acceptodds
Under review as a conference paper at ICLR 2027

Shared Address Learning for Amplitude Encoding in Quantum Machine Learning

Abstract

Quantum machine learning offers expressive Hilbert-space models, but loading classical data involves a trade-off between compactness and cost. Amplitude encoding is highly compact, yet state preparation can consume much of the entangling-gate budget on near-term hardware. This cost depends heavily on the address, which assigns data coordinates to computational-basis states and is typically fixed to a default order. Sharing one address across a dataset preserves all inner products and the ideal amplitude kernel, allowing us to reshape entanglement for efficient preparation. We prove an exponential fidelity separation between addresses on a family of maximally entangled states under limited gate budgets, and show that a correlation-based construction makes the family completely separable. Shared Address Learning (SAL) builds candidate addresses from the correlation structure of unlabeled data, refines them with a lightweight structural score, and selects one by shallow-circuit fidelity. Learned once, the address is applied to all inputs as classical preprocessing compatible with any loader. For images and pretrained vision and language features, SAL reaches equivalent preparation fidelity with significantly fewer entangling gates and consistently improves tensor-network and adaptive-compilation loaders. These gains extend to quantum kernel and quantum neural network (QNN) classifiers in simulation and persist for a QNN deployed on a superconducting processor.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
quantum machine learning, amplitude encoding, quantum state preparation, quantum data loading, quantum kernels, quantum neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EwBnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Shared Address Learning (SAL) to reduce the cost of amplitude encoding. SAL learns a permutation from unlabeled training data that preserves the amplitude kernel but can simplify state preparation. The method to find the permutation combines tests on correlation, local refinement, and selection by preparation fidelity. Experiments show improvements across several datasets and loaders.

Critical strengths

*The paper addresses an important bottleneck in quantum machine learning.

*The shared learned permutation preserves the ideal kernel. This provides a clear basis for comparing preparation methods.

*The theoretical results explain how changing the address can reduce preparation requirements.

*Improvements are consistent across several datasets, representations and loaders in experiments.

*The combination of correlation-based construction and successive refinement is interesting.

Critical weaknesses

*Comparison with alternative mapping strategies is limited. The paper discusses Jeon et al. and distinguishes their input-specific ordering from SAL’s shared permutation. However, no direct numerical comparison is provided. The experiments demonstrate improvements over natural addressing, but do not establish how SAL compares with other mapping strategies. This remains a major concern for assessing its relative benefits.

*The practical search has limited guarantees. SAL selects the best candidate from a small search set. The recovery guarantee applies to a specific constructed family. The quality on general datasets is supported empirically.

*Robustness to distribution shift is unknown. The learned permutation depends on training-data structure. Its preparation advantage may decrease on shifted data, although the ideal kernel remains preserved.

Questions for authors

The paper presents a useful approach with clear theoretical motivation and consistent empirical support. Its main value is improved state preparation under limited gate budgets. My main reservation is the limited numerical comparison with alternative mapping strategies to make the relevance of the SAL stronger. The comparisons with established loaders are useful, but address a different question: whether SAL improves those loaders relative to natural addressing. A controlled comparison with one or more mappings strategies would strengthen the evidence for SAL’s specific advantages.

Questions:

*Can SAL be compared with an alternative mapping strategy under comparable resource constraints? Jeon et al. (for example) is a relevant candidate, although its objective differs. Please explain how a fair comparison could account for preparation fidelity, classical search cost and reuse across inputs. If such a comparison is not doable, please clarify why. This discussion should be clear in the main text.

*Does a frozen learned permutation still has its loading advantage under distribution shift, such as translated or corrupted MNIST images?What about the other datasets?

To strenghten the paper content one can also (suggestions): *Add a short MNIST example showing the permutation and subsequent circuit fit. *Make the distinction between theoretical optimal fidelity and achieved fidelity more clear. *Emphasise the gate budget more clearly.

5 comments

We thank the reviewer for the constructive assessment. The main reservation, a controlled comparison with an alternative mapping strategy, is now in the main text (Sec. 4.2, Table 2), organized along the three axes the reviewer proposed.

W1 / Q1. Can SAL be compared with an alternative mapping strategy under comparable resource constraints? Jeon et al. (for example) is a relevant candidate, although its objective differs. Please explain how a fair comparison could account for preparation fidelity, classical search cost and reuse across inputs. [...] This discussion should be clear in the main text.

Yes. Jeon et al. search a qubit (tensor-axis) ordering of an MPS for each input, so we compare at matched tensor-train (TT) rank, the resource both methods share. All arms use Jeon et al.'s released routine unchanged, one normalized TT-SVD scorer at rank χ\chi and the 50 held-out targets per dataset of Table 10 (App. D.3); at χ=2\chi=2 the scorer reproduces the one-layer MPS loader of Table 10. SAL arms are mean ± std over three source runs; the other arms are run-independent.

DatasetχNaturalGrayJeon (per input)SALSAL, then Jeon
MNIST20.6190.6580.7000.689 ± 0.0070.776 ± 0.002
40.8780.8770.9160.896 ± 0.0000.947 ± 0.002
CIFAR-1020.3280.3310.3800.587 ± 0.0010.611 ± 0.004
40.4850.4880.5390.671 ± 0.0030.700 ± 0.002
SST-220.0620.0630.1080.265 ± 0.0030.298 ± 0.003
40.1870.1910.2500.365 ± 0.0040.409 ± 0.003

Preparation fidelity. On pretrained features the shared address alone exceeds per-input search by 0.12–0.21, with a gain over the natural order about three to five times larger. On MNIST, whose row-major order already carries spatial locality, the two are comparable. They compose: SAL followed by the search is best in every row.

