{"id":"e85a41ea-1e32-4205-8c23-62c33afb15bf","arxiv_id":"2412.11889","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A parameterized quantum embedding can be trained to make a circuit output invariant under a finite group, which the authors interpret as learning an equivariant map between representations.","lead":"This paper proposes using variational quantum circuits to learn symmetric data embeddings that map between two given representations of a finite group. The method is intended to let quantum machine learning models automatically respect symmetries rather than require a hand-chosen equivariant embedding.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Output-invariance loss (Eq. 28) does not imply equivariance of E_theta; the paper's own C2xC2 example trains non-equivariant parameters while pushing the loss low, so the central claim is unsupported.","rationale":"The reader's weakest_assumption matches the concern I find most load-bearing: Eq. (28) is a scalar-output loss, and the paper never proves that its minimizers are equivariant embeddings. The C2xC2 simulation is not merely a hypothetical failure: the reported converged parameters contradict the paper's own equivariance condition, yet the authors present the run as a success. This is an internal inconsistency, not a disagreement with an external consensus. Because the claimed reusable embedding and the covariant-channel application both depend on E_theta (or U_theta in Section IV) being genuinely equivariant, the central claim is unsupported. I agree with the reader's REJECT and recommend no verdict change. The direct equivariance-error check above would settle the issue cleanly.","tokens_in":15460,"tokens_out":14247,"duration_ms":125777,"concrete_test":"Reproduce the C2xC2 experiment with the reported trained parameters theta=(3.24,6.22,4.89,4.97). For both generators g=F_v,F_h and a grid of points x from the training distribution, directly compute the state-space equivariance error epsilon_g(x)=||E_theta(V(g)x)|0> - W(g)E_theta(x)|0>||_2, and compare its mean/max with the reported L_g. If epsilon is large while L_g is small, the loss does not certify equivariance. Then retrain while minimizing epsilon^2 (or adding it as a penalty) and check whether the recovered parameters match those reported; if they do not, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III's central step is the assertion that minimizing L_g(x)=(h_theta(V(g)x)-h_theta(x))^2 in Eq. (28) learns an equivariant embedding E_theta. This implication is invalid. h_theta is the expectation of one G-invariant observable after a G-invariant circuit, so the zero set of L_g contains embeddings for which E_theta(V(g)x)|0> != W(g)E_theta(x)|0>. The C2xC2 example makes the failure concrete. For E_theta=⊗_i Ry(theta_i+x_i), W(F_v)=SWAP01SWAP23 and W(F_h)=SWAP02SWAP13, exact state equivariance for both generators requires theta_0≡theta_1≡theta_2≡theta_3 (mod 2π). The paper reports successful training to theta≈(3.24,6.22,4.89,4.97)≈(π,2π,3π/2,3π/2), which does not satisfy this condition, while the validation loss is low. Thus the experiment demonstrably optimizes output invariance without producing an equivariant embedding. The D4 discussion likewise equates 'nearly equivariant' with near-identical expectation values, and Section IV's Proposition 2 inherits the same gap because Eq. (47) again constrains only expectation values. No identifiability argument or direct equivariance metric is supplied, so the paper's examples do not support the abstract claim of learning equivariant maps or covariant channels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a variational quantum circuit (VQC) method for learning an equivariant map between two given unitary representations of a finite group. The idea is to parameterize the data embedding E_theta, append a G-invariant circuit and observable, and train by minimizing the squared difference of expectation values L_g(x) = (h_theta(V(g)x) - h_theta(x))^2 over group generators and data. The authors claim that this procedure learns an equivariant embedding that can be reused in symmetric classifiers and that it extends to learning intertwining maps and covariant channels. The paper includes a review of symmetry-preserving VQCs and presents four examples: C2, C2 x C2, D4, and S6, with numerical simulations in PennyLane.","tokens_in":15811,"tokens_out":8366,"duration_ms":85857,"significance":"If the central claim were correct, the paper would contribute a practical way to learn symmetric embeddings rather than imposing them by construction, which is relevant for geometric quantum machine learning and for tasks involving covariant quantum channels. The review material in Section II (Proposition 1, Lemmas 1 and 2) is correct and clearly presented, and the examples are simple enough to be reproducible. However, the central methodological claim is not established: the loss function constrains only a single scalar expectation value, and the paper's own C2 x C2 experiment shows low loss with parameters that do not satisfy the equivariance conditions. The significance of the paper therefore depends on an inference that is invalid as written.","major_comments":[{"comment":"The central step is invalid: minimizing L_g(x) = (h_theta(V(g)x) - h_theta(x))^2 with h_theta the expectation value of a single observable does not imply that the embedding E_theta is equivariant. The zero set of L_g contains embeddings for which E_theta(V(g)x)|0> is not equal to