{"id":"2a12b982-b762-4cda-9ee7-32a808ffb7ae","arxiv_id":"2504.13232","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper derives a quaternion-calculus gradient rule for training quantum learning circuits and demonstrates it on a simulated 8-qubit probability-learning task, though the measurement model it relies on is not shown to match standard quantum measurement.","lead":"A quantum machine learning training rule is derived by encoding qubits as quaternions and applying quaternion calculus. The authors report convergent simulations, but the measurement probabilities used in the derivation need to be checked against the Born rule before the method can support real quantum hardware.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (7) fails the Born rule in elementary cases (P=4 for |0> vs |0>, P=0 for |0> vs |+x>), so the cost, gradient, and convergence proof built on it do not describe quantum measurements.","rationale":"The paper's central claim is that an adaptive quaternion-valued W trained via HR-calculus learns mappings between probability distributions encoded in multi-qubit systems, with a guaranteed convergence range. For this to be true, the cost function must be a function of actual quantum measurement probabilities. The only such link is Eq. (7). I verified the two test cases the reader cites. Using the augmented vector definition (3), for q=h=κ (|0>) one has q_a=h_a=[κ,κ,κ,−κ]; quaternion multiplication gives h_a^T q_a = −4, so (7) yields P=4 > 1. For h=ı (|+x>) and q=κ, h_a^T q_a = 0, so P=0 instead of 1/2. These are not edge cases of an approximation; they are elementary measurements that any measurement model must reproduce. Because (7) is unnormalized and has the wrong algebraic structure (transpose without conjugation, and an augmentation sum that overcounts), the objective (8), the gradient expression in Section IV, and the convergence interval imported from quaternion adaptive filtering all inherit the error. The numerical simulation converging is not evidence, because the data are generated by the same quaternion model and no code or data are provided for independent check. The second issue, that W is not constrained to be unitary, is also real: Proposition 1 proves only that actual quantum gates correspond to such linear maps, not the converse, and Remark 4 merely asserts the converse. However, the Eq. (7) failure is the primary, load-bearing defect. I agree with the reader's weakest-assumption assessment. A corrected Born-consistent probability expression would be necessary before the algorithm can be connected to quantum hardware; until that is supplied, the central claim is unsupported, and the appropriate verdict is REJECT rather than UNVERDICTED, because the defect is a concrete inconsistency with quantum-mechanical measurement, not merely missing evidence.","tokens_in":8406,"tokens_out":6526,"duration_ms":60198,"concrete_test":"Analytically verify Eq. (7) against the Born rule for all pairs from {|0>, |1>, |+x>, |−x>, |+y>, |−y>}: for q=h=κ it must give 1, and for q=κ, h=ı it must give 1/2; the current formula gives 4 and 0. If the authors supply a corrected derivation of Eq. (7), or a replacement P(q=h) that passes these cases, and then re-derive the Section IV gradient and the µ-bound under that replacement, the principal defect is settled; otherwise the learning claim cannot connect to measurements.","verdict_should_be":"REJECT","load_bearing_attack":"Equation (7) is the quantitative link between the quaternion model and actual measurement statistics, and Section IV's cost J(W), the gradient ∂d̂_s,n/∂W* ∝ Re{h_a^T W x_a} h_a x_a^H, and the step-size bound µ∈[0,(3⌈log2 z⌉)^−1) all rest on it. For the augmented vectors defined by (3), take q=h=κ (the |0> state): q_a=h_a=[κ,κ,κ,−κ], so h_a^T q_a = −4 and (7) gives P(q=h)=4, whereas the Born rule gives 1. Take h=ı (|+x>) and q=κ: h_a=[ı,−ı,ı,ı], q_a=[κ,κ,κ,−κ], so h_a^T q_a = (−ȷ)+ȷ+(−ȷ)+ȷ = 0 and (7) gives P=0, whereas Born gives 1/2. No normalization or phase convention in Remarks 2–3 repairs these factors, since the quaternions already have unit norm and are purely imaginary. Equation (7) is asserted without derivation and contradicts the quantum measurement theory it cites. Consequently, the training problem (8)–(9), its gradient, and the claimed convergence are disconnected from actual quantum measurement. A second unresolved premise is Remark 4: the learned W is unconstrained, so nothing enforces the unitarity that