Pith. sign in

REVIEW 3 major objections 4 minor 13 cited by

This paper claims generative quantum advantage: families of shallow quantum circuits, called instantaneously-deep QNNs, are both efficiently trainable from classical data and classically hard to sample, with experimental support on a 68-qub

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-04 19:47 UTC pith:VFR6WSAD

load-bearing objection Strong theory, solid small-scale experiment, but the headline beyond-classical claim is inferred, not measured, and the abstract overstates it. the 3 major comments →

arxiv 2509.09033 v1 pith:VFR6WSAD submitted 2025-09-10 quant-ph cs.LG

Generative quantum advantage for classical and quantum problems

Hsin-Yuan Huang , Michael Broughton , Norhan Eassa , Hartmut Neven , Ryan Babbush , Jarrod R. McClean This is my paper
classification quant-ph cs.LG MSC 68Q1281P6868T05 PACS 03.67.Lx03.67.Ac
keywords generative quantum advantageinstantaneously-deep QNNsshallow quantum circuitssewing techniquecross-entropy benchmarkingbarren plateauscircuit compressionrandom circuit sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper aims to close the loop on generative quantum advantage: showing that quantum computers can not only sample from classically intractable distributions, but also learn them efficiently from classical data. It introduces instantaneously-deep QNNs, shallow quantum circuits that are provably hard to simulate classically (assuming the non-uniform polynomial hierarchy does not collapse) yet have benign optimization landscapes, free of barren plateaus and proliferating local minima. The central technical move is an exact deep-to-shallow circuit mapping combined with a divide-and-conquer "sewing" technique that learns local inversions and stitches them into a global constant-depth unitary. The paper backs this with experiments on a 68-qubit superconducting processor: learning classically hard bitstring distributions (up to 816 effective shallow qubits, with inferred beyond-classical performance for 34,304 shallow qubits) and learning constant-depth representations of physical simulation circuits on 40 qubits.

Core claim

The central discovery is a family of generative quantum models — instantaneously-deep QNNs — that are universal for classical conditional distributions, efficiently trainable from classical examples, and classically hard to sample from under standard complexity assumptions. The paper proves (Theorem 13) that learning these models is quantumly easy but classically hard: a classical computer can find the parameters from data, but generating new samples requires a quantum computer. The same machinery yields a second result (Theorem 14): finding a constant-depth representation of a circuit promised to have one is classically hard unless BPP=BQP, so circuit compression for fast-forwardable simula

What carries the argument

The key mechanism is the instantaneously-deep QNN: a constant-depth circuit of Hadamard and CZ gates on a D-dimensional lattice whose output distribution exactly equals that of a depth-n deep circuit on a (D-1)-dimensional lattice, via quantum teleportation (measurement-based quantum computation). This exact deep-to-shallow mapping lets a 68-qubit processor sample from distributions that look like much larger shallow circuits. The other load-bearing tool is the sewing technique, which uses ancilla qubits to coherently combine learned local inversions of a shallow unitary into a global constant-depth circuit, provably rendering optimization landscapes benign (strongly convex or with a constan

Load-bearing premise

The load-bearing premise is that the trained shallow model, once learned by emulation, performs on real hardware as well as the original random circuit sampling data, because the exact mapping transfers the XEB fidelity; this transfer from emulation to hardware is not directly measured at the largest sizes.

What would settle it

Run the trained IDQNN at 816 shallow qubits (68 physical qubits) on hardware with no error mitigation and compare the measured XEB on input x=0^n with the XEB of the corresponding deep random circuit sampling data; if the trained circuit's XEB falls substantially below the proxy value, the claim of experimental beyond-classical performance would be falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the central claim holds, there exist distributions that are both efficiently learnable and beyond-classical to sample, so generative quantum advantage is achievable with near-term hardware.
  • The sewing technique implies that shallow quantum circuits can be learned in polynomial time from polynomial data with no barren plateaus, making a broad class of quantum models trainable.
  • The hardness of circuit compression means no efficient classical compiler can generally find constant-depth representations even when promised one, so quantum hardware is necessary for this generative task.
  • The experiments indicate that with about 10^6 samples an IDQNN can be learned to gate-precision below NISQ error, and inference can outperform classical simulation by a large margin, quantifying practical resource crossovers.
  • Both learning and sampling can be performed efficiently in the beyond-classical regime, completing the loop from random circuit sampling to a full generative quantum advantage.