Classical search cost. SAL is paid once per dataset, from unlabeled data: correlation statistics on 256 vectors and screening of at most 12 candidates on 128 inputs. A new input then costs one O(d)O(d) index gather. Per-input search is repeated for every input, at 10410^4–10510^5 SVD evaluations per feature vector in our runs.

Reuse across inputs. A shared address gives all inputs one coordinate system, so the ideal kernel and a single downstream model carry across the dataset. Per-input orderings give each input its own coordinates, so the same qubit carries different index bits for different inputs. This suits compressing individual states, but a learner across inputs must then undo each ordering before inputs can be compared.

Why the two differ. Qubit orderings form the n!n!-element bit-permutation subgroup of GL(n,2)\mathrm{GL}(n,2), the binary maps SAL searches. Optimal fidelity (Eq. 2) is invariant under qubit relabeling, so an ordering can adapt a fixed-chain loader to a state but cannot reduce the state's entanglement; a binary-linear address can.

  • Theorem 1 holds under every qubit ordering, even one chosen per input, while one shared XOR address reaches fidelity one at G=0G=0.
  • Proposition 4 (App. A.4) gives the TT analogue: one shared binary address maps sixteen orthogonal 15-qubit states to product states, whereas under any ordering, even one chosen per state, every rank-χ\chi TT approximation has fidelity at most min⁡(1,χ/8)\min(1,\chi/8).

W3 / Q2. Does a frozen learned permutation still has its loading advantage under distribution shift, such as translated or corrupted MNIST images? What about the other datasets?

Yes on MNIST, under every perturbation we tried. On the other datasets, the advantage carries over between datasets that share their dominant direction and is relearned otherwise. We froze the address and topology learned on clean MNIST and evaluated perturbed held-out images (App. G.5, Table 32; 128 fixed targets, five runs):

Condition4 CX: natural4 CX: SAL20 CX: natural20 CX: SAL
Clean0.470 ± 0.0000.587 ± 0.0070.716 ± 0.0200.793 ± 0.005
Translation, ±2 px (4 directions)0.475 ± 0.0000.549 ± 0.0040.713 ± 0.0180.774 ± 0.008
Gaussian noise, σ = 0.10.496 ± 0.0000.596 ± 0.0070.706 ± 0.0170.772 ± 0.004
Gaussian blur, σ = 1 px0.570 ± 0.0000.697 ± 0.0070.813 ± 0.0210.876 ± 0.004
Contrast × 0.50.634 ± 0.0000.711 ± 0.0060.797 ± 0.0140.846 ± 0.003
Salt-and-pepper, 5%0.408 ± 0.0000.497 ± 0.0060.598 ± 0.0160.656 ± 0.003
Rotation −10°0.481 ± 0.0000.600 ± 0.0110.723 ± 0.0220.807 ± 0.005
Rotation +10°0.484 ± 0.0000.610 ± 0.0070.743 ± 0.0190.810 ± 0.006

The size of the advantage follows the structure the address encodes:

  • Rotations and blur keep the digit centered and its local correlations intact, and largely preserve the gain.
  • Translations move content relative to the learned layout, and pixel-independent noise adds structure that no address can exploit; both reduce the gain moderately.
  • Contrast reduction flattens images toward the uniform state, which raises the natural fidelity and leaves less headroom.

Other datasets. For feature data we tested the strongest form of shift, a change of dataset under the same frozen encoder (App. G.6; details in our reply to tCd3, Q5). The frozen address keeps 89–95% of its gain between CIFAR-10 and CIFAR-100, most of it from MNIST to KMNIST at G=0G=0, and almost none between SST-2 and AG News. What decides this is whether the two datasets share the dominant direction u1u_1 of E[xx⊤]\mathbb E[xx^\top], on which SAL anchors its address; the cosine between source and target u1u_1 (0.985, 0.842 and 0.195) is computable from unlabeled samples before reuse.

W2. The practical search has limited guarantees. SAL selects the best candidate from a small search set. The recovery guarantee applies to a specific constructed family. The quality on general datasets is supported empirically.

We agree that exact recovery is proven only for a structured family (Theorem 2), as is usual for searches over ∣GL(10,2)∣≈3.7×1029|\mathrm{GL}(10,2)|\approx3.7\times10^{29} maps. The guarantees that matter for deployment, however, hold on every dataset, and where optimality can be measured, the search comes close to it.

  • Dataset-independent guarantees. Whatever address is selected, the ideal kernel is preserved exactly (Eq. 3). Eq. (4) and Corollary 3 bound how any address can change a kernel learner's predictions. Eq. (9) gives every address the floor FˉG⋆(P)≥λ1FG⋆(P;u1)\bar F^\star_G(P)\ge\lambda_1F^\star_G(P;u_1), which the principal-reference start is designed to raise.
  • Measured optimality. The screened address ties with the best member of its bank at GrefG_{\mathrm{ref}}; random family members, even screened, stay well below SAL (Table 25); and screening fidelity ranks candidates consistently with KRR validation accuracy (Spearman 0.79 on MNIST, 0.85 on CIFAR-10).

Suggestions. Add a short MNIST example showing the permutation and subsequent circuit fit. Make the distinction between theoretical optimal fidelity and achieved fidelity more clear. Emphasise the gate budget more clearly.

  1. MNIST example. Sec. 4.5 opens with the learned MNIST address as XOR rules over pixel row and column bits, a two-dimensional Gray code (all ten rules in Table 26). The fitted circuit of the same run is Fig. 7 (App. C.4), and prepared states at three budgets are in Fig. 14.
  2. Optimal versus achieved fidelity. Sec. 2.3, directly after Eq. (2), now states that the theory concerns the optimum FG⋆F^\star_G, while experiments report the achieved fidelity F^≤FG⋆\widehat F\le F^\star_G of a specific loader.
  3. Gate budget. GG counts the CX gates of the preparation circuit (Sec. 2.3). Sec. 4.1 now states the range in absolute terms, 0–80 CX at n=10n=10, under a tenth of the order-2n2^n CX of exact generic preparation.