Proposition 1 requires; the numerical example only feeds quaternion-synthesized data through the same model. These are internal correctness failures, not merely disagreements with standard conventions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a quaternion-valued model for qubits, quantum gates, and measurement, then uses the augmented quaternion representation and HR-calculus to formulate an adaptive learning problem on multi-qubit systems. The central claim is that a quaternion matrix W, trained by gradient descent, can learn a mapping between probability distributions encoded in multi-qubit states, with a guaranteed convergence step-size range. The learning objective is built on a quaternion measurement probability formula in Eq. (7), and the convergence and observability conditions are imported from the authors' prior work on quaternion distributed filtering. A numerical example with an 8-qubit system is presented to demonstrate learning.","tokens_in":8732,"tokens_out":3395,"duration_ms":32624,"significance":"If the central measurement model were correct, the paper would contribute a new algebraic framework for quantum machine learning, connecting quaternion adaptive filtering to quantum systems and potentially enabling new training algorithms. The use of HR-calculus for quaternion gradients is a legitimate and potentially useful technique. However, the core measurement probability formula is not consistent with the Born rule on elementary states, and the learned transformation is not constrained to be unitary. Because the cost function, gradient, and convergence analysis all rest on this formula, the main claims are not supported. The numerical validation is generated within the same quaternion model, so it does not provide independent evidence.","major_comments":[{"comment":"The measurement probability formula P(q=h) = (1/4)||Re(h_a^T q_a)||^2 contradicts the Born rule for simple states. For q = h = κ (the |0> state), the augmented vectors are q_a = h_a = [κ, -κ, -κ, κ]^T, so h_a^T q_a = -4 and Eq. (7) gives P = 4, whereas the Born rule gives P = 1. For h = ı (|+x>) and q = κ, h_a^T q_a is purely imaginary, yielding P = 0, whereas the Born rule gives 1/2. The factor 1/4 and the real-part extraction do not fix the normalization, since the quaternions involved are already unit-norm and purely imaginary. This formula is asserted without derivation, and because the cost function in (8), the gradient, and the convergence analysis in Section IV all use it, the learning algorithm is disconnected from actual quantum measurement statistics.","section":"Section III, Eq. (7)"},{"comment":"Proposition 1 requires multi-qubit computations to be represented by linear unitary maps in the augmented quaternion space, and Remark 4 claims that multiplication by the learned matrix W models unitary rotations and synthesizable quantum operations. However, W is learned as an unconstrained quaternion matrix via gradient descent; no step in the algorithm imposes unitarity or any structural constraint derived from quantum circuits. Consequently, the learned W need not correspond to any valid quantum operation, and the claimed connection to quantum computation is not established. The numerical example does not test whether the learned W is unitary or physically realizable.","section":"Section IV, Remark 4 and Proposition 1"},{"comment":"The observability condition in Theorem 1 and the step-size bound µ ∈ [0, (3⌈log2 z⌉)^(-1)) are imported from the authors' prior work on quaternion distributed filtering ([32], and also [27]), rather than proved within this quantum setting. Since the learning problem is built on the measurement model of Eq. (7), which is not a valid quantum measurement rule, transferring these conditions is not justified. Additionally, the numerical example generates both the data and the target from the same quaternion model, so it does not validate the approach against an independent quantum measurement model or a standard simulation of the Born rule.","section":"Section IV, Theorem 1 and convergence claim"}],"minor_comments":[{"comment":"There are several typographical errors that should be corrected: 'Theses' should be 'These' in the introduction, 'repressuring' should be 'representing' in Remark 3, 'effect' should be 'affect' in Remark 2, and 'Internatational' should be 'International' in reference [6].","section":"Throughout"},{"comment":"The phrase 'An elegant solution for repressuring the state' appears