Where Pith is reading between the lines

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

  • The exact deep-to-shallow mapping transfers hardness in both directions: if IDQNN sampling were classically simulable, so would be deep random circuit sampling; conversely, any improvement in deep-circuit sampling algorithms immediately weakens the generative advantage.
  • The resource estimate of 34,304 shallow qubits is an upper bound from one particular mapping; more efficient deep-to-shallow mappings could lower the bar for beyond-classical generative learning.
  • The sewing-based divide-and-conquer may generalize to learning other structured unitaries beyond constant depth, such as circuits with logarithmic depth or commuting gates, yielding a broader class of trainable quantum models.
  • The classical-hardness result for circuit compression could inform verifiable quantum advantage protocols, since it shows that even with a promise, finding a shallow representation is an inherently quantum task.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces instantaneously-deep quantum neural networks (IDQNNs), a family of shallow quantum generative models that are claimed to be efficiently trainable from classical data and classically hard to sample from under the assumption that the non-uniform polynomial hierarchy does not collapse (Theorems 12-13). It also proves that finding constant-depth circuit representations promised to exist is classically hard unless BPP=BQP (Theorem 14), and that a divide-and-conquer 'sewing' technique yields benign optimization landscapes (Theorems 4-5). Experimentally, the authors train IDQNNs on a 68-qubit superconducting processor to learn bitstring distributions and to compress a hidden-structure Hamiltonian evolution, reporting a scaling advantage over classical ML baselines. A further result infers beyond-classical performance at 34,304 effective shallow qubits by mapping circuits from Ref. [30] into IDQNN form and borrowing their reported XEB values.

Significance. The theoretical framework is substantial. If it holds, the paper would close an important loop between efficient learning of shallow quantum circuits and known hardness of sampling, with explicit proofs in Appendices F-H, concrete sample-complexity predictions (about 1e6 samples to reach 1e-3 gate fidelity), and a clean formal statement of the hardness assumption. The circuit-compression experiment on 40 qubits and the landscape numerics are also useful and support the trainability claims. The main weakness is experimental: the direct hardware data stop at 816 shallow qubits, a regime the authors themselves classify as classically simulatable, and the beyond-classical number is inferred from a proxy that is not validated against a noise model.

major comments (3)
  1. [Appendix C.2.c, Fig. 2e] The beyond-classical experimental claim is inferred rather than measured. The manuscript uses the reported XEB values of Ref. [30] as a proxy for the trained IDQNN's hardware performance, justified by an 'exact circuit correspondence.' However, the mapped IDQNN corresponds to 34,304 shallow qubits = 67 x 512 time steps, whereas the Ref. [30] XEB values were obtained at the native depth of those circuits. Transferring XEB requires the mapped deep circuit to have comparable depth, gate count, and error structure; no noise-model calculation or direct measurement of the mapped circuit is provided. Since XEB decays exponentially with accumulated gate error, the expected hardware XEB of a 512-layer mapped circuit is likely far below the quoted values. The caption claim that 'even with hardware noise, quantum models outperform...' is therefore unsupported.
  2. [Section III (Experiments on learning to generate bitstrings)] The only direct hardware data are for up to 816 shallow qubits, and the paper itself states that '816 shallow qubits are not yet expected to be in the beyond-classical regime.' Thus there is no direct experimental demonstration of generative quantum advantage in the beyond-classical regime; the abstract's claim that the 68-qubit processor demonstrates capabilities 'in the beyond-classical regime' relies entirely on the surrogate in Fig. 2e. This wording should be corrected, and the surrogate should either be supported by a noise-model calculation or removed.
  3. [Appendix C.2.c, step 1] The 'exact circuit correspondence' is not demonstrated at the level needed to license the proxy. Mapping arbitrary Ref. [30] random-circuit-sampling circuits into the H/RZ/CZ form of the IDQNN, and through a 1D deep layout, requires refitting and recompilation, which changes the two-qubit gate count, depth, and connectivity. The resulting overhead factor (512 time steps for 67 physical qubits) is quoted without derivation. The paper should provide the explicit compiled circuit, its total gate count and depth, and an estimate of its XEB under the device error rates in Appendix B, or present a direct measurement of the mapped circuit.
minor comments (4)
  1. [Abstract and Fig. 2e caption] The abstract states the processor demonstrates capabilities 'in the beyond-classical regime,' but the direct experiment reaches only 816 shallow qubits, which the text later admits is not beyond classical. Please rephrase to distinguish measured from inferred results.
  2. [Appendix C.2.c] The sentence 'specialized to the case of x=000..., y=000...' is immediately followed by 'in the case of an all 1 input state,' which is inconsistent. Please clarify whether the hard inference input is all-zero or all-one.
  3. [Appendix C.2.c] The formula 'FrontierFLOPS target fidelity = FrontierFLOPS perfect fidelity * F_target' is difficult to read; please format properly and define each quantity.
  4. [Appendix B / Fig. 2] XEB is used both as a fidelity metric and as a training-performance metric. Please define once in the main text how XEB values are computed for the trained models and how they relate to the hardware error rates reported in Appendix B.

Circularity Check

1 steps flagged

Fig. 2e 'beyond-classical' XEB is Ref [30]'s XEB reused as the trained IDQNN's performance; the 34,304-shallow-qubit advantage is a relabeled prior result rather than a measurement of the new model.

specific steps
  1. renaming known result [Appendix C.2.c 'Beyond classical performance', Fig. 2e; main text 'Experiments on learning to generate bitstrings']
    "As we are able to learn from data efficiently in these circuits to beyond the precision of the machine implementing these experiments in addition to an exact circuit correspondence, we use as a proxy for our performance on these circuits their reported XEB values."