We thank the reviewer again; the comparison has made SAL's specific advantages considerably clearer.

Thank you for the detailed response and additional experiments. The comparison with Jeon et al. at matched TT rank substantially addresses my main concern. The distribution-shift experiments also strengthen the paper. While the practical search remains supported mainly by empirical evidence on general datasets, I find the response convincing overall and maintain my weak acceptance. This is also due to my limited expertise in the specific field.

One small follow-up: could you give an approximate runtime comparison, including SAL’s initial learning cost? Roughly how many inputs are needed before reusing SAL becomes cheaper than running the per-input search?

Reviewer_tCd3OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper observes that amplitude encoding fixes a "natural address" - the assignment of data coordinates to basis states - and that this choice, not the data, determines the entanglement profile a loader must reproduce. A single shared permutation applied to every input preserves all inner products and the ideal amplitude kernel exactly, while reshaping entanglement. The authors prove an exponential fidelity separation between addresses on a family of maximally entangled states under bounded gate budgets, and give a correlation-based construction that makes the family separable. Shared Address Learning (SAL) then builds candidate addresses from the correlation structure of unlabeled data, refines them with a structural score, and selects one by shallow-circuit fidelity. Across eight datasets the address reduces entangling-gate cost (CIFAR-10 fidelity 0.315 -> 0.591 at 22 CX, exceeding the natural address at 88 CX), improves the independent MPS, AQCE and AQER loaders in all 91 dataset-budget-loader comparisons, and raises kernel and QNN classifier accuracy in all 32 and 6 comparisons respectively, with a hardware result on ibm_kingston (57.92% -> 71.04% on four-class MNIST).

Critical strengths
  1. A correct and unusually clean theoretical core with a matching construction. The separation result exhibits an exponential gap between addresses under bounded budgets, the correlation construction achieves separability, and the recovery guarantee matches the construction - a coherent chain rather than three separate claims.

  2. An empirical programme that is broad, controlled and consistent. Eight datasets, three independent loaders, a fast loader, 91/91, 32/32 and 6/6 consistency, construction-information ablations, the sorted-input negative control, and a hardware demonstration. The claim that all gains come from unlabeled data is supported by a properly designed intervention set, not just asserted.

Critical weaknesses
  1. There is no random-permutation control, so the experiments do not show that learning the address is what produces the gain. Every comparison is against the natural address. A learned permutation is only interesting if it beats a random one, and the paper never reports the fidelity of a random permutation from the same search space under the same budgets. The paper's own theory says some addresses are exponentially bad, which is precisely why "not the natural order" is not a sufficient explanation. The construction ablations do not substitute for this control, since they intervene on the construction input rather than on the selected permutation.

  2. The downstream benefit is bounded by the loader, and the paper's own numbers show the address does not help in the ideal limit. Because the shared address leaves the ideal kernel invariant, every downstream gain is a reduction in approximation penalty, not a modelling gain. On MNIST the ideal-SAL model is 0.88 points below the ideal-natural model, so a perfect loader would leave the address with nothing to offer. This is an honest and interesting decomposition, but it caps the QML contribution at the loader's budget.

  3. The closest published work on the same phenomenon is not cited. Salmenperä, Perkkola and Nurminen, "Feature permutation for quantum machine learning", Quantum Machine Intelligence 7(2):107 (2025), DOI 10.1007/s42484-025-00337-6, studies exactly this effect - that the order in which features are uploaded affects QML performance, and that the wrong permutation can ruin a model - and was published before the contemporaneous-work cutoff. Neither it nor Liu et al., "An efficient quantum algorithm for the preparation of arbitrary states", Quantum Information Processing 32:16 (2024), on Gray-code recursive preparation of amplitude-encoded states, appears in the references.

Questions for authors

1.What is the preparation fidelity of a random permutation drawn from the same candidate space, under the same budgets and seeds? Without this, the comparison against the natural address does not establish that learning the address matters.

  1. How much of the gain comes from the search family rather than from the search? What does a random member of the candidate family achieve, and what headroom does the family have?

  2. Please position SAL against Salmenperä et al. (QMI 2025) and against Liu et al. (QIP 2024). Is a shared, data-dependent address the natural continuation of the former, and how does the correlation-based construction compare with Gray-code ordering?

  3. Since the ideal kernel is invariant and the ideal-SAL model is slightly worse on MNIST, what is the theoretical bound on the downstream gain as a function of the loader budget? Is there any regime in which the address itself helps the ideal model rather than only the approximation penalty?

  4. Is the learned address transferable across datasets sharing feature semantics, and does it survive being re-learned on a disjoint subset of the same distribution?

4 comments

We thank the reviewer for a precise review. The requested controls make the role of learning explicit.

W1. There is no random-permutation control, so the experiments do not show that learning the address is what produces the gain. [...] The paper's own theory says some addresses are exponentially bad, which is precisely why "not the natural order" is not a sufficient explanation.

Q1. What is the preparation fidelity of a random permutation drawn from the same candidate space, under the same budgets and seeds?

