arxiv: 2604.06094 · v2 · submitted 2026-04-07 · 🪐 quant-ph · cs.LG

Recognition: 2 theorem links

· Lean Theorem

Pixel-Translation-Equivariant Quantum Convolutional Neural Networks via Fourier Multiplexers

Dmitry Chirkov , Igor Lobanov

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:16 UTC · model grok-4.3

classification 🪐 quant-ph cs.LG

keywords quantum convolutional neural networkstranslation equivariancequantum Fourier transformpixel cyclic shiftFourier multiplexerbarren plateausmeasurement-induced pooling

0 comments

The pith

Conjugation by the quantum Fourier transform diagonalizes pixel shifts, so every translation-equivariant quantum CNN layer reduces to a Fourier-mode multiplexer followed by its inverse.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that quantum convolutional layers can be forced to commute exactly with the cyclic shift on pixel indices that appears when images are loaded via address-amplitude encodings. Because the quantum Fourier transform turns that shift into a diagonal multiplication, any such layer factors into independent operations on each frequency mode. If the factorization is correct, it supplies a systematic way to build deeper networks that respect the spatial symmetry of the data while keeping the expected gradient norm bounded away from zero as depth grows. A reader should care because the construction removes the mismatch between typical qubit-permutation circuits and the actual translation symmetry induced by the encoding.

Core claim

Conjugation by the quantum Fourier transform diagonalizes translations, therefore any unitary commuting with the pixel cyclic shift operator is realized by applying a product of independent unitaries to the Fourier modes and then applying the inverse transform.

What carries the argument

The Fourier-mode multiplexer: after the quantum Fourier transform conjugates the shift operator to a diagonal phase operator, the multiplexer applies an arbitrary unitary independently to each eigenmode before the inverse transform restores the original basis.

If this is right

Deep PCS-QCNN architectures can be assembled using measurement-induced pooling and deferred conditioning.
Inter-layer cancellation of the quantum Fourier transform reduces total circuit depth.
The expected squared gradient norm remains bounded below by a positive constant independent of depth at random initialization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same diagonalization technique could be applied to other group symmetries by replacing the Fourier transform with the appropriate representation-theoretic transform.
Because the construction separates frequency modes explicitly, it may allow selective regularization or pruning of high-frequency components during training.
The absence of a depth-induced barren plateau in this regime suggests that scaling the network width rather than depth may be the more immediate bottleneck for expressivity.

Load-bearing premise

The data encoding produces an exact cyclic shift symmetry on the index register that every layer must preserve without approximation.

What would settle it

A direct matrix calculation on a two-qubit register showing that a constructed layer fails to commute with the shift operator, or a numerical sampling of the gradient variance at initialization that drops exponentially with added layers.

Figures

Figures reproduced from arXiv: 2604.06094 by Dmitry Chirkov, Igor Lobanov.

**Figure 1.** Figure 1: FIG. 1. Schematic comparison of qubit- vs pixel-translation symmetry for address/amplitude encodings. (a) A typical QCNN [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: FIG. 2. Fixed benchmark controls used to expose translation-sensitive inductive bias. Training uses a balanced subset of 1000 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3. Schematic of a multilayer PCS-QCNN. Each non-final layer maps the active index registers to the Fourier basis with [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Overview of the hybrid model used in our exper [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

**Figure 5.** Figure 5: FIG. 5. Infinite-shot test-accuracy dynamics for the PCS-QCNN sweeps. (a) Translated-MNIST architecture sweep over [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

**Figure 6.** Figure 6: FIG. 6. Initialization-time quantum-gradient diagnostics in [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

**Figure 7.** Figure 7: FIG. 7. Finite-shot reevaluation of one representative direct [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗

**Figure 8.** Figure 8: FIG. 8. Batch-mean test cross-entropy distributions for one representative direct-16 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗

read the original abstract

Convolutional neural networks owe much of their success to hard-coding translation equivariance. Quantum convolutional neural networks (QCNNs) have been proposed as near-term quantum analogues, but the relevant notion of translation depends on the data encoding. For address/amplitude encodings such as FRQI, a pixel shift acts as modular addition on an index register, whereas many MERA-inspired QCNNs are equivariant only under cyclic permutations of physical qubits. We formalize this mismatch and construct QCNN layers that commute exactly with the pixel cyclic shift (PCS) symmetry induced by the encoding. Our main technical result is a constructive characterization of all PCS-equivariant unitaries: conjugation by the quantum Fourier transform (QFT) diagonalizes translations, so any PCS-equivariant layer is a Fourier-mode multiplexer followed by an inverse QFT (IQFT). Building on this characterization, we introduce a deep PCS-QCNN with measurement-induced pooling, deferred conditioning, and inter-layer QFT cancellation. We also analyze trainability at random initialization and prove a lower bound on the expected squared gradient norm that remains constant in a depth-scaling regime, ruling out a depth-induced barren plateau in that sense.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper delivers a clean QFT-based characterization of all PCS-equivariant QCNN layers for FRQI-style encodings and a depth-independent gradient lower bound, but the measurement pooling step needs explicit verification to confirm end-to-end equivariance.

