whether the reported accuracy generalizes beyond the paper's examples.","Even if the tangent-space condition for the gradient were violated, the algorithm might still reduce the cost in practice, but its theoretical convergence claim would no longer be supported."],"forward_implications":["If the algorithm converges as claimed, it provides a practical lower bound on the Holevo capacity of any finite-dimensional quantum channel, including channels where regularization via n-fold tensor products is needed.","The method scales to substantially larger input and output dimensions than existing SDP-based estimators, as demonstrated on classical-quantum channels with $|X| = 100$ and output dimension 500.","On channels with known analytic capacity, such as d-dimensional depolarizing channels, the method achieves absolute errors around $10^{-13}$–$10^{-11}$ in seconds, versus roughly $10^4$ seconds for a prior first-order method.","Because the gradient on a tensor product satisfies a factorization identity, product-state starting points are stationary traps, which the authors identify as a caution for superadditivity searches.","Combined with existing semidefinite upper bounds, the lower bounds give bracketing intervals for the classical capacity of general channels, enabling numerical tests of additivity violations."],"supporting_citations":[{"why":"Establishes that the optimal ensemble for the Holevo capacity can be taken as pure states of cardinality at most $d_A^2$, fixing the dimensions of the product manifold.","marker":"[2]"},{"why":"Provides the additivity result for entanglement-breaking channels, used to equate classical capacity with Holevo capacity in the numerical experiments.","marker":"[4]"},{"why":"Supplies the exact Holevo capacity of depolarizing channels, used as the analytical benchmark in Table I.","marker":"[6]"},{"why":"A prior convex-optimization method for approximating channel capacities, providing the baseline for efficiency and accuracy comparisons.","marker":"[22]"},{"why":"An SDP-based estimator of the Holevo capacity of cq channels, the main comparison method for scalability in Tables II and III.","marker":"[27]"},{"why":"Supplies the probability partial derivative for cq channels that the gradient formula generalizes.","marker":"[31]"},{"why":"Gives the global $O(1/\\epsilon^2)$ convergence rate for nonconvex Riemannian optimization that certifies the algorithm's iteration count.","marker":"[32]"}],"fun_headline_variants":["Riemannian gradient descent gives Holevo lower bounds","Closed-form gradient speeds Holevo capacity estimation","Scaling quantum capacity via Riemannian optimization","Riemannian method outperforms SDP for Holevo capacity","Fast Holevo bound via product-manifold gradient"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the probability component of the computed gradient lies in the tangent space of the simplex, meaning its entries sum to zero; the paper's formula does not guarantee this sum vanishes.","fun_headline_variants_meta":{"raw":{"variants":["Riemannian gradient descent gives Holevo lower bounds","Closed-form gradient speeds Holevo capacity estimation","Scaling quantum capacity via Riemannian optimization","Riemannian method outperforms SDP for Holevo capacity","Fast Holevo bound via product-manifold gradient"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1422,"prompt_tokens":857,"completion_tokens":565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":491}},"tokens_in":473,"tokens_out":565,"duration_ms":6887,"temperature":1.0,"reasoning_tokens":491,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:06:06.590416+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $\\sum_j q_j$ from Proposition 1 at a random ensemble for a specific channel, say a 2-qubit depolarizing channel, and check whether the sum vanishes; if it does not, the gradient is not tangent, so the retraction step and the cited $O(1/\\epsilon^2)$ convergence do not apply.","supporting_citations":[{"cited_title":"Sending classical information via noisy quantum channels,","cited_arxiv_id":null,"evidence_quote":"Establishes that the optimal ensemble for the Holevo capacity can be taken as pure states of cardinality at most $d_A^2$, fixing the dimensions of the product manifold."},{"cited_title":"Additivity of the classical capacity of entanglement- breaking quantum channels,","cited_arxiv_id":null,"evidence_quote":"Provides the additivity result for entanglement-breaking channels, used to equate classical capacity with Holevo capacity in the numerical experiments."},{"cited_title":"The capacity of the quantum depolarizing channel,","cited_arxiv_id":null,"evidence_quote":"Supplies the exact Holevo capacity of depolarizing channels, used as the analytical benchmark in Table I."},{"cited_title":"A bregman proximal perspective on classical and quantum Blahut–Arimoto algorithms,","cited_arxiv_id":null,"evidence_quote":"Supplies the probability partial derivative for cq channels that the gradient formula generalizes."}],"review_version":1}