{"id":"19913133-3f0f-42fc-911e-cc0606d4a0d4","arxiv_id":"2412.20893","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An EQ-GAN architecture with random input states learns quantum circuit behavior and is used for equivalence checking and variational circuit optimization.","lead":"This paper presents an architecture that extends the EQ-GAN quantum state learning method to learn the full behavior of a quantum circuit by feeding it random inputs. It is applied to check whether two quantum circuits are equivalent and to recompile circuits, with demonstrations on a quantum processor.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Finite-sample completeness is not established: with the paper's own worst-case D, an n=20 circuit passes 1000 shots with ~99.8% probability, so the claims 'statistically zero' and 'large-scale complete' overstate what Theorem 1 proves.","rationale":"The paper presents a plausible architecture and substantial experimental work: the discriminator training, the 4-qubit hardware demonstrations, and the 4-9 qubit classical benchmarks are real evidence of heuristic usefulness. The issue is not the experiments but the gap between the claims ('complete', 'theoretically proved', 'large-scale') and the proof. Theorem 1's proof is invalid as written, and the worst-case analysis in the same section gives an exponentially small detection probability, so for large n the protocol cannot certify equivalence with few shots. The reader's weakest assumption identified the same finite-sampling issue; this pass also flags that the local product-state circuit does not satisfy the independence assumption invoked in Theorem 1. Because the reader already rendered a CONDITIONAL verdict and the concern reinforces rather than redirects it, my read does not change the verdict. The architecture is worth pursuing, but the completeness claim needs to be restated as a finite-sample statistical bound with explicit worst-case overhead, and a direct comparison to existing variational compiling and local Hilbert-Schmidt tests is needed.","tokens_in":15039,"tokens_out":7897,"duration_ms":83344,"concrete_test":"Run the worst-case benchmark from §3.2: choose reference U=I and generator U'=I^{⊕(n-1)}_2⊕U_s with U_s=R_x(1.23), for n=9 and n=20, using the local random circuit and m=1000 shots per input state, over 100 random states, in the MindSpore Quantum simulator. Compare the empirical miss rate to e^{-m/2^{n-1}} and to the Table-2 detection levels (pfailure_N≈0.15-0.33). If the miss rate is ~2% at n=9 and ~99.8% at n=20, then the finite-sample completeness claim fails while the paper's own limitation formula is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2's completeness claim is the load-bearing pillar: if it fails, RH is only a heuristic variational equivalence test. The proof of Theorem 1 is not a finite-sample proof. It assumes D|φ>=|φ> for arbitrary random states and then infers, from the single equation (6), that each side vanishes because the components are 'independent'. For a sampled state this does not follow; for example, with D=SWAP and |φ>=(|01>+|10>)/√2, D|φ>=|φ> though D≠I. Nor do the §3.3 local product states satisfy the independence premise, since their computational-basis amplitudes are multiplicative products of single-qubit amplitudes. The paper itself supplies the sharp counterexample to the finite-sample version: D=I^{⊕(n-1)}_2⊕U_s yields undetected-error probability ≈e^{-m/2^{n-1}}. With m=1000 shots, n=9 gives ≈2% misses; n=20 gives ≈99.8%. Thus the claim that several random states make non-equivalence 'statistically zero', and the abstract's 'large-scale complete' statement, are unsupported. This does not invalidate the 4-9 qubit demonstrations or the value of the idea as a heuristic, but it is exactly the assumption on which the central claim depends.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the RH architecture, which prepends two identical random circuits to the EQ-GAN framework to extend it from quantum state learning to unitary transformation learning. The authors claim a theoretical completeness proof (Theorem 1), introduce an efficient local random circuit, and demonstrate three applications—equivalence checking of non-parameterized circuits, equivalence checking of parameterized circuits, and variational circuit reconstruction—through classical simulation and experiments on the Zuchongzhi-2 processor, including benchmarks on VeriQBench circuits up to 9 qubits.","tokens_in":15272,"tokens_out":4559,"duration_ms":44140,"significance":"If the completeness claim were rigorously established, the RH architecture would offer a hardware-based alternative to classical simulation for circuit equivalence checking, with potential scalability to large circuits. The paper's empirical demonstrations on 4–9 qubits show clean separation between equivalent and non-equivalent circuits for typical cases, and the open-source code and real-hardware experiments are valuable assets. However, the theoretical completeness as stated is not established: the proof of Theorem 1 contains an invalid step, and the paper's own worst-case analysis shows an exponentially small detection probability for certain errors. The idea is promising as a variational heuristic, but the central 'large-scale complete' claim requires a substantially stronger finite-sample analysis or a deliberate weakening of the claim.","major_comments":[{"comment":"The proof