{"id":"af167b5e-455b-437c-b59b-18e13205ce0e","arxiv_id":"2507.01902","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Under local-operation circuit cutting, only the LUCJ ansatz among five tested electronic structure methods shows practical sampling overhead.","lead":"Researchers analyzed five quantum chemistry algorithms to see how well they can be split across multiple small quantum processors using circuit cutting. They found that most common algorithms require astronomically many extra runs, but one algorithm, LUCJ, can be distributed efficiently.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single fixed partition drives all overhead numbers; without a partition/ordering search, the infeasibility claim about the ansätze is not established.","rationale":"I identify the same weakest assumption as the reader: the fixed partition. This is the most load-bearing because the paper's headline numbers are the product 9^Ncuts, and Ncuts is entirely a function of where the circuit is cut. The paper's own package-based pipeline (Tequila/ffsim to Qiskit) leaves qubit ordering and partition as free choices; the only stated rationale for the chosen cut is that it is the midpoint. The failure to search partitions means the central claim over-generalizes from an instance. The zero-parameter initialization is a lesser issue: the nonlocal CNOT/CZ gates in UCC-type decompositions are not parameter-controlled, so zero parameters likely do not change Ncuts; that concern is weaker. The favorable LUCJ result is more robust because its CP angles are small and its layout (heavy-hex/all-to-all) is systematically varied. My recommended verdict remains CONDITIONAL: the analysis is a useful first measurement but the broad conclusion needs the partition-search test before it can be accepted as a property of the algorithms.","tokens_in":10090,"tokens_out":10796,"duration_ms":141261,"concrete_test":"For H2 (4 qubits), generate the zero-parameter UCCSD/JW and UpCCD/JW circuits used in the paper, then exhaustively evaluate every midpoint cut after applying all qubit permutations (or every nontrivial bipartition directly), transpiling to the same gate basis and counting crossing two-qubit gates. If any permutation yields Ncuts ≤ 6 for UCCSD/JW, the reported 9^18 overhead is not representative and the infeasibility conclusion for small molecules collapses. For the H2n chains, run a heuristic partition search (e.g., greedy swap or integer programming minimizing crossing two-qubit gates) and recompute the product overhead; if the minimum-overhead curve changes qualitatively for any ansatz, the general conclusion must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is that UCCSD, UpCCD, UpCCGSD, and SPA+GS are not efficiently parallelizable under LO because their sampling overhead scales as 9^Ncuts and Ncuts is large (Section IV). Every reported Ncuts is computed for one fixed cut: for the Tequila ansätze the circuit is 'cut in half, between qubits q1 and q2' with a predetermined orbital ordering, and for LUCJ the partition is the alpha/beta boundary. No search over qubit orderings, orbital orderings, or alternative bipartitions is reported. Since a two-qubit gate only contributes to Ncuts if its qubits lie on opposite sides of the partition, a different partition can change the count dramatically. The overhead is exponential in Ncuts, so e.g. reducing UCCSD/JW H2 from 18 cuts to 6 cuts changes the overhead from 1.5×10^17 to 5.3×10^5. Thus the conclusion that 'these algorithms cannot be efficiently parallelized using LO' is currently supported only for one point in a large configuration space. The paper presents this as a property of the algorithms, but the evidence is specific to a fixed compilation and cut choice.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the feasibility of executing five electronic structure ansätze (UCCSD, UpCCD, UpCCGSD, SPA+GS, and LUCJ) on distributed quantum computers using circuit cutting with local operations (LO). Circuits are generated with Tequila and ffsim, converted to Qiskit, and cut along a single fixed bipartition. The authors count the number of crossing two-qubit gates and compute LO sampling overheads using known decompositions (9^Ncuts for CX/CZ gates, (1+2|sin(theta/2)|)^2 per cut CP gate). They report that UCCSD, UpCCD, UpCCGSD, and SPA+GS have enormous overheads even for H2, while LUCJ has modest overheads that scale gently with system size, and conclude that most of these ansätze cannot be efficiently parallelized with LO and that new distributed electronic structure methods are needed.","tokens_in":10237,"tokens_out":7084,"duration_ms":85235,"significance":"If