W(g)E_theta(x)|0>, because the observable projects out almost all information about the state. No identifiability condition, completeness of the observable set, or direct equivariance error is supplied. Without such an argument, the paper does not establish that the procedure learns equivariant embeddings, despite the claim in the text that minimizing L_g learns the desired embedding.","section":"Section III, Eq. (28)"},{"comment":"The experimental results actually demonstrate the gap between L_g and equivariance. For the tensor-product ansatz E_theta(x) = ⊗_i Ry(theta_i + x_i) with W(F_v) = SWAP01 SWAP23 and W(F_h) = SWAP02 SWAP13, exact equivariance of the state T(x) = E_theta(x)|0> requires the parameters to be equal (up to the period of the Ry state and possible cancellations of global phases). The reported C2 x C2 parameters theta ≈ (3.24, 6.22, 4.89, 4.97) do not satisfy this condition (e.g., theta_0 - theta_2 ≈ -1.65, not a multiple of 2π), yet the validation loss is small. The paper's claim that these parameters correspond to an equivariant map is therefore contradicted by its own numbers.","section":"Section III, C2 and C2 x C2 examples"},{"comment":"The statement that equivariance is achieved when each pair (theta_0, theta_1) and (theta_2, theta_3) differs by an odd multiple of π is incorrect for state or operator equivariance. For example, Ry(theta + π)|0> is not proportional to Ry(theta)|0> as a vector-valued function of theta; the two are orthogonal for theta = 0. The condition only holds for products of Z-expectation values, which is precisely the loophole that allows the scalar loss to be small without the embedding being equivariant. This strengthens the concern that the method, as presented, optimizes output invariance rather than embedding equivariance.","section":"Section III, C2 example"},{"comment":"The same gap appears in the intertwining-map setting. Proposition 2 proves that h_theta is G-invariant under the hypothesis that U_theta is an intertwining map, but the training procedure minimizes L_g over expectation values only. The proof does not supply any mechanism by which minimizing L_g drives U_theta toward the hypothesis; Eq. (47) is identical in form to Eq. (28) and suffers from the same insufficiency. The discussion of learning intertwining maps therefore inherits the central flaw.","section":"Section IV, Proposition 2 and Eq. (47)"},{"comment":"The D4 example reports nearly identical expectation values for three transformed data points and concludes that the embedding is 'nearly equivariant,' while the S6 example reports parameter convergence. Neither example provides a direct measure of equivariance of the embedding, such as the state fidelity between E_theta(V(g)x)|0> and W(g)E_theta(x)|0>, or a comparison across a tomographically complete set of observables. Given the demonstrated failure in the C2 x C2 example, the absence of such a metric means the D4 and S6 results do not support the paper's central claim.","section":"Section III, D4 and S6 examples"}],"minor_comments":[{"comment":"The sentence 'the value of L_g(x) can be interpreted as a measure of how equivariant the embedding E_theta is' is misleading. L_g measures only the invariance of the chosen expectation value under the group action, not the equivariance of the embedding; this should be stated explicitly.","section":"Section III, text after Eq. (28)"},{"comment":"The symbol T(x) is used for the normalized amplitude-embedding state in Eq. (36), while earlier in the paper T denotes the equivariant map from the data space to the Hilbert space. This overloading of notation should be resolved for clarity.","section":"Section IV, Eq. (36)"},{"comment":"The horizontal axis is labeled 'Batch Number' and the caption says the curves show parameter values during training, but it is unclear whether the validation loss in Figures 5 and 8 is evaluated per batch, per epoch, or on a separate validation set. A clearer description of the training and evaluation protocol would improve reproducibility.","section":"Figure 4 and Figure 7"},{"comment":"The conclusion states that the work 'provides a method for measuring the similarity between two representations.' The loss L_g depends on the chosen observable and the data distribution and is not a well-defined metric on representations; this claim is stronger than what is demonstrated.","section":"Section V, Conclusion"}],"recommendation":"reject","confidential_remarks":"The paper contains a clear internal counterexample: the trained parameters in the C2 x C2 example violate the derived equivariance conditions while achieving low loss. This is not a missing proof but a falsification of the proposed method as stated. A salvageable version would need to replace the scalar loss with a direct equivariance error (e.g., a SWAP test or a sum over a tomographically complete set of observables) and then re-run all experiments with that new loss. That would be a substantially different paper. I recommend rejection in the current form, though the underlying idea—learning embeddings by symmetry constraints—may be worth pursuing with a correctly designed objective."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper reads well and the representation theory is handled cleanly. The idea of learning an equivariant embedding rather than hand-picking one is a natural and reasonable goal in geometric QML, and the authors cite the relevant classical work on learning invariances. The review of symmetric VQC construction in Section II is solid, and the examples are clearly described. That is the good part.