of Theorem 1 is not valid. From Eqs. (5) and (6), the fact that D|φ⟩ = |φ⟩ for a particular |φ⟩ does not imply that each side of Eq. (6) vanishes; the components of a random state are not independent in the sense of arbitrary variation, and the phrase 'mutually independent components' has no rigorous meaning for a fixed, normalized vector. The theorem as stated—if D|φ⟩ = |φ⟩ for every |φ⟩—is true, but the proof does not establish it. More importantly, the application uses only finitely many sampled states, and the paper provides no finite-sample bound relating the observed pfailure to a bound on the difference matrix D. The subsequent sentence that the probability of non-equivalence is 'statistically zero' is therefore unsupported by the presented argument.","section":"Section 3.2, Theorem 1 (Eq. (6))"},{"comment":"The local random circuit generates product states whose 2^n computational-basis amplitudes are functions of only 2n independent random parameters; these amplitudes are not 'mutually independent' as claimed in the text. The set of product states is a measure-zero subset of the Hilbert space, so the premise of Theorem 1—'arbitrary random state'—is not satisfied. Even in an infinite-sample limit, the paper's own worst-case analysis in Section 3.2 with D = I^{⊕(n-1)} ⊕ U_s gives an undetected-error probability of e^{-m/2^{n-1}} for m measurement shots. For n = 20 and m = 1000, this is approximately 99.8%, so the Abstract's claim of a 'large-scale complete' architecture is not supported by the provided analysis.","section":"Section 3.3, Eq. (7)"},{"comment":"The experimental benchmark is encouraging but only demonstrates separation for circuits of 4–9 qubits with errors that are single-qubit rotations (Ry(δ), Rx(1.23)) or simple gate insertions. These errors induce typical fidelity reductions that a few hundred shots can detect; the worst-case errors identified in Section 3.2 are not included in the benchmark. Without a finite-sample completeness bound, the empirical results cannot support the claim that the architecture is a complete verification method at scale. This is not a criticism of the heuristic utility, but the manuscript's central claim must be either proven with rigorous sample complexity or explicitly reframed as a heuristic.","section":"Section 4 and Appendix A.2 (Table 2)"}],"minor_comments":[{"comment":"The text reports initial and final pfailure values (3.148 × 10^-4 decreasing to 2.232 × 10^-17) that are inconsistent with Table 1, which lists 1.291 × 10^-3 and 1.214 × 10^-9 for the same quantities; the authors should reconcile these numbers.","section":"Section 4.1"},{"comment":"The caption contains a typo: 'Refence' should be 'Reference'.","section":"Figure 5 caption"},{"comment":"The term 'SW AP' appears with a spurious space in several places (e.g., 'SW AP test'); it should be 'SWAP test'.","section":"Section 2.2"},{"comment":"The notation bin(i)_k is undefined; it presumably denotes the k-th bit of the binary representation of i, but this should be stated explicitly.","section":"Eq. (7)"},{"comment":"The pfailure_Y values around 10^-31 are floating-point artifacts of exact simulation and should be reported as exactly zero, with a note that the nonzero values arise from numerical precision; otherwise readers may misinterpret machine precision as a physical detection limit.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is worth publishing after a rigorous revision. I recommend that the editor require the authors to either supply a proper finite-sample completeness guarantee (for example, using concentration bounds on the overlap for Haar-random states or for their product-state ensemble) or explicitly weaken the abstract and Section 3.2 to present RH as a heuristic variational equivalence test. The current wording overstates the theoretical guarantees, and the proof of Theorem 1 as written is not recoverable without significant additional analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the RH paper. The core trick is clean: run two circuits on the same random input states and use a trained destructive SWAP test to compare outputs, extending EQ-GAN from state learning to unitary learning by adding a random prepending circuit. That is worth something. The authors demonstrate the idea on 4–9 qubit benchmarks from VeriQBench, with clear separation between equivalent (pfailure ~1e-31) and non-equivalent (pfailure ~0.15–0.33) circuits, and they also run real hardware on Zuchongzhi-2. Code and data are public. That is solid empirical work for the small-scale regime.\n\nThe soft spot is the completeness claim. Theorem 1 assumes D|φ⟩=|φ⟩ for arbitrary random states and then uses \"independence\" to conclude the components vanish term-wise. That step is not justified. A single sampled state can be an eigenstate of a non-identity D—the SWAP example (|01⟩+|10⟩)/√2 is a counterexample. And the local random circuit in §3.3 produces product states whose amplitudes are multiplicative, so they do not satisfy the independence premise either. The authors' own worst-case analysis (D = I⊕(n−1) ⊕ U_s) gives an undetected-error probability e^{-m/2^{n-1}}; with m=1000 shots and n=20, that is ~99.8% misses. So the abstract's \"large-scale complete\" and the text's \"statistically zero\" overstate what is actually proven. What is left is a reasonable heuristic that works well on the tested circuits, which is still useful but not the certified method the paper claims.