    The paper's inferred beyond-classical performance (34,304 shallow qubits) is defined to be the XEB values measured in Ref. [30] on the original RCS circuits. The trained IDQNN is never run on hardware in this regime, and no noise-model calculation for the mapped deep circuit is provided. Thus the headline 'generative quantum advantage' in the beyond-classical regime reduces by construction to the prior RCS XEB result, re-expressed as the learned generative model's performance. The mapping is asserted as exact, but the XEB proxy is an identity substitution, not an independent prediction.

full rationale

Most of the paper is self-contained and not circular. Theorems 12 and 13 are proven in Appendices F and H, including a full proof of the shallow-circuit learning theorem (Theorem 8) and standard complexity reductions (Proposition 4); the citations to Ref. [45] are backed by in-paper proofs. The 816-shallow-qubit hardware experiment is a genuine measurement on 68 physical qubits. However, the paper's marquee beyond-classical experimental evidence (Fig. 2e, 'inferred results beyond 34,000 shallow qubits') is not a measurement of the trained IDQNN: the plotted XEB equals Ref. [30]'s reported XEB, by the paper's own statement. This is a renaming/importation of a prior result rather than a derivation from the new model, and the 'exact circuit correspondence' that would license the proxy is asserted without a hardware check or noise analysis. One significant circular/imported step, partial circularity, warrants a 6 rather than a lower score. The paper's transparency about the proxy is a limitation, but it does not remove the by-construction equivalence.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The central claims rest on standard complexity assumptions (non-uniform PH not collapsing, BPP != BQP), on the promise that a constant-depth representation exists in the compression task, and on the local decoupling property of the input distribution for the learning algorithm. No new physical entities are introduced.

free parameters (2)
  • IDQNN parameters beta_j (one per qubit) = learned from classical samples to precision 1e-3 in experiments
    Trainable parameters of the generative model; the central claim is that they can be learned efficiently from classical data.
  • Input distribution D(x) mixture weights = 1/3 for x=0^n, 1/3 i.i.d. p=0.6, 1/3 i.i.d. p=0.2
    Chosen by hand to satisfy the local decoupling property required for efficient classical training of the IDQNN; affects experimental sample complexity but not the asymptotic claims.
axioms (5)
  • domain assumption Non-uniform polynomial hierarchy does not collapse (Conjecture 1)
    Used in Proposition 4 and Theorems 12,13 to establish classical hardness of sampling from IDQNN outputs.
  • domain assumption BPP != BQP
    Used in Theorem 14 to prove classical hardness of learning compressed simulation circuits.
  • domain assumption Promise that a constant-depth representation exists for the target circuit
    Used in Theorem 14 and the experimental design of the physical simulation task; without this promise the task is ill-posed.
  • domain assumption Local decoupling property of the input distribution D(x)
    Invoked in Appendix C.2.a to ensure the beta parameters can be estimated from local marginals; the experimental D(x) is constructed to satisfy it.
  • standard math Universality of H, RZ, CZ gate set with polynomial overhead (Fact 1)
    Used in Corollary 2 to embed arbitrary circuits into the IDQNN family; standard result from universality.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative quantum advantage for classical and quantum problems." pith.science (2026). https://pith.science/paper/VFR6WSAD

@misc{pith2026250909033,
  author       = {Pith},
  title        = {Pith review of: Generative quantum advantage for classical and quantum problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VFR6WSAD}},
  note         = {Machine review of arXiv:2509.09033}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent breakthroughs in generative machine learning, powered by massive computational resources, have demonstrated unprecedented human-like capabilities. While beyond-classical quantum experiments can generate samples from classically intractable distributions, their complexity has thwarted all efforts toward efficient learning. This challenge has hindered demonstrations of generative quantum advantage: the ability of quantum computers to learn and generate desired outputs substantially better than classical computers. We resolve this challenge by introducing families of generative quantum models that are hard to simulate classically, are efficiently trainable, exhibit no barren plateaus or proliferating local minima, and can learn to generate distributions beyond the reach of classical computers. Using a $68$-qubit superconducting quantum processor, we demonstrate these capabilities in two scenarios: learning classically intractable probability distributions and learning quantum circuits for accelerated physical simulation. Our results establish that both learning and sampling can be performed efficiently in the beyond-classical regime, opening new possibilities for quantum-enhanced generative models with provable advantage.

Figures

Figures reproduced from arXiv: 2509.09033 by Hartmut Neven, Hsin-Yuan Huang, Jarrod R. McClean, Michael Broughton, Norhan Eassa, Ryan Babbush.