read the letter

The core advance is a constructive description of every unitary that commutes with the pixel cyclic shift induced by address encodings: conjugate by the QFT, apply an arbitrary diagonal operator on the Fourier modes, then conjugate back. This is new relative to earlier QCNN work that only handled qubit permutations, and it directly gives a practical layer template. They stack these layers into a deeper network, insert measurement-induced pooling with deferred conditioning, cancel the QFTs between layers to reduce gate count, and prove that the expected squared gradient norm at random initialization stays bounded below by a positive constant independent of depth. That last claim is useful because it rules out a depth-driven barren plateau under their model assumptions. The unitary characterization itself rests on standard facts about the cyclic group and the QFT, so it looks internally solid and reproducible. The gradient bound is presented as a derivation rather than a numerical fit, which is the right kind of evidence. The softer spot is the non-unitary part. The abstract and stress-test note both flag that equivariance must survive the pooling measurements and conditioning; if the chosen basis and deferred operations do not commute with the shift in the required way, the end-to-end map loses the symmetry even though each unitary layer preserves it. The paper claims to handle this, but without seeing the explicit commutation check it is the least-secured step. The trainability result also assumes noise-free random initialization, so its practical reach on hardware remains open. This work is aimed at researchers building symmetry-aware quantum circuits for image-like data or studying barren-plateau avoidance in structured ansatze. A reader who already works on QCNNs or group-equivariant quantum machine learning will find the characterization and the gradient bound worth their time. The math is grounded enough and the claims are specific enough that the paper deserves a serious referee rather than a desk rejection, even if the pooling details require tightening in revision.

Referee Report

1 major / 2 minor

Summary. The paper formalizes the mismatch between pixel cyclic shift (PCS) symmetry induced by address/amplitude encodings (e.g., FRQI) and standard qubit-permutation equivariance in QCNNs. It gives a constructive characterization of all PCS-equivariant unitaries: any such unitary is a QFT, followed by a Fourier-mode multiplexer, followed by an IQFT. Building on this, the authors construct a deep PCS-QCNN architecture that incorporates measurement-induced pooling with deferred conditioning and inter-layer QFT cancellation. They additionally prove a lower bound on the expected squared gradient norm at random initialization that remains constant with depth in a specified scaling regime, indicating the absence of a depth-induced barren plateau.

Significance. If the characterization and the extension to the full architecture hold, the work supplies a symmetry-principled design rule for QCNNs that respects the translation symmetry native to common quantum image encodings. The constructive nature of the unitary characterization (leveraging the diagonalizing action of the QFT on cyclic shifts) and the explicit gradient-norm lower bound are clear strengths that could guide future equivariant quantum models. These elements address both architectural consistency and trainability, which are central concerns for near-term quantum machine learning on structured data.

major comments (1)

[§4.2] §4.2 (Measurement-Induced Pooling and Deferred Conditioning): The unitary characterization (Theorem 1) is internally consistent, but the central claim that the end-to-end PCS-QCNN remains equivariant requires an explicit argument that the non-unitary pooling step preserves the symmetry—i.e., that a PCS on the input produces a correspondingly shifted output distribution after measurement and conditioning. The manuscript invokes inter-layer QFT cancellation but does not supply a self-contained lemma or calculation showing commutation for the chosen measurement basis; this step is load-bearing for the deep-architecture claim.

minor comments (2)

[§3.1] The definition of the 'Fourier-mode multiplexer' is introduced conceptually but would benefit from an explicit operator equation (e.g., as a diagonal gate in the Fourier basis with mode-dependent phases) to make the construction fully reproducible from the text.
[§5] In the gradient analysis, the precise depth-scaling regime (number of layers relative to qubit count, and whether measurements are included in the random-initialization ensemble) should be stated more explicitly to allow direct verification of the constant lower bound.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading of the manuscript and for identifying a point that requires clarification to fully substantiate the end-to-end equivariance claim. We address the major comment below and have revised the manuscript to incorporate an explicit supporting argument.