to be a typo for 'representing'. The remark would be clearer if the proposed mapping from α and β to quaternion components were written explicitly with equations.","section":"Section III, Remark 3"},{"comment":"The notation P(q = h) for measurement outcomes is ambiguous, since quantum measurement outcomes are labeled by basis states, not by the full pre-measurement state. The authors should clarify the intended meaning and relate it to standard projection-valued measures.","section":"Section III, Eq. (7) and surrounding text"},{"comment":"The displayed gradient expression contains a fraction: ∂d̂_s,n/∂W* = (1/8) Re{h_a^T W x_a} h_a x_a^H. The factor 1/8 is not derived; it likely depends on the specific measurement formula in (7). Since (7) is in question, the gradient would need to be re-derived if the measurement model is repaired.","section":"Section IV, gradient expression"},{"comment":"The numerical example reports convergence and a likelihood estimate, but it does not state the learning rate µ used, the number of iterations, or the initialization of W. These details are needed to reproduce the experiment. More importantly, the example never checks whether the learned W is unitary or whether the measurement statistics satisfy the Born rule.","section":"Section V, numerical example"}],"recommendation":"reject","confidential_remarks":"The paper's central measurement formula is not merely unconventional; it fails on elementary states, and the learned W is not constrained to be unitary. The theoretical scaffolding relies heavily on the authors' own prior results, and the numerical validation is generated within the same quaternion model. In its current form, the manuscript does not establish a valid connection to quantum learning. I would not encourage resubmission without a complete reworking of the measurement model and an independent validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's central measurement formula is wrong, so the training algorithm does not describe quantum measurements. The idea is worth a second look, but this version should not be published as is.\n\nWhat is new: using quaternions to encode qubits and then borrowing HR-calculus and quaternion adaptive-filtering results to train a quantum learning machine is a legitimate new combination. The exposition is clear, and the authors are honest about the assumptions. The step-size bound from quaternion adaptive filtering is a nice transfer.\n\nWhere it falls apart: Eq. (7) is asserted without derivation and contradicts the Born rule in elementary cases. For q=h=κ (the |0> state), the augmented vectors give h_a^T q_a = -4, so P=4, not 1. For h=ı (|+x>) and q=κ, h_a^T q_a = 0, so P=0, not 1/2. No normalization or phase convention in Remarks 2–3 fixes this. Since the cost in (8) and the gradient in Section IV are built on (7), the learning curve in Fig. 3 is learning the authors' quaternion model, not quantum statistics. The numerical example is self-referential: data are generated by the same model, and the convergence condition is imported from the authors' own [32]. The learned W is not constrained to be unitary, so even within the model it need not correspond to a quantum circuit. These last two are secondary; the measurement formula is the load-bearing flaw.\n\nCredit where due: the quaternion encoding on the Bloch sphere is standard, the HR-calculus derivation of the gradient is clean, and the problem formulation is sensible. None of this saves (7).\n\nWho this is for: a reader interested in quaternion-valued signal processing applied to quantum machine learning might find the construction instructive, but only as a cautionary example of what happens when the measurement model is not checked against the Born rule.\n\nRecommendation: as submitted, reject. If a revised version replaces (7) with a correct quaternion form of the Born rule and enforces unitarity, it would deserve another look. I would send this version to a referee anyway, because the error is precise and the authors might fix it; but the referee report should be blunt about (7).","headline":"The quaternion adaptive-learning setup is a fresh combination, but Eq. (7) gives P=4 for measuring |0> against |0> and P=0 for |+x>, so the learning loop is not connected to quantum measurements as written.","tokens_in":9255,"tokens_out":5676,"would_cite":false,"duration_ms":51820,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","15A33"],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"An