Figure 1
Figure 1. Figure 1: Generative quantum advantage overview. (a) Generative AI enabled by quantum technology allows learning and sampling from distributions that are classically hard to sample from. (b) Standard approaches to learning generative quantum AI models suffer from poor training landscapes due to issues such as proliferation of suboptimal local minima and barren plateaus, in contrast to the techniques proposed here th… view at source ↗
Figure 2
Figure 2. Figure 2: Training and inference on classically-hard-to-sample distributions. (a) A round of CZ gates on the device layout used for this experiment. There are 68 physical qubits on the grid, which correspond to up to 816 shallow qubits in the 3D shallow representation. (b) Schematic correspondence from a 2D shallow circuit to a 1D deep circuit, illustrating the mapping used experimentally between 3D shallow circuits… view at source ↗
Figure 3
Figure 3. Figure 3: Learning quantum generative models. (a) The layout of system and ancilla qubits embedded into the superconducting qubit device used for learning with the sewing technique. Arrows indicate the qubits whose results are plotted in panels (c) and (d). (b.1) Decomposition of the QNN into circuit pieces, each containing a block of 4 output qubits. (b.2) Each circuit piece is trained to locally invert the target … view at source ↗
Figure 4
Figure 4. Figure 4: Efficient learning from improved optimization landscapes. (a) The probability of success for learning the SWAP-4 circuit using randomly restarted gradient descent with a local ansatz as a function of the number of qubits. Our divide-and-conquer-based learning algorithm based on learning local circuit pieces avoids the exponential decay in success probability observed with standard approaches. (b) Visualiza… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 13 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Lie Group Diffusion Models for Hardware-Aware Quantum Circuit Synthesis

    quant-ph 2026-06 unverdicted novelty 7.0

    Lie group diffusion models combine a discrete circuit skeleton selector with continuous diffusion on SU(2) ≃ S³ to synthesize hardware-aware quantum circuits, outperforming baselines on three-qubit Hamiltonian simulat...

  2. Quantum Fourier Generative Models Trainable at Large Scale

    quant-ph 2026-06 unverdicted novelty 7.0

    Quantum Fourier generative models are trained classically at over 1000-qubit scale using log-likelihood loss from Parseval's identity and deployed on superconducting hardware for fast sampling that preserves multi-mod...

  3. Gated QKAN-FWP: Scalable Quantum-inspired Sequence Learning

    cs.LG 2026-05 unverdicted novelty 7.0

    Gated QKAN-FWP combines fast weight programming with quantum-inspired Kolmogorov-Arnold networks via single-qubit DARUAN activations and gated updates to deliver a 12.5k-parameter model that outperforms larger classic...

  4. Adversarial Effects on Expressibility and Trainability in Distributed Variational Quantum Algorithms

    quant-ph 2026-05 unverdicted novelty 7.0

    Adversaries perturbing shared entanglement in distributed VQAs can manipulate a new Kraus expressibility metric to keep gradients large but steer training to incorrect solutions.

  5. Minimizing classical resources in variational measurement-based quantum computation for generative modeling

    quant-ph 2026-04 unverdicted novelty 7.0

    A VMBQC model restricted to one extra trainable parameter generates distributions that the corresponding unitary model cannot learn.

  6. Exponential quantum advantage in processing massive classical data

    quant-ph 2026-04 unverdicted novelty 7.0

    A polylog-sized quantum computer achieves exponential advantage over classical machines in classification and dimension reduction of massive classical data using quantum oracle sketching combined with classical shadows.

  7. Toward Generative Quantum Utility via Correlation-Complexity Map

    cs.LG 2026-03 unverdicted novelty 7.0

    A pre-training diagnostic map based on spectral correlation resemblance to IQP circuits and excess structural complexity identifies suitable datasets like turbulence data for quantum generative models, yielding compet...

  8. Quantum computation at the edge of chaos

    quant-ph 2026-04 unverdicted novelty 6.0

    Topological entanglement entropy regularizes variational quantum algorithms to enforce quantum sparsity and operate at the edge of chaos for better trainability.

  9. Scaling Quantum Machine Learning without Tricks: Full-Resolution and Diverse Image Generation

    quant-ph 2026-02 conditional novelty 6.0

    A single end-to-end quantum generator using an image-tailored circuit and learnable multimodal noise achieves state-of-the-art simulated FID scores on full MNIST and Fashion-MNIST without tricks.

  10. Software Between Quantum and Machine Learning -- And Down to Pulses

    quant-ph 2026-05 unverdicted novelty 5.0

    A JAX-based framework extending quantum machine learning to pulse-level control with composable ansatzes, end-to-end optimization, and Fourier diagnostics.

  11. Spectral methods: crucial for machine learning, natural for quantum computers?

    quant-ph 2026-03 unverdicted novelty 5.0

    Quantum computers may enable more natural manipulation of Fourier spectra in ML models via the Quantum Fourier Transform, potentially leading to resource-efficient spectral methods.

  12. Generative IQP Circuit Learning with Physics-Informed Latent Initialization

    quant-ph 2026-07 conditional novelty 4.0

    Initializing the latent block of an IQP generative circuit with a PINN-learned latent vector lowers reconstruction error at unseen viscosities versus random initialization on Burgers' equation.

  13. Software Between Quantum and Machine Learning -- And Down to Pulses

    quant-ph 2026-05 unverdicted novelty 4.0

    Introduces a JAX-based framework for pulse-level QML with composable ansatze, end-to-end pulse optimization, and Fourier-analytic diagnostics.

