REVIEW 3 major objections 5 minor 1 cited by
Measurement-based quantum convolutional neural network for deep learning
T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash
Pith's one-line read This paper claims that a general QCNN can be implemented exactly by preparing a cluster state and training only local measurement bases, replacing deep coherent gate sequences.
desk verdict The exact cluster-state construction for QCNNs is invalid as written because Eq. (2) is angle-independent, but the variational square-lattice numerics are a salvageable idea. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the cluster-state gadget pair: the line cluster $|L_4\rangle$ (four qubits chained, with the first as input and the fourth as output) for single-qubit rotations, and the T-shaped cluster $|E_8\rangle$ (eight qubits, two inputs and two outputs) for a CNOT. These are the elementary logic blocks of the construction: composing them according to known two-qubit unitary decompositions yields the $U_{ij}$ and $V_{ij}$ clusters, and wiring those in the QCNN pattern produces the full resource state. Training is carried by the local projectors $\langle 0|R_z(\theta)\rangle$ (and, in the square-lattice examples, $\langle 0|R_y(\alpha)R_z(\beta)\rangle$), whose angles are updated by gradient descent on the mean-square loss defined in Eq. (5).
What would settle it
Compute the left-hand side of Eq. (2) for a general input state $|\psi\rangle$: because $R_z(\theta)|0\rangle$ is proportional to $|0\rangle$ for every $\theta$, each projector $\langle 0|R_z(\theta)\rangle$ is, up to an overall phase, just $\langle 0|$, so the post-measurement state cannot depend on $\theta$, $\zeta$, or $\xi$. Checking the output state numerically for two different angle triples would settle whether the claimed rotation identity holds; the calculation predicts no angle dependence, contradicting Eq. (2).
Extended reading notes
Core claim
On the paper's own terms, the central discovery is that the convolutional and pooling layers of a QCNN can be rebuilt as fixed entangled sub-clusters of a larger cluster state. The elementary blocks are the four-qubit line cluster $|L_4\rangle$, which is claimed to perform a single-qubit rotation when its first three qubits are measured in bases $\langle 0|R_z(\theta)\rangle$, and the eight-qubit T-shaped cluster $|E_8\rangle$, which is claimed to perform a CNOT when most of its qubits are measured in $\langle 0|$. Standard decompositions of two-qubit unitaries into single-qubit rotations and CNOTs let these blocks be wired into a $U_{ij}$-cluster for each convolutional gate and a $V_{ij}$-cluster for each pooling control-rotation; connecting those clusters in the same zig-zag pattern as the QCNN circuit gives a total state whose measurement angles are the trainable parameters. The numerical demonstrations then replace the exact wiring with experimentally convenient square-lattice cluster states, which are not circuit-equivalent but still reproduce the learning behaviour.
Load-bearing premise
The entire construction depends on one identity: measuring three qubits of a four-qubit chain in the stated rotated bases is supposed to produce a controllable rotation on the fourth qubit. If that identity does not hold, the two-qubit gate gadgets built from it do not perform their assigned gates.
Editorial extensions
If this is right
- A QCNN of arbitrary depth can in principle be executed by preparing one fixed entangled resource state and measuring it, so the difficulty moves from coherent multi-gate control to cluster-state preparation and local measurement.
- Because the same physical cluster state is reprogrammed by changing measurement angles, one hardware resource could be reused for many learning tasks without changing the entanglement structure.
- Square-lattice cluster states, which have been generated at large scale, appear sufficient for the demonstrated learning tasks, giving a concrete path from current experiments to quantum deep learning.
- In the reported benchmarks the measurement-based model converges at least as fast as a circuit QCNN and faster than a classical CNN of the same parameter count, and it reaches higher iris-classification accuracy than the circuit QCNN.
Reading between the lines
- A natural extension the paper leaves implicit is that the same mapping could translate other layered quantum circuits into cluster-state measurements, trading circuit depth for cluster-state width and pre-computable entanglement.
- The local, angle-only training landscape suggests that gradient estimates could be obtained from measurement statistics alone, for example by finite differences across repeated preparations, which is how the paper itself computes gradients numerically.