adaptive quaternion-valued filter can learn a mapping between data and hypothesis distributions encoded on qubits.","keywords":["quaternion algebra","quantum machine learning","qubit representation","HR-calculus","quantum measurement","adaptive filtering","gradient descent","Bloch sphere"],"falsifier":"Evaluate Eq. (7) for a single qubit: measuring the state $|0\\rangle$ in the $|+x\\rangle$ basis gives probability $1/2$ by quantum mechanics, while Eq. (7) gives $0$; measuring $|0\\rangle$ against itself gives $4$ instead of $1$. If this mismatch persists in the paper's own model, the update rule's probabilities do not describe physical measurements.","tokens_in":8129,"feed_emoji":"⚛️","tokens_out":5688,"duration_ms":56137,"temperature":0.7,"pith_summary":"The paper tries to establish that quaternion algebra is the right language for training quantum learning machines, replacing the complex-valued formulation of quantum computation. It models every single-qubit gate as a quaternion involution and every multi-qubit computation as a linear unitary map in an augmented quaternion space. On that footing it derives an adaptive algorithm, using quaternion gradients, that learns a mapping from a data distribution encoded in multi-qubit states to a hypothesis distribution encoded in measurement outcomes. If the model holds, a quantum circuit can be trained with the same adaptive-filtering toolkit used for classical quaternion-valued signals, complete with convergence and observability conditions.","feed_headline":"Quaternion math can train a quantum learning machine","feed_subtitle":"Treating quantum gates as quaternion rotations makes circuits trainable with classical adaptive-filtering tools.","key_machinery":"The load-bearing object is the augmented quaternion vector $q_a=[q,\\,q^{\\imath},\\,q^{\\jmath},\\,q^{\\kappa}]^{\\mathrm{T}}$ and the identity that rotations of a quaternion become linear maps in this augmented space. Proposition 1 turns quantum gates into quaternion involutions, so multi-qubit computations become linear unitary maps in that space. The measurement model $P(q=h)=\\tfrac{1}{4}\\|\\Re\\{h_a^{\\mathrm{T}}q_a\\}\\|^2$ converts those linear maps into probability distributions, and the HR-calculus gradient provides the update rule. Together these pieces make a quantum circuit look like a quaternion-valued linear system that classical adaptive-filtering theory can train.","core_discovery":"The central claim is that an adaptive quaternion-valued matrix $W$, updated by HR-calculus gradient descent on a measurement-based cost, learns the mapping $f: P(D_n\\mid\\zeta)\\to P(\\zeta\\mid D_n)$ when data and hypothesis distributions are encoded in multi-qubit systems. Proposition 1 asserts that every quantum gate is a quaternion involution and every multi-qubit computation is a linear unitary transformation in augmented quaternion space, so the optimization problem becomes a quaternion-valued linear estimation problem. The paper further claims that the mapping is recoverable when expected measurement outcomes are available together with $3\\lceil\\log_2(z)\\rceil$ linearly independent measurement vectors, and that, with an encoding satisfying Assumption 2, step sizes $\\mu\\in[0,(3\\lceil\\log_2(z)\\rceil)^{-1})$ give convergent training. The numerical example on an 8-qubit system shows the cost decreasing and the estimated circuit reproducing the likelihood of held-out hypotheses.","pith_inferences":["If the measurement formula in Eq. (7) were replaced by a map that matches the Born rule exactly, the same augmented-quaternion machinery would connect the algorithm to measurements on real quantum hardware; the paper's own simulation uses Eq. (7) rather than the Born rule.","Requiring the learned matrix $W$ to be unitary on the augmented space would turn the proposed algorithm into a training rule for parameterized quantum circuits, a constraint the paper asserts but never enforces.","A natural testable extension is to rerun the same update rule with measurement probabilities computed by the standard Born rule and compare the convergence behavior, since the cost function would then reflect physical measurements rather than Eq. (7)."],"forward_implications":["A quantum learning machine can be trained by quaternion adaptive filtering, inheriting convergence and observability criteria from that theory.","The mapping between data and hypothesis distributions is learnable from expected