\n\nI also noticed the related work is thin: the idea of probing a unitary with random input states and comparing outputs is very close to the local Hilbert-Schmidt test and variational quantum compiling approaches (Sharma, Khatri, etc.), and those are not cited. A revision should compare against those and clarify the finite-sample guarantees, or drop the completeness language.\n\nBottom line: the experimental work is genuine, the proof gap is real and load-bearing, and the claims need to be moderated. This deserves a serious referee—it is not a desk reject—but I would expect heavy revision.","headline":"Useful heuristic for circuit equivalence checking on hardware, but the completeness theorem is not proven as stated and the 'large-scale complete' claim needs to be walked back.","tokens_in":15806,"tokens_out":2599,"would_cite":false,"duration_ms":22998,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","68Q12"],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"The paper claims that prepending identical random quantum circuits to an entangling quantum generative adversarial network turns a state-learning method into a unitary-learning architecture, letting quantum hardware verify and reconstruct…","keywords":["quantum circuit redesign","equivalence checking","entangling quantum generative adversarial network","unitary learning","random quantum circuits","variational circuit reconstruction","destructive SWAP test","noisy intermediate-scale quantum"],"falsifier":"Run the RH test on a pair of circuits whose difference matrix is the paper's worst-case form, $D = I_2^{\\oplus(n-1)} \\oplus U_s$ with a nontrivial $U_s$ acting on an exponentially small subspace, and measure the false-negative rate as a function of shot count $m$ and qubit number $n$; if the observed miss probability does not approach $e^{-m/2^{n-1}}$, or if at large $n$ the error is routinely missed with feasible shot counts, the architecture's practical completeness claim fails.","tokens_in":14828,"feed_emoji":"🔀","tokens_out":13962,"duration_ms":135422,"temperature":0.7,"pith_summary":"This paper sets out a way to redesign quantum circuits directly on quantum hardware, without classically simulating the full unitary. The idea is to prepend two identical random quantum circuits to an entangling quantum generative adversarial network, so the network no longer learns just a single output state but the whole unitary transformation implemented by a reference circuit. The paper argues this gives a complete verification procedure: if the trained discriminator says two circuits match for random input states, then their difference matrix $D = U^{\\dagger}U'$ must be the identity. It also introduces a shallow local random circuit that makes the required random inputs practical, and demonstrates the architecture on equivalence checking of parameterized and non-parameterized circuits and on variational circuit reconstruction, both in simulation and on a superconducting processor.","feed_headline":"Random inputs turn a quantum state learner into a circuit checker","feed_subtitle":"Prepending identical random circuits lets one adversarial network verify and rebuild whole unitaries on noisy quantum chips.","key_machinery":"The machinery has four parts. The load-bearing identity is $D = U^{\\dagger}U'$, the difference matrix, together with the fidelity condition $|\\langle\\varphi|D|\\varphi\\rangle|^2 = 1$; Theorem 1 says a unitary satisfying this for arbitrary random states is the identity. The probe states are prepared by two copies of a random circuit placed before the reference and generator circuits, and the paper's practical proposal is a local random circuit made of single-qubit rotations that produces product states with mutually independent amplitudes. The comparator is the parameterized destructive SWAP test, whose failure probability $p_{\\mathrm{failure}} = 1 - |\\langle\\phi|\\psi\\rangle|^2$ is trained to be robust against coherent CZ errors. The paper also explicitly analyzes a worst-case difference matrix $D = I_2^{\\oplus(n-1)} \\oplus U_s$, where an error acting on an exponentially small subspace is missed with probability about $e^{-m/2^{n-1}}$ for $m$ measurements.","core_discovery":"The central claim is that the RH architecture, built from two identical random state-preparation circuits followed by a reference circuit, a generator circuit, and a trained parameterized destructive SWAP-test discriminator, learns and verifies arbitrary unitary transformations rather than merely quantum states. The paper defines the difference matrix $D = U^{\\dagger}U'$ between the reference and generator unitaries and proves that if $D$ fixes every random input state, then $D$ is the identity, so the two circuits are equivalent. This completeness argument turns a collection of state-overlap tests into a proof of circuit identity, and the paper supplements it with the observation that the trained discriminator can compensate for coherent two-qubit gate errors. The architecture is then demonstrated on three tasks: checking equivalence of non-parameterized and parameterized circuits, and variationally reconstructing a target circuit from a structurally different ansatz, with benchmarks reported on a superconducting processor and on circuits of four to nine qubits in classical simulation.","pith_inferences":["Beyond the paper: the theoretical completeness is a continuum