read point-by-point responses

Referee: [§4.2] §4.2 (Measurement-Induced Pooling and Deferred Conditioning): The unitary characterization (Theorem 1) is internally consistent, but the central claim that the end-to-end PCS-QCNN remains equivariant requires an explicit argument that the non-unitary pooling step preserves the symmetry—i.e., that a PCS on the input produces a correspondingly shifted output distribution after measurement and conditioning. The manuscript invokes inter-layer QFT cancellation but does not supply a self-contained lemma or calculation showing commutation for the chosen measurement basis; this step is load-bearing for the deep-architecture claim.

Authors: We agree that the equivariance of the full architecture, including the non-unitary measurement-induced pooling, requires an explicit verification beyond the unitary characterization in Theorem 1. In the revised manuscript we have added a self-contained Lemma 2 in §4.2 that directly addresses this gap. The lemma shows that, for the computational-basis measurement performed after the final IQFT and under the deferred-conditioning protocol, a PCS applied to the input state produces a correspondingly shifted distribution over the measured outcomes. The argument proceeds by (i) using the fact that the preceding Fourier-mode multiplexer is diagonal in the QFT basis, (ii) verifying that the inter-layer QFT/IQFT cancellation restores the computational-basis measurement to a PCS-equivariant operation, and (iii) confirming that the classical conditioning step, which depends only on the measured index, commutes with the residual shift. We have also updated the surrounding discussion in §4.2 and the architecture overview to reference Lemma 2 when asserting end-to-end PCS equivariance. This addition makes the load-bearing step fully rigorous while preserving the original construction. revision: yes

Circularity Check

0 steps flagged

No circularity: central characterization follows from standard QFT diagonalization of cyclic shifts

full rationale

The paper's main technical result is a constructive characterization of PCS-equivariant unitaries via QFT conjugation and Fourier-mode multiplexing. This is a direct application of the known fact that the quantum Fourier transform diagonalizes the cyclic shift operator on the index register (a standard result in quantum computing and group Fourier analysis, not derived from or fitted to the paper's own data or definitions). The subsequent construction of the deep PCS-QCNN (including measurement-induced pooling and inter-layer QFT cancellation) and the proof of the constant lower bound on expected squared gradient norm are presented as explicit derivations and proofs without reducing any claimed prediction or equivariance property to a fitted parameter, self-definition, or load-bearing self-citation. The end-to-end equivariance extension is claimed via the unitary characterization plus additional architectural choices, but does not collapse by construction to the inputs. No steps match the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The construction assumes standard quantum circuit axioms (unitary evolution, QFT properties) and a domain-specific assumption about the encoding-induced symmetry; no free parameters or new physical entities are introduced.

axioms (2)

domain assumption Address/amplitude encodings such as FRQI induce modular addition (pixel cyclic shift) on the index register.
Stated in the abstract as the relevant notion of translation for the symmetry.
standard math The quantum Fourier transform diagonalizes the cyclic shift operator.
Invoked as the key technical step in the characterization.

invented entities (1)

Fourier-mode multiplexer no independent evidence
purpose: To construct all PCS-equivariant layers after QFT conjugation.
New circuit primitive introduced to realize the equivariant unitaries.

pith-pipeline@v0.9.0 · 5506 in / 1363 out tokens · 41097 ms · 2026-05-10T19:16:48.567928+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Theorem 1 (Fourier-multiplexer form of PCS-equivariant unitaries). ... U = (F†_N ⊗ I) B (F_N ⊗ I), where B is block-diagonal in the Fourier basis
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean embed_eq_pow unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The shift operator T is diagonal in the Fourier basis: F_N T F†_N = diag(ω^k)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

59 extracted references · 6 canonical work pages

[1]

These histograms use the same trained model

and for several shot budgets. These histograms use the same trained model. For each epoch and shot budget, the 10,000-sample test set is split into 40 nonoverlapping batches of 250 images; each finite-shot histogram entry is one batch-mean cross-entropy under one of 100 indepen- dent resampling passes, giving 40×100 = 4000 entries per shot budget. The inf...
[2]

LeCun, B

Y. LeCun, B. E. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel, Neural Compu- tation1, 541 (1989)

1989
[3]

Goodfellow, Y

I. Goodfellow, Y. Bengio, and A. Courville,Deep Learn- ing(MIT Press, 2016)

2016
[4]