We agree that this control was missing; it shows that learning produces the gain. We drew uniformly random invertible binary maps on SAL's two base orders, i.e., from the same space {PQ,A:Q∈{I,Qref}, A∈GL(n,2)}\{P_{Q,A}: Q\in\{I,Q_{\mathrm{ref}}\},\ A\in\mathrm{GL}(n,2)\}, and evaluated them with SAL's loader, budgets, seeds and screening (App. F.5, Table 25). Fidelity on 128 fixed held-out targets at GrefG_{\mathrm{ref}}:

AddressMNIST (20 CX)CIFAR-10 (22 CX)SST-2 (20 CX)
Natural0.716 ± 0.0200.313 ± 0.0070.103 ± 0.001
Fixed Gray code0.767 ± 0.0060.356 ± 0.0040.102 ± 0.001
Random member0.375 ± 0.0310.364 ± 0.0150.157 ± 0.006
Random, screened by SAL0.509 ± 0.0230.400 ± 0.0190.230 ± 0.008
Unscreened member of SAL's bank0.682 ± 0.0030.511 ± 0.0050.182 ± 0.007
SAL0.793 ± 0.0050.591 ± 0.0020.269 ± 0.004
  • As the reviewer infers from Theorem 1, a random member is worse than the natural order on MNIST (0.375 vs 0.716).
  • Screening random members with SAL's unlabeled procedure narrows the gap but stays below SAL on every dataset. When random members are mixed into SAL's bank, screening prefers the constructed candidates.
  • SAL stays above every control at GlowG_{\mathrm{low}} as well (Table 25).

We also agree that the construction ablations "intervene on the construction input rather than on the selected permutation"; this control intervenes on the permutation itself.

What learning produces. On MNIST the correlation construction learns a Gray code by itself, in two dimensions. Write a pixel's row and column as r4⋯r0r_4\cdots r_0 and c4⋯c0c_4\cdots c_0. Eight of the ten learned qubits are parities of adjacent bits, interleaved across the two axes: q0=r0⊕r1q_0=r_0\oplus r_1, q1=c0⊕c1q_1=c_0\oplus c_1, q2=r1⊕r2q_2=r_1\oplus r_2, q3=c1⊕c2q_3=c_1\oplus c_2, and so on (Table 26). Neighboring pixels then differ in one address bit (two across three row boundaries), against up to five under the row-major index. A fixed Gray code of the row-major index also makes horizontal steps one bit, but its bit c4⊕r0c_4\oplus r_0 couples the column field to the lowest row bit, so every vertical step costs two.

Q2. How much of the gain comes from the search family rather than from the search? What does a random member of the candidate family achieve, and what headroom does the family have?

Most of the gain comes from the search, not from the family. Family versus search. Measured from the natural order, a random family member (drawn on both of SAL's base orders, including the data-derived QrefQ_{\mathrm{ref}}) contributes none of SAL's gain on MNIST (it loses 0.34) and 18% (CIFAR-10) and 33% (SST-2) on the features. The rest comes from the data-driven construction and selection. The local search contributes as well: from the same starts and with the same number of moves, score-guided XOR refinement proposes better candidates than random moves (held-out fidelity 0.688 vs 0.633 on MNIST and 0.522 vs 0.468 on CIFAR-10, marginal on SST-2).

Headroom. Within the constructed bank, essentially none is left: in hindsight, the best bank member ties with the screened address at GrefG_{\mathrm{ref}}, where selection is made, and exceeds it by less than 0.01 at every other budget. For the whole family we cannot certify the optimum, since ∣GL(10,2)∣≈3.7×1029|\mathrm{GL}(10,2)|\approx3.7\times10^{29}. What can be measured is that the family contains many addresses far worse than the natural order, while the constructed bank sits in a high-fidelity region that random sampling, even screened, does not reach.

W2. [...] Because the shared address leaves the ideal kernel invariant, every downstream gain is a reduction in approximation penalty, not a modelling gain. On MNIST the ideal-SAL model is 0.88 points below the ideal-natural model, so a perfect loader would leave the address with nothing to offer. [...] it caps the QML contribution at the loader's budget.

Q4. [...] what is the theoretical bound on the downstream gain as a function of the loader budget? Is there any regime in which the address itself helps the ideal model rather than only the approximation penalty?

For kernel learners we agree, and the revision states the bound. For gate-based readouts the premise does not hold, and the answer to the second question is yes. The 0.88-point figure, which our original text did not attribute clearly, comes from the QNN (77.50% vs 76.62%, within one std; Table 15). The ideal kernel model is identical under both addresses by construction (96.72%).

Kernel learners: the bound. Eq. (4) bounds the kernel error by the mean infidelity, and the resolvent bound of App. A.2 turns it into a logit perturbation ηP,G(x)\eta_{P,G}(x). The ideal predictor and its top-two margin m(x)m(x) are shared by every address, so Corollary 3 gives

∣AccSAL(G)−AccN(G)∣≤Pr⁡x[m(x)≤2max⁡(ηN,G(x),ηSAL,G(x))].\lvert\mathrm{Acc}^{\mathrm{SAL}}(G)-\mathrm{Acc}^{\mathrm{N}}(G)\rvert\le\Pr_x\big[m(x)\le 2\max(\eta_{\mathrm{N},G}(x),\eta_{\mathrm{SAL},G}(x))\big].

The gain is confined to small-margin inputs, at a scale set by the budget, and vanishes with the infidelity. Budgets up to 8n8n show this (App. E.4, Table 16):

  • MNIST: 10.64 points at 4 CX, 0.86 at 20 CX and zero at 80 CX, where both addresses reach 96.56% (ideal 96.72%).
  • CIFAR-10: the natural address is still at fidelity 0.54 at 88 CX, and the gain is still 4.80 points.

Exact preparation of generic states needs order 2n2^n CX, so for hard-to-load data such as CIFAR-10 features, the cap is far from binding.

Gate-based readouts: the address changes the ideal model. A QNN computes ⟨x∣P†Uθ†OUθP∣x⟩\langle x|P^\dagger U_\theta^\dagger O U_\theta P|x\rangle (Eq. 1), and its model family is not invariant under PP. With exact inputs, i.e., a perfect loader:

Exact-input QNN (%)NaturalSAL
MNIST77.50 ± 0.9376.62 ± 1.24
CIFAR-1054.90 ± 2.0860.14 ± 0.55
SST-279.50 ± 2.3981.38 ± 1.97
Parity task (App. E.5)90.09 ± 20.91100.00 ± 0.00

On CIFAR-10 the QNN gain also persists at every budget from 6 to 88 CX (5.26–15.22 points).