Reference graph

Works this paper leans on

46 extracted references · cited by 12 Pith papers

  1. [1]

    Sample an input product state|𝜓ℓ⟩= ⨂︀𝑛 𝑖=1|𝜓ℓ,𝑖⟩, with each|𝜓 ℓ,𝑖⟩being a uniformly random single-qubit stabilizer state from{|0⟩,|1⟩,|+⟩,|−⟩,|𝑦+⟩,|𝑦−⟩}

  2. [2]

    Apply the unknown unitary𝑈to|𝜓 ℓ⟩

  3. [3]

    The measurement collapses the state to a product state|𝜑ℓ⟩= ⨂︀𝑛 𝑖=1|𝜑ℓ,𝑖⟩

    Measure each qubit of𝑈|𝜓ℓ⟩in a random Pauli basis chosen from{𝑋, 𝑌, 𝑍}. The measurement collapses the state to a product state|𝜑ℓ⟩= ⨂︀𝑛 𝑖=1|𝜑ℓ,𝑖⟩. This dataset can be represented efficiently on a classical computer using𝒪(𝑁 𝑛)bits. F.2. Algorithm 1: Known structure with direct Heisenberg sewing We begin by presenting an algorithm for learning a shallow qu...

  4. [4]

    Use Lemma 1 to learn the Heisenberg-evolved Pauli observables^𝑂𝑖,𝑃≈𝑈†𝑃𝑖𝑈

  5. [5]

    For each qubit𝑖, optimize a local inversion𝑉𝑖 such that𝑉† 𝑖 ^𝑂𝑖,𝑃 𝑉𝑖≈𝑃 𝑖 as in Algorithm 2

  6. [6]

    Apply local inversion sewing as in Algorithm 2 to obtain the learned channel^ℰ(𝜌). In the following, we state the main lemma for learning the Heisenberg-evolved observables from the dataset𝒯𝑈(𝑁) with an arbitrary circuit architecture for the unknown shallow quantum neural network. Lemma 1(Learning Heisenberg-evolved Pauli observables).Given an unknown𝑛-qu...

  7. [7]

    For all𝑄∈ {𝐼, 𝑋, 𝑌, 𝑍}⊗𝑛 with|𝑄| ≤𝑘, i.e.,𝑄acts as non-identity on at most𝑘qubits, compute the coefficient: ^𝛼𝑄 = 3|𝑄| 𝑁 𝑁∑︁ ℓ=1 3⟨𝜑 ℓ,𝑖|𝑃|𝜑 ℓ,𝑖⟩·⟨𝜓 ℓ|𝑄|𝜓 ℓ⟩,(S.4.18) then apply a threshold function to the coefficient: ^𝛽𝑄 = {︃ ^𝛼𝑄,|^𝛼 𝑄|≥0.5𝜀/(2 √ 2)𝑘 0,|^𝛼 𝑄|<0.5𝜀/(2 √ 2)𝑘 .(S.4.19)

  8. [8]

    barren plateau problem

    Construct the approximation as: ^𝑂𝑖,𝑃 = ∑︁ 𝑄:|𝑄|≤𝑘 ^𝛽𝑄𝑄.(S.4.20) This algorithm guarantees that with probability at least1−𝛿: ‖ ^𝑂𝑖,𝑃−𝑈†𝑃𝑖𝑈‖∞≤𝜀andsupp( ^𝑂𝑖,𝑃 )⊆supp(𝑈 †𝑃𝑖𝑈)(S.4.21) for all𝑖∈{1, . . . , 𝑛}and𝑃∈{𝑋, 𝑌, 𝑍}, wheresupp(𝐴)is the set of qubits that𝐴acts nontrivially on. F.5. Main learning theorem Theorem 8(Learning shallow quantum circuits).Given...

  9. [9]

    Every local minimum is global

  10. [10]

    Proof.These properties follow directly from strong convexity: strictly convex functions have a unique global mini- mizer, and any stationary point must be this minimum

    Every critical point satisfies⃗ 𝛼=⃗ 𝛼*. Proof.These properties follow directly from strong convexity: strictly convex functions have a unique global mini- mizer, and any stationary point must be this minimum. G.2.b. Constant local minima in learning local inversions We now analyze the landscape for the non-convex optimization problem underlying Algorithms...

  11. [11]

    Encodes the input bitstring𝑥∈{0,1} 𝑛 into anℓ-qubit input product state|𝜓 𝑥⟩and a local measurement basisℳ 𝑥 over𝑚≤ℓqubits

  12. [12]

    Evolves the input product state|𝜓𝑥⟩under anℓ-qubit parameterized quantum circuit𝐶( ⃗𝜃). 41

  13. [13]

    The QNN generates an𝑚-bit string𝑦∈{0,1} 𝑚 according to a distribution𝑝QNN(𝑦|𝑥; ⃗𝜃)that depends on the input bitstring𝑥∈{0,1} 𝑛 and the trainable circuit parameters⃗𝜃

    Measures the first𝑚qubits of theℓ-qubit output state𝐶( ⃗𝜃)|𝜓 𝑥⟩in the basisℳ 𝑥. The QNN generates an𝑚-bit string𝑦∈{0,1} 𝑚 according to a distribution𝑝QNN(𝑦|𝑥; ⃗𝜃)that depends on the input bitstring𝑥∈{0,1} 𝑛 and the trainable circuit parameters⃗𝜃. Even when the measurement is generalized to a POVM conditioned on𝑥,ℳ𝑥 forms a probability distribution for𝑝(𝑦|...