statement; any practical run uses finitely many product states and finitely many shots, so the true guarantee is statistical. A natural extension is a full false-negative analysis as a function of the number of input states, the shot count, and the error's subspace dimension.","Beyond the paper: the local random circuit produces only product states, so it may be a weaker probe for errors concentrated on entangled subspaces. One could test whether inserting a few entangling layers into the random module improves worst-case detection at modest depth.","Beyond the paper: because the discriminator is trained on the same hardware that later runs the verification, the method's reliability is coupled to the noise it is meant to detect. An experiment with deliberately injected coherent errors of varying strength would map the smallest detectable error for a given shot budget.","Beyond the paper: collecting output fidelities over a tomographically complete set of inputs could turn the framework from a pass/fail checker into a characterization tool that reconstructs the difference matrix $D$, giving more information than equivalence alone."],"forward_implications":["Equivalence checking of quantum circuits can be moved onto the quantum device: instead of classically simulating a large circuit, one runs the two circuits side by side with identical random inputs and reads off the test-failure probability.","Parameterized circuit equivalence becomes checkable by sampling random parameter values, so compiled variational quantum algorithms can be verified without symbolic reasoning.","The same trained discriminator can be reused across applications, and the generator can be optimized to match a reference circuit, giving a variational circuit-optimization loop that works directly on noisy hardware.","Because the test only misses real differences probabilistically and cannot fabricate differences, a detected mismatch is trustworthy under ideal operations; the residual risk is an undetected error, whose worst-case probability the paper bounds.","The architecture can plug into existing circuit-optimization tools as a hardware substitute for the classical simulation step that currently bottlenecks them."],"supporting_citations":[{"why":"Supplies the entangling quantum generative adversarial network that the RH architecture extends, including its trained discriminator and convergence properties.","marker":"Niu et al. [2022]"},{"why":"Extends EQ-GAN to multiqubit state learning, the starting point that the random-circuit module upgrades to unitary learning.","marker":"Rasmussen and Zinner [2022]"},{"why":"Provides the destructive SWAP test used as the discriminator, whose failure probability equals state infidelity.","marker":"Garcia-Escartin and Chamorro-Posada [2013]"},{"why":"Introduces the SWAP test relation that underlies the comparison between the reference and generator outputs.","marker":"Buhrman et al. [2001]"},{"why":"Justifies checking parameterized circuits by randomly instantiating their parameters and gives the baseline approach for that application.","marker":"Peham et al. [2023]"},{"why":"Provides the VeriQBench dataset of common quantum circuits used for the larger-scale equivalence-checking benchmarks.","marker":"Chen et al. [2022]"},{"why":"Demonstrates calibrating coherent CZ errors with single-qubit Rz rotations, the mechanism the trained discriminator exploits.","marker":"Quantum et al. [2020]"}],"fun_headline_variants":["Randomizing inputs lets one quantum network verify circuits","Quantum circuit checking with a GAN trained on random inputs","From state learning to unitary verification via random prepends","A quantum GAN that checks circuit equivalence after random state prep","Random circuits enable a GAN to learn and verify unitaries"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The completeness guarantee rests on the assumption that the random input states used in the test are rich enough that a unitary fixing every one of them must be the identity; with finitely many states and finitely many shots, a non-equivalent pair can slip through undetected.","fun_headline_variants_meta":{"raw":{"variants":["Randomizing inputs lets one quantum network verify circuits","Quantum circuit checking with a GAN trained on random inputs","From state learning to unitary verification via random prepends","A quantum GAN that checks circuit equivalence after random state prep","Random circuits enable a GAN to learn and verify unitaries"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000553,"raw_usage":{"total_tokens":2605,"prompt_tokens":884,"completion_tokens":1721,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":1641}},"tokens_in":500,"tokens_out":1721,"duration_ms":11435,"temperature":1.0,"reasoning_tokens":1641,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:07:58.837072+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the RH test on a pair of circuits whose difference matrix is the paper's worst-case form, $D = I_2^{\\oplus(n-1)} \\oplus U_s$ with a nontrivial $U_s$ acting on an exponentially small subspace, and measure the false-negative rate as a function of shot count $m$ and qubit number $n$; if the observed miss probability does not approach $e^{-m/2^{n-1}}$, or if at large $n$ the error is routinely missed with feasible shot counts, the architecture's practical completeness claim fails.","supporting_citations":[],"review_version":1}