Mechanism (Sec. 3.2, App. E.3). Under a binary address, the reduced state of qubit qq is ρq=12(I+cy(vq)X+zy(wq)Z)\rho_q=\tfrac12\big(I+c_y(v_q)X+z_y(w_q)Z\big), and its smaller eigenvalue is exactly the term of SAL's structural score. Lowering the score for loading therefore lengthens these Bloch vectors and moves strong shared correlations onto single-qubit observables, which a shallow readout accesses most directly.

  • For (∣00⟩±∣11⟩)/2(|00\rangle\pm|11\rangle)/\sqrt2, the separating observable X⊗XX\otimes X becomes a single XX under (a,b)↦(a⊕b,b)(a,b)\mapsto(a\oplus b,b), with all inner products preserved.
  • Row-major pixels already place neighbors on low address bits, which is why the two ideal MNIST QNNs agree. On encoder features, the address decides which correlations are local.

The contribution is therefore capped by the budget for kernel learners only; for gate-based readouts the address remains part of the model even with perfect loading.

W3 / Q3. Please position SAL against Salmenperä et al. (QMI 2025) and against Liu et al. (QIP 2024). Is a shared, data-dependent address the natural continuation of the former, and how does the correlation-based construction compare with Gray-code ordering?

Salmenperä et al. (now in Sec. 2.2). Yes, SAL can be read as a constructive continuation, carried over to amplitude encoding, where the mechanism changes. They show that in angle-encoded re-uploading models a feature permutation amounts to a change of ansatz and observable, and they leave open how to construct good dataset-specific permutations. In amplitude encoding:

  • an address acts on all 2n2^n basis states instead of relabeling nn qubits, so it can reshape entanglement and preparation cost;
  • it leaves the ideal kernel exactly invariant (Eq. 3), consistent with their remark that kernels are agnostic to input-independent permutations;
  • the model-level effect they describe reappears in our exact-input QNN (reply to Q4).

Liu et al. We take this to be Liu, Cheng, Loh and Li (QIP 23:161, 2024), now cited in Sec. 2.1; we will add another paper if one was meant. It Gray-codes the pixel-position register of NEQR, a basis encoding, to merge multi-controlled gates in exact preparation, as Möttönen et al. (2005) do for the CNOTs of exact amplitude preparation. In both, the Gray code acts on the circuit rather than on the data address.

Correlation construction versus Gray ordering. As a data address, a Gray code is one fixed member of SAL's family. Under the same loader it stays below SAL on every dataset (table under Q1) and at tensor-train rank 2 (Table 11). It gains little on features without spatial order, while on MNIST the correlation construction learns a two-dimensional Gray code by itself (under Q1), a layout beyond any fixed Gray code of the row-major index.

Q5. Is the learned address transferable across datasets sharing feature semantics, and does it survive being re-learned on a disjoint subset of the same distribution?

Relearning: yes. Transfer: yes when two datasets share their dominant direction, which shared feature semantics alone does not guarantee, and which can be checked without labels.

Relearning on a disjoint subset reproduces the held-out fidelity: 0.793 ± 0.006 vs 0.793 ± 0.005 on MNIST at 20 CX, and 0.591 ± 0.002 for both on CIFAR-10 at 22 CX (Table 27). On MNIST the relearned addresses are again XORs within row bits and within column bits, mostly the Gray code under Q1 (App. G.1).

What decides transfer. Let u1u_1 be the leading eigenvector of the uncentered second moment E[xx⊤]\mathbb E[xx^\top], the amplitude profile the inputs share most strongly. SAL anchors its address on it: the reference order sorts the coordinates by u1u_1, and Eq. (9) bounds every address's average fidelity below by λ1FG⋆(P;u1)\lambda_1F^\star_G(P;u_1). A source address therefore carries over when the target has nearly the same u1u_1, as ∣cos⁡(u1S,u1T)∣|\cos(u_1^S,u_1^T)| measures from the 256 unlabeled vectors SAL already draws (independent samples of one dataset agree to at least 0.94).

Transfer. Each source package (address and topology) was frozen and evaluated on 128 held-out targets of the target dataset (App. G.6, Tables 33–34; std ≤ 0.019). The first three fidelity columns are at G=0G=0, where only the address matters, the last three at GrefG_{\mathrm{ref}}:

Source → target∣cos⁡∣\lvert\cos\rvertnaturalfrozenrelearnednaturalfrozenrelearned
CIFAR-10 → CIFAR-1000.9850.2060.3600.3690.2950.5370.549
CIFAR-100 → CIFAR-100.9850.2270.3880.4080.3110.5740.591
MNIST → KMNIST0.8420.3830.4890.5030.6790.7260.752
SST-2 → AG News0.1950.0320.0340.0850.1020.1050.148
AG News → SST-20.1950.0320.0350.2040.1020.1050.269
  • Aligned u1u_1 (CIFAR-10/100). The frozen address keeps 89–95% of the in-domain gain and raises target-label KRR at 6 CX from 23.10% to 26.64% on CIFAR-100, matching the relearned 26.76% (Table 35).
  • Shared semantics, different u1u_1. SST-2 and AG News share the RoBERTa encoder, yet their u1u_1 are nearly orthogonal, consistent with pretrained language models placing domains in distinct regions (Aharoni & Goldberg, 2020); the frozen address stays at the natural level already at G=0G=0.
  • Low alignment. SAL is relearned from unlabeled target data; 32 construction vectors already recover most of the in-domain gain (Table 21).

We would be glad to discuss any remaining point.