  14. [14]

    Preparing the𝑛-qubit product input state|𝜓𝑥⟩according to bitstring𝑥

  15. [15]

    Applying a layer of𝑅𝑍(𝜃𝑖) =𝑒−𝑖𝜃𝑖𝑍/2 rotation gates to all𝑛qubits, where{𝜃𝑖}𝑛 𝑖=1 are𝑛trainable parameters

  16. [16]

    Applying a layer of𝐶𝑍gates on all edges in the connectivity graph𝐺

  17. [17]

    This quantum circuit is shallow with𝒪(1)circuit depth regardless of the number of qubits

    Measuring the𝑛-qubit final state in the Pauli-𝑋basis. This quantum circuit is shallow with𝒪(1)circuit depth regardless of the number of qubits. The following lemma shows that despite their constant depth, these shallow QNNs can effectively implement deep random quantum circuits when arranged with an appropriate connectivity structure. This lemma uses exac...

  18. [18]

    ·𝑛𝐷)-qubit state|𝜑⟩to be the product state given by𝑥restricted to the coordinates (1, 𝑖2,

    Initialize an(𝑛2·𝑛 3·. . .·𝑛𝐷)-qubit state|𝜑⟩to be the product state given by𝑥restricted to the coordinates (1, 𝑖2, . . . , 𝑖𝐷), where the mapping is0↦→|+⟩and1↦→|0⟩for each qubit

  19. [19]

    This parameter indexes both the first coordinate of the shallow quantum circuit and the circuit layer of the corresponding deep quantum circuit

    Set𝑖 1←1. This parameter indexes both the first coordinate of the shallow quantum circuit and the circuit layer of the corresponding deep quantum circuit

  20. [20]

    , 𝑖𝐷)with rotation angles𝜃(𝑖1,𝑖2,...,𝑖𝐷) determined by the QNN, for all𝑖2,

    Apply a layer of𝑅𝑍 rotations to qubits at positions(𝑖1, 𝑖2, . . . , 𝑖𝐷)with rotation angles𝜃(𝑖1,𝑖2,...,𝑖𝐷) determined by the QNN, for all𝑖2, . . . , 𝑖𝐷

  21. [21]

    , 𝑖𝐷)and(𝑖 1, 𝑖′ 2,

    Apply a layer of𝐶𝑍gates to|𝜓⟩on edges between(𝑖 1, 𝑖2, . . . , 𝑖𝐷)and(𝑖 1, 𝑖′ 2, . . . , 𝑖′ 𝐷)in𝐺for all coordinates 1≤𝑖 2, 𝑖′ 2≤𝑛 2, . . . ,1≤𝑖𝐷, 𝑖′ 𝐷≤𝑛 𝐷

  22. [22]

    , 𝑖𝐷)for all𝑖 2,

    If𝑖 1 < 𝑛1, generate the bits for𝑦𝑖 with𝑖= (𝑖 1, 𝑖2, . . . , 𝑖𝐷)for all𝑖 2, . . . , 𝑖𝐷 through the following steps: •If𝑥 (𝑖1+1,𝑖2,...,𝑖𝐷) = 0, apply an𝐻gate irrespective of𝑦 𝑖, set𝑦 𝑖 to be a random bit, and apply a𝑍gate to the(𝑖 2, . . . , 𝑖𝐷)-th qubit of|𝜑⟩if𝑦 𝑖 = 1(do nothing if𝑦 𝑖 = 0). •If𝑥 (𝑖1+1,𝑖2,...,𝑖𝐷) = 1, measure the(𝑖2, . . . , 𝑖𝐷)-th qubit o...

  23. [23]

    , 𝑖𝐷)for all𝑖 2,

    If𝑖 1 =𝑛 1, generate the random bits for𝑦 𝑖 with𝑖= (𝑛 1, 𝑖2, . . . , 𝑖𝐷)for all𝑖 2, . . . , 𝑖𝐷 by measuring all 𝑛2·𝑛 3·. . .·𝑛𝐷 qubits in the state|𝜑⟩in the𝑋basis. Proof.We show that the shallow QNN simulates a depth-𝑛1 circuit on a(𝐷−1)-dimensional lattice of size𝑛 2× ···×𝑛 𝐷. The first dimension coordinate𝑖 1 in our shallow QNN effectively encodes the c...