- The square-lattice results raise the possibility that cluster-state geometry itself is a trainable inductive bias; testing the scheme on larger classical datasets would show whether the accuracy gain over circuit QCNNs persists outside the iris benchmark.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a measurement-based quantum convolutional neural network (MBQCNN) in which a cluster state is prepared and trained by tuning local measurement bases rather than by applying deep gate sequences. The central theoretical claim is an "exact cluster state solution to general QCNNs," built from |L4> line-cluster gadgets for single-qubit rotations and |E8> cluster gadgets for CNOT operations, assembled into U_ij and V_ij clusters. Numerical experiments are then reported for two tasks: identifying Haldane ground-state phases using a 2-by-5 square-lattice cluster state, and classifying the iris dataset using a 2-by-4 square-lattice cluster state. The numerical model uses parameterized projectors <0|R_y(alpha)R_z(beta), not the <0|R_z(theta) projectors of the exact construction.
Significance. If the exact mapping were correct, the paper would offer a practical route to executing deep QCNNs as static cluster-state preparations followed by local measurements, avoiding deep coherent control. The numerical results, with disjoint training and testing sets and parameter-matched comparison models, are a useful heuristic demonstration of a square-lattice variational model. However, the central identity in Eq. (2) is algebraically false, and the numerical sections do not implement the claimed exact construction. The paper therefore does not establish its advertised theoretical contribution; the numerical evidence is for a different model.
major comments (3)
- [Section II, Eq. (2)] The identity in Eq. (2) is algebraically incorrect: since R_z(theta)|0> = e^{-i theta/2}|0>, the projector <0|R_z(theta) is proportional to <0| and carries no theta dependence. Measuring the first three qubits of |L4> in the stated bases leaves the fourth qubit in |psi> up to a global phase, not in U_H R_z(xi) R_x(zeta) R_z(theta)|psi>. Because the U_ij and V_ij clusters in Fig. 1(b)-(c) and Supplementary S1-S2 are composed of such |L4> gadgets, the claimed exact cluster-state implementation of a general QCNN collapses.
- [Section III, Eq. (7) and Figs. 2-4] The numerical demonstrations replace the Section II construction with square-lattice cluster states measured in parameter-dependent bases <0|R_y(alpha_i)R_z(beta_i). These simulations therefore do not validate the exact mapping advertised in the Abstract and Section II; they validate a different variational model. The Introduction explicitly notes that the square-lattice states do not have a one-to-one correspondence with the QCNN circuit, but the paper still claims the numerical results support the exact scheme, which they do not.
- [Discussion and Conclusion] The Discussion states that the paper provides "a strict cluster state that can realize the function of every detail of the circuit." This conclusion depends on the invalid Eq. (2) and is therefore unsupported. The paper's own caveat in Section II that the V_ij-cluster constraints are unnecessary in practice further distances the numerical results from the strict construction, but the decisive issue remains the invalid rotation gadget in Eq. (2).
minor comments (5)
- [Section III] The text refers to "the ground states of Eq. (7)" where the Hamiltonian is defined in Eq. (6); this appears to be a typo.
- [Section III, Eq. (7)] The state in Eq. (7) is written as |S13> although it was defined as |S13^(i)> earlier in the same section; the notation should be made consistent.
- [Section I] The abbreviation "MBQCC" appears once in the Introduction; it should be "MBQCNN."
- [Section III] Ref. [18] is cited for the QCNN treatment of Haldane ground states, but the relevant prior work appears to be Cong et al., Ref. [27]; the citation should be checked and corrected.
- [Appendix C] In the definition of the CV CZ operator, the text reads "q_i and q_i"; the second index should be j.