Dear Reviewer tCd3,

We know the last days of the discussion are busy, so just a short note. Your main point was that a learned permutation only matters if it beats a random one. We ran exactly that comparison: random members of the same family fall well below SAL on all three datasets, and on MNIST even below the natural order. Picking the best random one with our own screening still doesn't close the gap (Table 25).

The two papers you pointed us to are now cited in Sec. 2, and the other questions are answered in our earlier replies. If anything is still unclear, we're happy to explain before the deadline.

Reviewer_4W73OpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper learns a shared permutation of input coordinates before quantum amplitude encoding. The permutation preserves the ideal kernel but change state-preparation difficulty. The authors proposes a correlation-based search and proves guarantees for a constructed state family. Experiments covers eight datasets, several loaders, classifier, and quantum hardwares. Results show improved preparation fidelity and classification accuracy. The main contribution is a method for reducing quantum loading costs while retaining the original representation’s ideal pairwise similarity structure.

Critical strengths

The intervention is well motivated. A shared permutation preserves ideal similarity while making approximated loading easier. This gives the experiments a clear reference point. Gains across several dataset and independent loaders is convincing. The work addresses a real quantum bottleneck, and the theoretical construction helps explain why reordering matter.

Critical weaknesses

My main concern is its contribution to machine learning. The paper mainly improves state preparation. The learning experiment show that a better approximation to the input can improve classification. That is useful, but it gives limited new insight into learning, generalization, or model design. The strongest results concern fidelity and gate counts. I see a clear contribution to quantum computation but not to a top ML conference.

The practical ML case is also incomplete. Preserving the ideal amplitude kernel is a clean constraint. But why should an ML practitioner chooses that representation in the first place? The paper does not establish a compelling accuracy–cost trade-off against practical classical approaches. Full cost includes address learning, per-input circuit fitting, loading, and prediction. Without that comparison, the method improving one stage of a pipeline leaves overall value uncertain. The authors should make that broader case, rather than rely mainly on loading improvement.

The quantum preparation result is credible. The case for a substantial ML contribution is not yet strong enough in my feeling.

Questions for authors

My questions:

  • What is the main ML claim beyond making amplitude-encoded states easier to prepare? Is there a finding about learning behavior that would remain useful outside this particular loading setup? Please distinguish the state-preparation contribution from the learning contribution.
  • Can you show one convincing end-to-end accuracy–cost comparison? Including address learning, target-specific fitting, and prediction. Compare against practical classical models on the same fixed input features. This would help establishing if the pipeline is worth using.
  • Is fidelity the right selection objective for prediction? Your results show that approximate kernels can sometimes outperform the ideal kernel. What does that imply about preserving the ideal representation as the central design goal?
6 comments

We thank the reviewer for pressing on the learning side of the paper. Following the suggestion, the introduction now states the two contributions separately.

W1. [...] The learning experiment show that a better approximation to the input can improve classification. That is useful, but it gives limited new insight into learning, generalization, or model design.

Q1. What is the main ML claim beyond making amplitude-encoded states easier to prepare? Is there a finding about learning behavior that would remain useful outside this particular loading setup? Please distinguish the state-preparation contribution from the learning contribution.

Main ML claim. The input address is a learnable, label-free part of the representation with an exact invariance, and it enters two standard classes of quantum learners differently: kernel learners only through loading error, gate-based readouts also through which correlations are local. The revision characterizes both.

  • Kernel learners see only inner products, which the address preserves (Eq. 3). Corollary 3 bounds the change in predictions by the loading error and confines it to small-margin inputs, and budgets up to 8n8n confirm this: the MNIST gain falls from 10.64 points at 4 CX to zero at 80 CX, where both reach 96.56% (ideal 96.72%), while on CIFAR-10, still far from exact loading, it persists (83.84% vs 79.04% at 88 CX).
  • Gate-based readouts act on qubits. Under a binary address, the reduced state of qubit qq is ρq=12(I+cy(vq)X+zy(wq)Z)\rho_q=\tfrac12\big(I+c_y(v_q)X+z_y(w_q)Z\big), and its smaller eigenvalue is exactly the term SAL's structural score minimizes (App. E.3). Purifying qubits for loading therefore places strong shared correlations on single-qubit observables, which a shallow readout accesses most directly. This goes beyond a better approximation of the input: on CIFAR-10 the QNN gain persists with exact inputs (60.14 ± 0.55% vs 54.90 ± 2.08%), where no approximation is left, and at every budget from 6 to 88 CX.

Contributions, separated (Sec. 1).

  • State preparation: the exponential separation between addresses, which no qubit ordering closes (Theorem 1); correlation-based recovery (Theorem 2); the SAL procedure; fidelity and gate savings across four loaders and eight datasets.
  • Learning: a label-free input reparameterization with an exact kernel invariance; its budget-bounded effect on kernel learners (Corollary 3), traced up to near-exact loading; its effect on gate-based readouts, explained by the identity above and present with exact inputs.

Findings that hold outside this loading setup.

  1. Without a loader. The readout effect holds with exact inputs.
  2. A property of the data. The address summarizes second-order structure. On MNIST it recovers, without labels or the pixel grid, a two-dimensional Gray code (Table 26), a locality prior of the kind convolutional architectures build in by design. It reproduces on disjoint samples (Table 27), and its transfer to another dataset follows whether the two share their dominant second-moment direction, measured without labels (Table 34).
  3. Beyond quantum states. The same address raises the classical tensor-train approximability of the data vectors. At TT rank 2, fidelity rises from 0.619 to 0.689 on MNIST, from 0.328 to 0.587 on CIFAR-10 and from 0.062 to 0.265 on SST-2 (Table 2). In TT and quantics representations, the bit ordering of the index governs the required rank (Oseledets, 2011; Khoromskij, 2011); SAL chooses it without labels and beyond axis permutations.