  24. [24]

    , 𝑖𝐷)is a component of|𝜑 (𝑖1)⟩, |𝜑(𝑖1)⟩=|0⟩⊗| ̃︀𝜑(𝑖1) 0 ⟩+|1⟩⊗| ̃︀𝜑(𝑖1) 1 ⟩,(S.2.2) for some unnormalized states|̃︀𝜑(𝑖1) 0 ⟩,| ̃︀𝜑(𝑖1) 1 ⟩

    The qubit at position(𝑖1, 𝑖2, . . . , 𝑖𝐷)is a component of|𝜑 (𝑖1)⟩, |𝜑(𝑖1)⟩=|0⟩⊗| ̃︀𝜑(𝑖1) 0 ⟩+|1⟩⊗| ̃︀𝜑(𝑖1) 1 ⟩,(S.2.2) for some unnormalized states|̃︀𝜑(𝑖1) 0 ⟩,| ̃︀𝜑(𝑖1) 1 ⟩. 44

  25. [25]

    , 𝑖𝐷)is initialized to|+⟩= 1√ 2(|0⟩+|1⟩)as specified by the condition on the input bit𝑥(𝑖1+1,𝑖2,...,𝑖𝐷) = 0

    The qubit at position(𝑖 1 + 1, 𝑖2, . . . , 𝑖𝐷)is initialized to|+⟩= 1√ 2(|0⟩+|1⟩)as specified by the condition on the input bit𝑥(𝑖1+1,𝑖2,...,𝑖𝐷) = 0

  26. [26]

    The𝐶𝑍gate between these qubits creates the entangled state: |0⟩⊗ |0⟩+|1⟩√ 2 ⊗| ̃︀𝜑(𝑖1) 0 ⟩+|1⟩⊗ |0⟩−|1⟩√ 2 ⊗| ̃︀𝜑(𝑖1) 1 ⟩(S.2.3) =|+⟩+|−⟩√ 2 ⊗|0⟩+|1⟩√ 2 ⊗| ̃︀𝜑(𝑖1) 0 ⟩+ |+⟩−|−⟩√ 2 ⊗|0⟩−|1⟩√ 2 ⊗| ̃︀𝜑(𝑖1) 1 ⟩(S.2.4) =|+⟩ 2 ⊗ (︁ |0⟩⊗ (︁ | ̃︀𝜑(𝑖1) 0 ⟩+| ̃︀𝜑(𝑖1) 1 ⟩ )︁ +|1⟩⊗ (︁ | ̃︀𝜑(𝑖1) 0 ⟩−| ̃︀𝜑(𝑖1) 1 ⟩ )︁)︁ (S.2.5) +|−⟩ 2 ⊗ (︁ |0⟩⊗ (︁ | ̃︀𝜑(𝑖1) 0 ⟩−| ̃︀𝜑(𝑖1...

  27. [27]

    Measuring the first qubit in the𝑋basis yields a uniformly random outcome𝑦(𝑖1,𝑖2,...,𝑖𝐷) and collapses the state to one of two cases: ∗If𝑦 (𝑖1,𝑖2,...,𝑖𝐷) = 0: the remaining state collapses to |+⟩⊗| ̃︀𝜑(𝑖1) 0 ⟩+|−⟩⊗| ̃︀𝜑(𝑖1) 1 ⟩.(S.2.8) ∗If𝑦 (𝑖1,𝑖2,...,𝑖𝐷) = 1: the remaining state collapses to |+⟩⊗| ̃︀𝜑(𝑖1) 0 ⟩−|−⟩⊗| ̃︀𝜑(𝑖1) 1 ⟩.(S.2.9) The first qubit meas...

  28. [28]

    Applying the𝐻gate irrespective of𝑦 (𝑖1,𝑖2,...,𝑖𝐷) turns the state into |0⟩⊗| ̃︀𝜑(𝑖1) 0 ⟩+|1⟩⊗| ̃︀𝜑(𝑖1) 1 ⟩if𝑦 (𝑖1,𝑖2,...,𝑖𝐷) = 0,(S.2.10) |0⟩⊗| ̃︀𝜑(𝑖1) 0 ⟩−|1⟩⊗| ̃︀𝜑(𝑖1) 1 ⟩if𝑦 (𝑖1,𝑖2,...,𝑖𝐷) = 1.(S.2.11)

  29. [29]

    , 𝑖𝐷)-th qubit of|𝜑⟩if𝑦 (𝑖1,𝑖2,...,𝑖𝐷) = 1maps the state to|0⟩⊗ | ̃︀𝜑(𝑖1) 0 ⟩+|1⟩⊗| ̃︀𝜑(𝑖1) 1 ⟩=|𝜑 (𝑖1)⟩in both cases of𝑦 (𝑖1,𝑖2,...,𝑖𝐷)

    Applying a𝑍gate to the(𝑖 2, . . . , 𝑖𝐷)-th qubit of|𝜑⟩if𝑦 (𝑖1,𝑖2,...,𝑖𝐷) = 1maps the state to|0⟩⊗ | ̃︀𝜑(𝑖1) 0 ⟩+|1⟩⊗| ̃︀𝜑(𝑖1) 1 ⟩=|𝜑 (𝑖1)⟩in both cases of𝑦 (𝑖1,𝑖2,...,𝑖𝐷). In Case 1, we have teleported the qubit at position(𝑖 1, 𝑖2, . . . , 𝑖𝐷)to the qubit at position(𝑖 1 + 1, 𝑖2, . . . , 𝑖𝐷)while maintaining the state|𝜑 (𝑖1)⟩unchanged. –Case 2 (𝑥 (𝑖1+1,𝑖...