Preskill, Quantum2, 79 (2018)

J. Preskill, Quantum2, 79 (2018)

2018
[5]

Vidal, Physical Review Letters99, 220405 (2007)

G. Vidal, Physical Review Letters99, 220405 (2007)

2007
[6]

I. Cong, S. Choi, and M. D. Lukin, Nature Physics15, 1273 (2019)

2019
[7]

S. Oh, J. Choi, J. K. Kim, and J. Kim, inProceedings of the 2021 International Conference on Information Net- working (ICOIN)(IEEE, 2021) pp. 50–52

2021
[8]

Y. Jing, X. Li, Y. Yang, C. Wu, W. Fu, W. Hu, Y. Li, and H. Xu, Quantum Information Processing21, 101 (2022)

2022
[9]

Huang, W.-J

S.-Y. Huang, W.-J. An, D.-S. Zhang, and N.-R. Zhou, Optics Communications533, 129287 (2023)

2023
[10]

Wei, Y.-H

S.-J. Wei, Y.-H. Chen, Z.-R. Zhou, and G.-L. Long, APL Photonics7, 056103 (2022)

2022
[11]

Li, R.-G

Y. Li, R.-G. Zhou, R. Xu, J. Luo, and W. Hu, Quantum Science and Technology5, 044003 (2020)

2020
[12]

Li, P.-C

W. Li, P.-C. Chu, G.-Z. Liu, Y.-B. Tian, T.-H. Qiu, and S.-M. Wang, Quantum Engineering2022, 5701479 (2022)

2022
[13]

Easom-McCaldin, A

P. Easom-McCaldin, A. Bouridane, A. Belatreche, and R. Jiang, IEEE Access9, 65127 (2021)

2021
[14]

Das and F

S. Das and F. Caruso, Quantum Science and Technology 10, 015030 (2024)

2024
[15]

Larocca, F

M. Larocca, F. Sauvage, F. M. Sbahi, G. Verdon, P. J. Coles, and M. Cerezo, PRX Quantum3, 030341 (2022)

2022
[16]

J. J. Meyer, M. Mularski, E. Gil-Fuster, A. A. Mele, F. Arzani, A. Wilms, and J. Eisert, PRX Quantum4, 010328 (2023)

2023
[17]

Schatzki, M

L. Schatzki, M. Larocca, Q. T. Nguyen, F. Sauvage, and M. Cerezo, npj Quantum Information10, 12 (2024)

2024
[18]

Chinzei, Q

K. Chinzei, Q. H. Tran, K. Maruyama, H. Oshima, and S. Sato, Physical Review Research6, 023042 (2024). 17

2024
[19]

P. Q. Le, F. Dong, and K. Hirota, Quantum Information Processing10, 63 (2011)

2011
[20]

Zhang, K

Y. Zhang, K. Lu, Y. Gao, and M. Wang, Quantum In- formation Processing12, 2833 (2013)

2013
[21]

Gong, J.-J

L.-H. Gong, J.-J. Pei, T.-F. Zhang, and N.-R. Zhou, Op- tics Communications550, 129993 (2024)

2024
[22]

J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Nature Communications9, 4812 (2018)

2018
[23]

Pesah, M

A. Pesah, M. Cerezo, S. Wang, T. Volkoff, A. T. Sorn- borger, and P. J. Coles, Physical Review X11, 041011 (2021)

2021
[24]

R. M. Gray, Foundations and Trends in Communications and Information Theory2, 155 (2006)

2006
[25]

G. H. Golub and C. F. Van Loan,Matrix Computations, 4th ed. (Johns Hopkins University Press, 2013)

2013
[26]

Y. L¨ u, Q. Gao, J. L¨ u, M. Ogorza lek, and J. Zheng, inPro- ceedings of the 40th Chinese Control Conference (CCC) (IEEE, 2021) pp. 6329–6334

2021
[27]

M. A. Nielsen and I. L. Chuang,Quantum Computa- tion and Quantum Information: 10th Anniversary Edi- tion(Cambridge University Press, Cambridge, 2010)

2010
[28]

Cleve and J

R. Cleve and J. Watrous, inProceedings of the 41st An- nual Symposium on Foundations of Computer Science (IEEE, 2000) pp. 526–536

2000
[29]

Bergholm, J

V. Bergholm, J. J. Vartiainen, M. M¨ ott¨ onen, and M. M. Salomaa, Physical Review A71, 052330 (2005)

2005
[30]