\n\nThe problem is the central claim, and it is not a minor issue. Equation (28) minimizes (h_theta(V(g)x) - h_theta(x))^2, which is a difference of expectation values after a G-invariant circuit and a G-invariant observable. The zero set of that loss contains many embeddings that are not equivariant; it merely makes the scalar output invariant. The paper explicitly asserts that this loss is 'a measure of how equivariant the embedding E_theta is,' but that implication does not hold. The C2xC2 example makes the failure concrete: the reported trained parameters roughly (pi, 2pi, 3pi/2, 3pi/2) do not satisfy the derived exact equivariance condition (all four equal mod 2pi), yet the validation loss is low. So the experiment demonstrably optimizes output invariance without producing an equivariant embedding. The D4 discussion makes the same conflation, calling near-identical expectation values 'nearly equivariant,' and Proposition 2 inherits the same gap because Eq. (47) again constrains only expectation values.\n\nWhat would fix this? Train on a direct equivariance loss, e.g., comparing E_theta(V(g)x)|0> with W(g)E_theta(x)|0>, or at least verify the embedding equivariance on held-out points and report the state-space error. Without that, the paper's main claim—that you can learn an equivariant map—is unsupported. Also note there is no downstream classification or covariant channel task, so the practical utility is not demonstrated even in a toy setting.\n\nIs it worth serious refereeing? I think yes. The flaw is instructive, the topic is of moderate interest, and the fix is reasonably clear. A good referee could guide the authors to either strengthen the objective or reframe the claims to be about learning invariant outputs. As it stands, I would not accept it, but I would not desk-reject it either; it deserves a careful review that can identify this gap and demand the missing verification.","headline":"The paper is clear and the setup is sensible, but the loss it minimizes does not learn equivariant embeddings, and the paper's own C2xC2 example shows the gap.","tokens_in":16256,"tokens_out":2029,"would_cite":false,"duration_ms":21331,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","81R05","81P45"],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"A variational quantum circuit can learn the equivariant map between two group representations.","keywords":["variational quantum circuits","equivariant maps","geometric quantum machine learning","representation theory","covariant quantum channels","intertwining maps","quantum embeddings","finite group symmetries"],"falsifier":"After training on a fixed pair of representations, choose several points $x$ and group elements $g$, prepare the states $E_\\theta(V(g)x)|0\\rangle$ and $W(g)E_\\theta(x)|0\\rangle$, and compare them with full state tomography or a SWAP test; if the states differ while $L_g$ is near zero, the loss is not a faithful proxy for equivariance and the paper's central implication fails.","tokens_in":15293,"feed_emoji":"⚛️","tokens_out":6822,"duration_ms":57349,"temperature":0.7,"pith_summary":"This paper proposes that a variational quantum circuit can learn an equivariant map between two unitary representations of a finite group, instead of requiring the symmetric embedding to be designed in advance. The method parameterizes the data embedding and trains it so that the circuit output does not change when the input is transformed by the group, using the loss $L_g(x)=(h_\\theta(V(g)x)-h_\\theta(x))^2$ over sampled points and group generators. The authors present examples with $C_2$, $C_2\\times C_2$, $D_4$, and $S_6$ symmetries, and extend the procedure to linear intertwining maps by exploiting the equivalence between normalized and unnormalized amplitude embeddings. If the proposal holds, geometric quantum machine learning can learn its symmetry-respecting embeddings from data, and the learned maps double as covariant quantum channels.","feed_headline":"Quantum circuits learn symmetric maps by training the embedding","feed_subtitle":"A variational circuit learns the symmetric embedding from data, then reuses it for classification and covariant channels.","key_machinery":"The central mechanism is the paired loop between a trainable embedding and a fixed $G$-invariant measurement: because the circuit's expectation value is invariant exactly when the embedding respects the group action, the scalar loss $L_g$ acts as a differentiable measure of equivariance. The intertwining-map extension rests on Lemma 3, which identifies equivariance of the normalized amplitude embedding with equivariance of its unnormalized linear counterpart, allowing a linear map to be trained by the same circuit-based loss.","core_discovery":"The central claim is that the construction of symmetry-respecting variational circuits can be inverted: given two unitary representations $V$ and $W$ of a finite group $G$, a parameterized embedding $E_\\theta$ can be trained to be equivariant between them by minimizing $L_g(x)=(h_\\theta(V(g)x)-h_\\theta(x))^2$, where $h_\\theta$ is the expectation value of a circuit whose unitary and observable are $G$-invariant with respect to $W$. The paper argues that an equivariant $E_\\theta$ makes $h_\\theta$ $G$-invariant, and that minimizing this loss over data points and generators produces an almost-equivariant embedding usable for classification. It