measurement outcomes plus $3\\lceil\\log_2(z)\\rceil$ linearly independent measurement vectors.","Training converges for step sizes $\\mu\\in[0,(3\\lceil\\log_2(z)\\rceil)^{-1})$ when the encoding keeps $\\Re\\{h_a^{\\mathrm{T}*} y_a^n\\}>0$ for all measurement vectors and samples.","The trained circuit estimates $P(\\zeta'\\mid D_n)$ for held-out hypotheses, as demonstrated on an 8-qubit numerical example."],"supporting_citations":[{"why":"Supplies the HR-calculus and the augmented-quaternion linearization that the gradient update and Proposition 1 rely on.","marker":"[21]"},{"why":"The textbook reference for quantum measurement theory that Eq. (7) is claimed to follow from.","marker":"[16]"},{"why":"Establishes observability conditions for distributed quaternion state estimation, which Theorem 1 invokes for recoverability of the mapping.","marker":"[32]"},{"why":"Defines the quaternion gradient operator used to derive the training update.","marker":"[25]"},{"why":"Provides augmented statistics of quaternion random variables supporting the augmented-space formulation.","marker":"[26]"},{"why":"Justifies that the permitted operations (rotations plus addition with re-normalisation) are implementable by known quantum circuitry.","marker":"[11]"},{"why":"Lecture notes used for the claim that multi-qubit operations are rotations and re-normalisations, and for circuit synthesis.","marker":"[17]"}],"fun_headline_variants":["Quantum of Learning: Quaternion Training for Qubits","Quaternion algebra trains quantum learning machines","HR-calculus adapts quantum circuits via quaternions","Learning quantum mappings with quaternion gradient descent","Adaptive quantum circuits from quaternion involution gates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training procedure stands on the measurement probability formula in Eq. (7), which is stated without derivation and whose probabilities must match quantum measurement outcomes for the learned circuit to mean anything on real hardware.","fun_headline_variants_meta":{"raw":{"variants":["Quantum of Learning: Quaternion Training for Qubits","Quaternion algebra trains quantum learning machines","HR-calculus adapts quantum circuits via quaternions","Learning quantum mappings with quaternion gradient descent","Adaptive quantum circuits from quaternion involution gates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1387,"prompt_tokens":838,"completion_tokens":549,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":475}},"tokens_in":454,"tokens_out":549,"duration_ms":6695,"temperature":1.0,"reasoning_tokens":475,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:18:31.572922+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate Eq. (7) for a single qubit: measuring the state $|0\\rangle$ in the $|+x\\rangle$ basis gives probability $1/2$ by quantum mechanics, while Eq. (7) gives $0$; measuring $|0\\rangle$ against itself gives $4$ instead of $1$. If this mismatch persists in the paper's own model, the update rule's probabilities do not describe physical measurements.","supporting_citations":[{"cited_title":"The HR-Calculus: Enabling Information Processing with Quaternion Algebra","cited_arxiv_id":"2311.16771","evidence_quote":"Supplies the HR-calculus and the augmented-quaternion linearization that the gradient update and Proposition 1 rely on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The textbook reference for quantum measurement theory that Eq. (7) is claimed to follow from."},{"cited_title":"Quaternion-valued dis- tributed filtering and control,","cited_arxiv_id":null,"evidence_quote":"Establishes observability conditions for distributed quaternion state estimation, which Theorem 1 invokes for recoverability of the mapping."},{"cited_title":"A quaternion gradient operator and its applications,","cited_arxiv_id":null,"evidence_quote":"Defines the quaternion gradient operator used to derive the training update."},{"cited_title":"Augmented statistics of quaternion random variables: A lynchpin of quaternion learning machines,","cited_arxiv_id":null,"evidence_quote":"Provides augmented statistics of quaternion random variables supporting the augmented-space formulation."},{"cited_title":"Quantum computing through the lens of control: A tutorial introduction,","cited_arxiv_id":null,"evidence_quote":"Justifies that the permitted operations (rotations plus addition with re-normalisation) are implementable by known quantum circuitry."}],"review_version":1}