V. V. Shende, S. S. Bullock, and I. L. Markov, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems25, 1000 (2006)

2006
[31]

LeCun, C

Y. LeCun, C. Cortes, and C. J. C. Burges, MNIST hand- written digit database,http://yann.lecun.com/exdb/ mnist/(2010), accessed 2026-04-23

2010
[32]

Schuld and N

M. Schuld and N. Killoran, PRX Quantum3, 030101 (2022)

2022
[33]

Cerezo, G

M. Cerezo, G. Verdon, H.-Y. Huang, L. Cincio, and P. J. Coles, Nature Computational Science2, 567 (2022)

2022
[34]

Khoshaman, W

A. Khoshaman, W. Vinci, B. Denis, E. Andriyash, H. Sadeghi, and M. H. Amin, Quantum Science and Tech- nology4, 014001 (2018)

2018
[35]

I. V. Oseledets, SIAM Journal on Scientific Computing 33, 2295 (2011)

2011
[36]

Chirkov and I

D. Chirkov and I. Lobanov, Pixel-Translation- Equivariant Quantum Convolutional Neural Net- works via Fourier Multiplexers, Zenodo (2026), DOI: 10.5281/zenodo.19814130

work page doi:10.5281/zenodo.19814130 2026
[37]

Bowles, S

J. Bowles, S. Ahmed, and M. Schuld, Better than clas- sical? the subtle art of benchmarking quantum machine learning models (2024), arXiv:2403.07059 [quant-ph]

work page arXiv 2024
[38]

Mattern, D

D. Mattern, D. Martyniuk, H. Willems, F. Bergmann, and A. Paschke, Variational quanvolutional neu- ral networks with enhanced image encoding (2021), arXiv:2106.07327 [cs.CV]

work page arXiv 2021
[39]

Huggins, P

W. Huggins, P. Patil, B. Mitchell, K. B. Whaley, and E. M. Stoudenmire, Quantum Science and Technology4, 024001 (2019)

2019
[40]

Zheng, Q

J. Zheng, Q. Gao, J. L¨ u, M. Ogorza lek, Y. Pan, and Y. L¨ u, Journal of the Franklin Institute360, 13761 (2023)

2023
[41]

T. Hur, L. Kim, and D. K. Park, Quantum Machine In- telligence4, 3 (2022)

2022
[42]

Huang, X

F. Huang, X. Tan, R. Huang, and Q. Xu, Physica A: Statistical Mechanics and its Applications605, 128067 (2022)

2022
[43]

Discovering transforms: A tutorial on circulant matrices, circular convolution, and the discrete fourier transform.arXiv:1805.05533,

B. Bamieh, Discovering transforms: A tutorial on cir- culant matrices, circular convolution, and the discrete fourier transform (2022), arXiv:1805.05533 [eess.SP]

work page arXiv 2022
[44]

Z. Li, Y. Zhang, and S. Arora, inProceedings of the 9th International Conference on Learning Representa- tions (ICLR)(2021) arXiv:2010.08515 [cs.LG]

work page arXiv 2021
[45]

Lahoti, S

A. Lahoti, S. Karp, E. Winston, A. Singh, and Y. Li, inProceedings of the 12th International Conference on Learning Representations (ICLR)(2024)

2024
[46]

Coppersmith, An approximate fourier transform useful in quantum fac- toring, arXiv:quant-ph/0201067 (2002)

D. Coppersmith, An approximate fourier transform use- ful in quantum factoring (2002), arXiv:quant-ph/0201067 [quant-ph]. 18 Supplemental Material for “Pixel-Translation-Equivariant Quantum Convolutional Neural Networks via Fourier Multiplexers” This document includes a literature table for QCNN-style MNIST results (Sec. A), additional experiments and ana...

work page arXiv 2002
[47]

Hyperparameter sweep and parameter accounting The experiments reported here use two sweep families. Figure 5(a) is a translated-MNIST architecture sweep with digits resized to 16×16, placed on a 32×32 canvas, translated by at most 8 pixels along each axis, evaluated with full readout, and trained for 2000 epochs over 3 seeds, withQ∈ {1,2,3,4,5}andn f ∈ {1...