Position. The encoding largely determines a quantum model's expressivity and generalization (Schuld, 2021). SAL acts on the encoding before any circuit, and it improves even AQER (ICLR 2026), whose own bound is set by the target's entanglement, at every budget we tested (Table 10).

W2. [...] Preserving the ideal amplitude kernel is a clean constraint. But why should an ML practitioner chooses that representation in the first place? The paper does not establish a compelling accuracy–cost trade-off against practical classical approaches. Full cost includes address learning, per-input circuit fitting, loading, and prediction.

Q2. Can you show one convincing end-to-end accuracy–cost comparison? Including address learning, target-specific fitting, and prediction. Compare against practical classical models on the same fixed input features.

If we read the question correctly, it asks when a practitioner should choose this classical–quantum pipeline over a purely classical one. That presupposes a quantum part that beats classical learners somewhere, an open question for the field, well beyond the scope of a loading method: provable advantages so far rely on data built around classically hard problems (Liu, Arunachalam & Temme, 2021), and whether advantage is even the right goal for QML is itself debated (Schuld & Killoran, 2022). On fixed explicit features a classical model is the efficient choice: on our four-class hardware features, logistic regression reaches 99.4% and the exact amplitude kernel, itself classical, 98.8%, while at 4 CX SAL lifts simulated KRR from 89.4% to 96.3% (Table 19). We nonetheless adapted parity, a task commonly used to illustrate what quantum classifiers can learn, into an amplitude-encoded version of our own, in which the address decides what a restricted readout learns (below). First we give the end-to-end account within an amplitude-encoded pipeline.

Full cost. We count every stage in units that do not depend on whether circuits are simulated or executed: classical circuit fits and logical CX gates.

  • Address learning, once per dataset and without labels: Walsh statistics on 256 vectors and at most 12×128=1,53612\times128=1{,}536 circuit fits at GrefG_{\mathrm{ref}} (App. B.4), fewer than a quarter of the roughly 7,000 per-input fits that one KRR run needs at the same budget under either address.
  • Per-input fitting, loading and prediction: the same procedure, sweeps, CX budget and model under both addresses; SAL adds one O(d)O(d) index gather per input.

Accuracy against CX (Tables 12, 15, 16; test accuracy %, five runs):

DatasetGG (CX)KRR, naturalKRR, SALQNN, naturalQNN, SAL
MNIST475.82 ± 0.4086.46 ± 0.7664.84 ± 1.7873.18 ± 0.99
2094.04 ± 0.8394.90 ± 0.3974.12 ± 2.8676.08 ± 2.20
4095.48 ± 0.3496.34 ± 0.2176.78 ± 1.7177.18 ± 2.59
8096.56 ± 0.2796.56 ± 0.3677.36 ± 1.4077.32 ± 2.30
exact96.72 ± 0.5496.72 ± 0.5477.50 ± 0.9376.62 ± 1.24
CIFAR-10653.36 ± 0.6663.04 ± 2.0232.92 ± 1.0348.14 ± 2.92
2268.08 ± 0.7972.90 ± 2.0942.74 ± 1.4248.00 ± 4.33
4471.34 ± 2.1779.00 ± 1.1239.04 ± 1.7650.24 ± 2.48
8879.04 ± 0.4483.84 ± 0.6539.78 ± 1.8653.76 ± 1.59
exact89.92 ± 0.4089.92 ± 0.4054.90 ± 2.0860.14 ± 0.55
  • Bounded budgets: learning substitutes for gates. On CIFAR-10 the natural address needs about twice the CX to match SAL's kernel accuracy (51 CX for SAL's 22-CX accuracy, 88 CX for its 44-CX accuracy) and does not reach SAL's 66-CX accuracy within 88 CX. On MNIST the factor is 1.6–2.3 up to 40 CX.
  • Near-exact loading: kernel learners no longer need it. Both MNIST kernels reach 96.56% at 80 CX, as Corollary 3 requires. Screening tells in advance, without labels, whether learning pays: the gain is large only where the natural address is still far from exact loading and the learned address adds fidelity. At 80 CX on MNIST the first fails (natural fidelity 0.93); on AG News, the second.
  • Gate-based readouts: learning can pay at any budget. On CIFAR-10 the QNN gain persists at every budget and with exact inputs, and SAL at 88 CX nearly matches the natural readout with exact inputs (53.76% vs 54.90%).

Q2 (continued): an adapted parity task. Parity is a common probe of what quantum classifiers can learn (Farhi & Neven, 2018; Du et al., 2023); the amplitude-encoded version below is our own construction, in which the input address matters (App. E.5, Table 17).

Setup. Each input is a dense, normalized 256-dimensional vector that factors into eight single-qubit states in a binary coordinate system shared across the dataset and hidden from every learner; the label is the parity of eight latent signs. Every model gets 128 labeled training and 64 validation examples and is tested on 1,024; SAL additionally uses unlabeled vectors. The QNN is the three-layer template of Sec. 4.3 with 24 readout CX. Ten independent layouts:

ModelPreparation CXTest accuracy (%)
SAL + QNN0100.00 ± 0.00
Natural address + same QNN054.80 ± 14.40
865.10 ± 24.16
2489.38 ± 22.01
Natural, exact inputs + same QNN—90.09 ± 20.91
SAL, exact inputs + same QNN—100.00 ± 0.00
MLP, validation-selected—50.65 ± 1.20
Logistic regression—48.74 ± 1.10
Exact amplitude kernel (classical)—49.42 ± 1.27
  • SAL recovers the hidden coordinate system from unlabeled vectors in every layout. Every input then loads as a product state at 0 CX (fidelity 1.000) and carries the label in its single-qubit marginals.
  • Natural address. The same information is spread over many-qubit correlations. With exact inputs it reaches 90.09%, close to its 24-CX accuracy, so most of its small-budget shortfall is a loading effect; the rest is the readout, which fails to train on some layouts under the natural address and on none under SAL.
  • Classical learners on the same vectors and labels stay near chance; the exact kernel cannot be changed by any address.