Circularity Check
No significant circularity: the exact mapping is asserted via standard MBQC identities, and numerical claims are tested on disjoint data; the apparent flaw in Eq. (2) is a correctness issue, not circularity.
full rationale
I examined the derivation chain for circular reductions. The only self-citation is Ref. [25] (same authors), and it appears in the introduction merely as an example of quantum deep transfer learning, not as load-bearing evidence for any gadget, uniqueness claim, or numerical result. The central exact mapping from QCNN circuits to cluster states is built on standard measurement-based quantum computing constructions (Refs. [34-46,50]) and on the algebraic identity in Eq. (2). Even if Eq. (2) is mathematically incorrect, because Rz(theta)|0> equals |0> up to a global phase so the projectors carry no angle information, that is an internal algebraic error rather than a circularity: the claimed output is not assumed in the input, and no fitted parameter is relabeled as a prediction. The numerical examples use parameterized projectors of the form <0|Ry(alpha)Rz(beta) on square-lattice cluster states, train on one dataset, and test on completely different samples, so the reported accuracies are not forced by construction. The paper also explicitly states that the square-lattice states used numerically do not have a one-to-one correspondence with the QCNN circuit, so the numerical success is not merely a restatement of the exact-mapping claim. No instance of the seven circularity patterns was found, and the appropriate finding is therefore no significant circularity.
Assumptions & free parameters
free parameters (2)
- trainable measurement angles (alpha_i, beta_i) of each qubit in the numerical cluster states =
26 angles for the Haldane 13-qubit example; 28 angles for the iris 14-qubit example
- cluster graph topology per task =
2-by-5 square lattice for Haldane; 2-by-4 square lattice with two extra qubits and one diagonal for iris
assumptions (5)
- ad hoc to paper Measuring the first three qubits of the |L4> line cluster in bases <0|R_z(theta) produces a single-qubit rotation with angle theta (Eq. (2)).
- domain assumption The |E8> T-shaped cluster, with all qubits except 5 and 8 measured in <0|, implements a CNOT gate (Eq. (4)).
- standard math Any general QCNN circuit can be compiled exactly into a cluster state by concatenating the U_ij and V_ij gadgets.
- ad hoc to paper The V_ij-cluster measurement constraints can be relaxed without harming learning performance.
- domain assumption Iris data can be amplitude-encoded into a 4-qubit input by padding each 4-vector to dimension 16.
Cite this review
Pith. "Pith review of Measurement-based quantum convolutional neural network for deep learning." pith.science (2026). https://pith.science/paper/HMNXB7XM
@misc{pith2026241208207,
author = {Pith},
title = {Pith review of: Measurement-based quantum convolutional neural network for deep learning},
year = {2026},
howpublished = {\url{https://pith.science/paper/HMNXB7XM}},
note = {Machine review of arXiv:2412.08207}
}
read the original abstract
Recently, quantum convolutional neural networks (QCNNs) are proposed, harnessing the power of quantum computing for faster training compared to the classical counterparts. However, this framework for deep learning also relies on multiple processing layers to capture the representation of data, which necessitates precise dynamical control. Given the current stage of quantum computing, achieving this level of control at a large scale remains challenging. Here, we propose an alternate approach to implementing QCNNs by utilizing cluster states. The training process of the method involves tuning the projection basis of each qubit in cluster states, rather than adjusting the parameters of layers of operators in deep quantum circuits. Hence, the whole system is easier to stabilize by avoiding the complex controls. Leveraging techniques in measurement-based quantum computing, we present an exact cluster state solution to general QCNNs. Followingly, we provide numerical evidence that both quantum and classical data can be learned by measuring cluster states, and a faster convergence of the method is observed. The cluster states we consider in our learning examples are merely square-lattice cluster states, whose implementation at large scale have been reported recently. It indicates that our method has the potential for realizing the advance of quantum deep learning for practical uses.
Figures
Forward citations
Cited by 1 Pith paper
-
One Polynomial Strategy for Computing Local Projections on Square-Lattice Cluster States
The note conjectures a polynomial-time recursive method for computing arbitrary local projections on 2D square-lattice cluster states, but the core 2D recursion is not proved and the numerical evidence is too small to...
Reference graph
Works this paper leans on
-
[1]
M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, 10th anniversary ed (Cambridge University Press, Cambridge; New York, 2010)
work page 2010
-
[2]
G. Vidal and C. M. Dawson, Universal Quantum Circuit for Two- Qubit Transformations with Three Controlled-NOT Gates, Phys. Rev. A 69, 010301 (2004)
work page 2004
-
[3]
I. Cong, S. Choi, and M. D. Lukin, Quantum Convolutional Neural Networks, Nat. Phys. 15, 1273 (2019)
work page 2019
Reviewed August 11, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.