then shows that with amplitude embeddings, equivariance of the normalized embedding is equivalent to the unnormalized linear map being an intertwining map, so the same loss can train a true linear equivariant map.","pith_inferences":["The loss $L_g$ compares only expectation values under a single observable, so a near-zero loss does not certify that the embedding operator itself is equivariant; a direct check with a tomographically complete set of observables would test that stronger claim, which the paper does not perform.","The same training objective could be relaxed to learn approximately equivariant maps for data with approximate or partial symmetries, a direction the paper leaves implicit.","When several parameter settings satisfy equivariance (as in the $C_2$ example), the optimizer's final choice may be influenced by initialization and architecture; an added classification objective could select among them."],"forward_implications":["A learned equivariant embedding can be substituted for the hand-designed symmetric embeddings used in geometric quantum machine learning, without changing the downstream classification circuit.","Since a learned linear equivariant map is an intertwining map, the procedure trains covariant quantum channels directly from data.","Sampling the loss over a generating set of $G$ suffices, so training cost is governed by the size of the generating set rather than the full group.","The loss value itself provides a quantitative score of how equivariant a given embedding is, enabling direct comparison of architectures on the same symmetry task."],"supporting_citations":[{"why":"It supplies the representation-theoretic framework for geometric quantum machine learning that this work inverts.","marker":"[15]"},{"why":"It provides the theory of equivariant quantum neural networks whose symmetry-preserving circuits are reused here.","marker":"[14]"},{"why":"It gives the alternative penalty-based way to enforce symmetry and motivates testing G-invariance efficiently.","marker":"[38]"},{"why":"It supports the claim that the data embedding is a nontrivial choice with consequences for the model.","marker":"[36]"},{"why":"It backs the assertion that embedding choice can strongly affect trainability.","marker":"[37]"},{"why":"It is the simulation backend used to implement and train the example circuits.","marker":"[40]"},{"why":"It supplies the simultaneous-perturbation gradient approximation used in optimization.","marker":"[41]"},{"why":"It connects equivariant maps to covariant quantum channels, the main information-theoretic application.","marker":"[20]"},{"why":"It defines intertwining maps and representation equivalence used in Section IV.","marker":"[28]"}],"fun_headline_variants":["Quantum circuits learn equivariant embeddings from data","Train symmetric embeddings with variational quantum circuits","Variational quantum circuits can learn equivariant maps","Learn the symmetric embedding instead of assuming it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that minimizing the scalar loss $L_g(x)=(h_\\theta(V(g)x)-h_\\theta(x))^2$ drives the embedding $E_\\theta$ to be equivariant, even though the loss sees only the expectation value of one observable.","fun_headline_variants_meta":{"raw":{"variants":["Quantum circuits learn equivariant embeddings from data","Train symmetric embeddings with variational quantum circuits","Variational quantum circuits can learn equivariant maps","Learn the symmetric embedding instead of assuming it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000329,"raw_usage":{"total_tokens":1775,"prompt_tokens":821,"completion_tokens":954,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":911}},"tokens_in":437,"tokens_out":954,"duration_ms":8946,"temperature":1.0,"reasoning_tokens":911,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:29:36.606297+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"After training on a fixed pair of representations, choose several points $x$ and group elements $g$, prepare the states $E_\\theta(V(g)x)|0\\rangle$ and $W(g)E_\\theta(x)|0\\rangle$, and compare them with full state tomography or a SWAP test; if the states differ while $L_g$ is near zero, the loss is not a faithful proxy for equivariance and the paper's central implication fails.","supporting_citations":[{"cited_title":"Nguyen, Louis Schatzki, Paolo Braccia, Michael Ragone, Patrick J","cited_arxiv_id":null,"evidence_quote":"It provides the theory of equivariant quantum neural networks whose symmetry-preserving circuits are reused here."},{"cited_title":"LaBorde, Soorya Rethinasamy, and Mark M","cited_arxiv_id":null,"evidence_quote":"It gives the alternative penalty-based way to enforce symmetry and motivates testing G-invariance efficiently."},{"cited_title":"Robust data encodings for quantum classifiers","cited_arxiv_id":null,"evidence_quote":"It backs the assertion that embedding choice can strongly affect trainability."},{"cited_title":"The data, which is shuffled at every epoch, is generated in Python","cited_arxiv_id":null,"evidence_quote":"It supplies the simultaneous-perturbation gradient approximation used in optimization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It connects equivariant maps to covariant quantum channels, the main information-theoretic application."},{"cited_title":"Steinberg","cited_arxiv_id":null,"evidence_quote":"It defines intertwining maps and representation equivalence used in Section IV."}],"review_version":1}