This is a constructed task, not a claim of advantage. It shows that when the label depends on structure an address can make local, learning the address decides whether a small budget suffices at all, and whether the readout learns reliably even from exact inputs.

Q3. Is fidelity the right selection objective for prediction? Your results show that approximate kernels can sometimes outperform the ideal kernel. What does that imply about preserving the ideal representation as the central design goal?

For kernel learners, fidelity is the matched objective. For gate-based readouts it is a good proxy, but not the whole objective.

  1. Matched to kernel learners. The ideal kernel is address-invariant, so an address changes predictions only through approximation error, which Eq. (4) and Corollary 3 bound by infidelity. Empirically, screening fidelity ranks candidates consistently with validation accuracy (Spearman 0.79 on MNIST, 0.85 on CIFAR-10), and the selected address is on average within 1.1 and 2.3 points of the most accurate candidate in its bank (App. E.1).
  2. Label-free and model-agnostic. One address per dataset, learned once, serves four loaders, kernel learners and QNNs.
  3. Standard in state preparation. MPS, AQCE and AQER all target fidelity, keeping the comparison with that literature clean.

Approximate kernels above the ideal. On Galaxy10-128 at GrefG_{\mathrm{ref}}, SAL (42.66%) > natural (40.96%) > ideal (39.88%). We conjecture a mild regularizing effect: a bounded circuit keeps the dominant, low-entanglement part of each input and drops fine-scale detail, much as a low-rank truncation would. The ideal kernel is thus a reference rather than a ceiling (Sec. 4.3), and preserving it is a safeguard rather than the objective. It guarantees that the address never removes information (instance-wise sorting, which breaks it, loads almost exactly yet drops ideal MNIST accuracy to 34.80%; Tables 23–24), and that the learner converges to the ideal model as loading becomes exact, as MNIST shows at 80 CX.

Gate-based readouts. Here the address also shapes what the readout can access (Q1). A fidelity-selected address already serves the QNN well on CIFAR-10 and on the parity task, but it is not tuned to the readout: the two ideal MNIST QNNs agree within run-to-run variation. A task objective for gate-based readouts is a natural extension, now noted in the conclusion; we keep fidelity because it isolates the effect on loading and keeps one address reusable across models.

We thank the reviewer again; the three questions have sharpened both the learning claim and its scope.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to all reviewers and summary of revisions

We thank the reviewers for careful and constructive reports, which agree on the premise: a shared address "preserves ideal similarity while making approximated loading easier" (4W73), the theory forms "a coherent chain rather than three separate claims" (tCd3), and the work "addresses an important bottleneck in quantum machine learning" (EwBn). The reviews converge on four questions, which the revision answers with new experiments under the original protocol. For ease of review, revisions in the updated PDF are marked in dark orange.

1. Does the gain come from learning the address, or from any reordering? (tCd3) From learning. Uniformly random members of SAL's own family fall far below SAL and stay below it even after the same screening; on MNIST they fall below the natural order, as Theorem 1 anticipates (Table 25). What SAL learns is also visible: from 256 unlabeled MNIST vectors, without being told they are images, it recovers a two-dimensional Gray code over the row and column bits of the pixel index (Table 26).

2. How does it compare with other mapping strategies? (EwBn, tCd3) At equal tensor-train rank, the shared address alone exceeds per-input qubit-order search (Jeon et al.) by 0.12–0.21 in fidelity on pretrained features and is comparable on MNIST pixels, with one index gather per input instead of a search; SAL followed by that search is best everywhere (Table 2). It also exceeds a fixed Gray code (Table 25). Both are members of SAL's family, and Theorem 1 holds under every qubit ordering, even one chosen per input.

3. What does it contribute to learning? (4W73, tCd3) The address enters two classes of learners differently.

  • Kernel learners see only inner products, which the address preserves, so they gain only through reduced loading error, and the gain vanishes as loading becomes exact (new Corollary 3). It closes on MNIST at 80 CX and persists on CIFAR-10 at 88 CX (83.84% vs 79.04%).
  • Gate-based readouts act on qubits, so the address decides which correlations are local. On CIFAR-10 the QNN gain persists with exact inputs (60.14% vs 54.90%). In a parity task we adapted to amplitude encoding, SAL makes the label readable with no preparation CX (100% vs 54.80%; Table 17).

4. Where does it apply? (EwBn, tCd3) A frozen MNIST address keeps its advantage under translation, noise, blur, contrast and rotation (Table 32). Across datasets, transfer follows whether two datasets share the dominant direction u1u_1 of E[xx⊤]\mathbb E[xx^\top], on which SAL anchors its address, and this can be checked from unlabeled data before reuse. CIFAR-10 and CIFAR-100 share it and keep 89–95% of the in-domain gain; SST-2 and AG News do not, and there SAL is relearned from unlabeled samples (Tables 33–34).

Revision overview, by concern.

  • Learning (tCd3): random addresses (Sec. 4.4, App. F.5); learned code, relearning, transfer (Sec. 4.5, App. G).
  • Other mappings (EwBn, tCd3): order search and Gray code (Sec. 4.2, App. D.3); Proposition 4 (App. A.4); related work (Secs. 2.1–2.2).
  • Contribution to learning (4W73, tCd3): Sec. 1; Corollary 3 (App. A.2); readout access, budgets up to 8n8n, parity (Sec. 4.3, App. E.3–E.5); classical references (App. E.6).
  • Shift and presentation (EwBn): App. G.5; Secs. 2.3 and 4.1.

Update (Nov 12, AoE). In response to a follow-up question from EwBn, App. B.4 (new Table 4b) now reports classical runtimes, including the one-time cost of address learning and its break-even point against per-input qubit-order search.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.