2000
[48]

The article-wide convention (a, b) = (0, π) was chosen a priori as a one-period real FRQI color-map convention and was not selected by optimizing this sweep

Brightness-range sensitivity sweep for encoder scaling Besides the main translated-MNIST architecture sweep and the full-MNIST size sweep, we ran a separate low-data sensitivity sweep for the encoder angle scale. The article-wide convention (a, b) = (0, π) was chosen a priori as a one-period real FRQI color-map convention and was not selected by optimizin...
[49]

2(a) are specified here for reproducibility (text Secs

Classical baseline architectures for benchmark comparisons The classical reference models used in Fig. 2(a) are specified here for reproducibility (text Secs. V F and VI C). For the CNN baseline (Fig. S2(a)), the input is a 32×32 grayscale canvas produced by the shared preprocessing protocol. In Fig. 2(a), this architecture is evaluated on the translated ...
[50]

The model architectures were kept identical to those specified in Sec

Full-MNIST classical control without translations The same CNN and MLP baselines were also evaluated on the full standard MNIST split (60,000 training images and 10,000 test images) after resizing each digit directly to a 32×32 grayscale image, with no canvas translation. The model architectures were kept identical to those specified in Sec. B 3, so the c...
[51]

6, distinguishing the norm of the empirical-loss gradient from the per-sample RMS gradient norm

Readout entropy diagnostic The main text reports initialization-time quantum-gradient diagnostics in Fig. 6, distinguishing the norm of the empirical-loss gradient from the per-sample RMS gradient norm. The readout entropy diagnostic below checks how finite-shot sampling changes the classifier input. Figure S4 shows the Shannon entropy of the full readout...

2000
[52]

Geometric interpretation: local loss landscape in readout space The loss-landscape diagnostic probes the classical head under finite-shot perturbations expected from the multi- nomial readout model. Figure S5 uses the representative direct-16×16 full-MNIST reference PCS-QCNN after 100 and 800 training epochs, shot budgetN= 128, and an 81×81 grid of coordi...

2000
[53]

Qualitative error structure Figure S6 reports the remaining classification errors for the representative final direct-16×16 full-MNIST reference model after 2000 training epochs (confusion matrix and examples of misclassified digits). These errors are not dominated by a single failure mode; rather, they reflect the expected ambiguity between visually simi...

2000
[54]

In many spatial problems, the same local pattern is applied at every position: each output feature at locationkdepends on inputs in the same relative way at any location

Classical convolution and translation symmetry A classical feedforward network is a composition of layers of the formz=f(Ax), whereAis a linear operator andf is a non-linear activation. In many spatial problems, the same local pattern is applied at every position: each output feature at locationkdepends on inputs in the same relative way at any location. ...
[55]

[5] and used in several subsequent works

MERA-QCNN as a QCS-equivariant architecture Figure 1(a) shows a MERA-inspired QCNN, originally proposed in Ref. [5] and used in several subsequent works. The architecture consists of repeated local unitaries arranged in a multiscale pattern, interleaved with pooling that reduces the number of active qubits, typically implemented by measurements followed b...
[56]

Fourier cancellation at the interface of PCS-QCNN layers Figure 3 shows that consecutive PCS layers contain adjacent inverse and forward Fourier transforms on the index registers. In our architecture the pooling step is implemented after the inverse QFT of a layer: in the computational index basis we split the fine index asr= 2q+s, measure the parity bits...
[57]

We begin with the qubit count

Complexity analysis In the NISQ regime it is important to keep both qubit requirements and circuit depth in view. We begin with the qubit count. For anN×Nimage withN= 2 nidx, address encoding uses two index registers withn idx = log 2 N 32 qubits each. In addition, we use a feature register of sizen f qubits, with the grayscale/color qubit included inn f ...
[58]

Each axis initially hasn idx qubits

PCS-QCNN family and parameter count We work with an index register consisting ofdspatial axes. Each axis initially hasn idx qubits. A depth-QPCS- QCNN pools after the firstQ−1 layers, each time measuring and discarding one computational-index qubit per axis. After pooling, the remaining index register has nl :=n idx −Q+ 1 (S11) qubits per axis, hence Didx...
[59]

For head input dimensionD out, this means Waz ∼Unif −D−1/2 out , D−1/2 out , b a ∼Unif −D−1/2 out , D−1/2 out , a∈[M], z∈[D out],(S33) independently over all entries

Aggregate gradient-energy estimate The benchmark initialization samples every quantum Pauli coefficient independently as θ(ℓ) k,m,α ∼Unif(0,2π).(S32) Independently, the classical head uses fan-in uniform initialization. For head input dimensionD out, this means Waz ∼Unif −D−1/2 out , D−1/2 out , b a ∼Unif −D−1/2 out , D−1/2 out , a∈[M], z∈[D out],(S33) in...