  30. [30]

    , 𝑖𝐷)is initialized to|0⟩

    The qubit at position(𝑖1 + 1, 𝑖2, . . . , 𝑖𝐷)is initialized to|0⟩

  31. [31]

    The𝐶𝑍gatebetweenthecurrentqubitand|0⟩hasnoeffectsince𝐶𝑍|𝑏⟩|0⟩=|𝑏⟩|0⟩forany𝑏∈{0,1}

  32. [32]

    , 𝑖𝐷)in the𝑋basis is equivalent to measuring the qubit at position(𝑖 2,

    Measuring the qubit at position(𝑖1, 𝑖2, . . . , 𝑖𝐷)in the𝑋basis is equivalent to measuring the qubit at position(𝑖 2, . . . , 𝑖𝐷)in the state|𝜑 (𝑖1)⟩in the𝑋basis. This generates a measurement outcome 𝑦(𝑖1,𝑖2,...,𝑖𝐷)∈{0,1}and collapses the state|𝜑 (𝑖1)⟩according to the measurement outcome

  33. [33]

    , 𝑖𝐷)in the state|𝜑 (𝑖1)⟩to|0⟩

    We reset the qubit at position(𝑖2, . . . , 𝑖𝐷)in the state|𝜑 (𝑖1)⟩to|0⟩. In Case 2, the qubit at position(𝑖1, 𝑖2, . . . , 𝑖𝐷)is not teleported, but the qubit at position(𝑖2, . . . , 𝑖𝐷)in the state|𝜑 (𝑖1)⟩is measured and reset to|0⟩. •State Evolution: After processing all qubits at layer𝑖1, the state|𝜑(𝑖1+1)⟩correctly represents the state after 𝑖1 + 1laye...

  34. [34]

    Qubits that received teleported states (at positions where𝑥(𝑖1+1,𝑖2,...,𝑖𝐷) = 0) carry forward the quantum information from the previous layer

  35. [35]

    Qubits initialized to|0⟩(at positions where𝑥 (𝑖1+1,𝑖2,...,𝑖𝐷) = 1) match the mid-circuit measurement and qubit reset in the deep QNN

  36. [36]

    45 By induction, after all𝑛1 layers, the final state|𝜑(𝑛1)⟩is the same as the state after all𝑛1 layers of the deep QNN

    The pattern of entanglement created by the𝐶𝑍gates within each slice matches the entangling operations in the corresponding layer of the deep QNN. 45 By induction, after all𝑛1 layers, the final state|𝜑(𝑛1)⟩is the same as the state after all𝑛1 layers of the deep QNN. Final Measurement At the final slice (𝑖1 =𝑛 1), all qubits are measured in the𝑋basis. The m...

  37. [37]

    a layer of Hadamard gates on all𝑚qubits

  38. [38]

    a layer of𝑅𝑍(𝜃)gates on all qubits with independent𝜃on each of the𝑚qubits

  39. [39]

    We now define the family of deep quantum circuits, which corresponds to injecting different patterns of single- qubit𝑍gates in each round of the circuit layers

    a layer of𝐶𝑍gates on some choices of nearest neighbors on the 1D line, such that the unitaries implemented by𝐶and𝐷differ by at most𝜖error in diamond distance. We now define the family of deep quantum circuits, which corresponds to injecting different patterns of single- qubit𝑍gates in each round of the circuit layers. Definition 17(A family of deep quantu...

  40. [40]

    Prepare a single-qubit|+⟩state,

  41. [41]

    Apply an𝑅 𝑍(𝜃𝑖)rotation with an unknown𝜃 𝑖,

  42. [42]

    Measure the single-qubit state in the𝑋basis to obtain𝑦𝑖. In this single-qubit experiment, we have: Pr[𝑦𝑖 = 1] = 1−cos(2𝜃 𝑖) 2 .(S.4.13) Hence, we can learn𝜃𝑖 up to a small error using the estimator: ^𝜃𝑖 = 1 2 arccos ⎛ ⎝1−2 1 𝑁sp 𝑁sp∑︁ 𝑡=1 𝑦(𝑡) 𝑖 ⎞ ⎠ .(S.4.14) Because the specified condition that all adjacent𝑥𝑖′ are1while the target𝑥 𝑖 is0is satisfied with...

  43. [43]

    Prepare the𝑛-qubit product input state according to bitstring𝑥

  44. [44]

    Apply a single layer of𝑅𝑍(𝜃𝑖)rotation gates to all𝑛qubits, where{𝜃 𝑖}𝑛 𝑖=1 are the𝑛trainable parameters, and𝑅 𝑍(𝜃𝑖) =𝑒−𝑖𝜃𝑖𝑍/2

  45. [45]

    Apply a layer of𝐶𝑍gates on all edges in the connectivity graph𝐺

  46. [46]

    This quantum circuit is shallow with𝒪(1)circuit depth regardless of the number of qubits

    Measure the𝑛-qubit final state in the Pauli basis according to bitstring𝑥. This quantum circuit is shallow with𝒪(1)circuit depth regardless of the number of qubits. Suppose that it is classically easy to learn generative models created by any polynomial-size tomographically- complete shallow QNNs. Because a polynomial-size tomographically-complete instant...

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.