the reported overheads were representative of the ansätze in general, the paper would provide a useful negative benchmark for distributed quantum chemistry and a concrete motivation for designing new ansätze that are more amenable to circuit cutting. The work is transparent: the overhead arithmetic is straightforward, the external decompositions are referenced, and the authors provide a GitHub repository with the data. The favorable scaling of LUCJ is an interesting and falsifiable finding. However, the strength of the central claim is currently limited by the absence of any exploration over circuit partitions, qubit orderings, and parameter values; the paper mostly demonstrates that one particular compilation of each ansatz is expensive to cut, not that the ansatz itself is inherently incompressible under LO.","major_comments":[{"comment":"Every reported Ncuts is computed for a single fixed bipartition: the Tequila circuits are 'cut in half, between qubits q1 and q2' and the LUCJ circuits are partitioned at the alpha/beta boundary. Since Ncuts is the exponent in the sampling overhead 9^Ncuts, the overhead numbers cited in the paper are not properties of the ansatz alone but of one chosen partition and orbital ordering. A different ordering can move two-qubit gates entirely into one partition, reducing Ncuts dramatically; for example, reducing the H2 UCCSD/JW count from 18 cuts to 6 cuts would lower the overhead from about 1.5e17 to about 5.3e5. Thus the conclusion that these ansätze 'cannot be efficiently parallelized using LO' is currently established only for the specific partition and compilation tested. I request that the authors either perform a scan over partitions and qubit/orbital orderings for at least H2 and H10, or explicitly restrict the conclusion to the tested configuration rather than to the algorithms in general.","section":"Section IV, fixed bipartition"},{"comment":"The paper states that initializing all Tequila parameters to zero does not affect the sampling overhead analysis because the non-local gate cuts do not correspond to parameterized gates. This assumption is not verified. With all parameters set to zero, parameterized rotations become identities and entire excitation operators may simplify away or cancel adjacent two-qubit gates during compilation, which would reduce the number of crossing CX/CZ gates; conversely, nonzero amplitudes could introduce additional two-qubit structure. The authors should compare cut counts for at least one nonzero parameter set (for example MP2 or random amplitudes) and explicitly report the Qiskit optimization level used, so that the reader can see that Ncuts is stable under parameter choices.","section":"Section III, zero-parameter initialization"}],"minor_comments":[{"comment":"The text reports a 'sampling overhead of 10^4 = 6.5610×10^3' for the UpCCD/JW circuit; the correct expression is 9^4 = 6.5610×10^3, and the typo should be fixed.","section":"Section IV, H2 UpCCD/JW overhead"},{"comment":"Reference [25] is misattributed: the UpCCD and UpCCGSD methods are introduced in Y. Lee et al., 'Generalized unitary coupled cluster wave functions for quantum computation,' J. Chem. Theory Comput. 15, 311 (2019), which is listed as reference [39], not in the npj Quantum Information article cited as [25].","section":"References [25] and [39]"},{"comment":"Labeling every sampling overhead above the NumPy double-precision maximum (1.7977e308) as 'infinite' conflates numerical overflow with a mathematical divergence. I suggest reporting log10 overheads or saying 'exceeds the double-precision range' instead, since the qualitative point is already made by the enormous finite values.","section":"Section IV and Fig. 3, 'infinite' overhead"},{"comment":"The figure is difficult to read as printed: the legend and line styles for the different H2n chain lengths are not differentiated in the caption, and it is not stated whether the vertical axes are log-scaled. Please add a clear caption and, if possible, use distinguishable markers or separate panels.","section":"Fig. 3"},{"comment":"The concluding statement that 'new methods must be developed' for distributed electronic structure should consistently be qualified as applying to the LO setting, since Eq. (13) notes that LOCC can reduce sampling overhead and the paper only analyzes LO because of the current Qiskit addon implementation.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The fixed-partition issue is the main scientific risk: the paper's headline claim is currently supported only for one point in a large space of possible partitions and orderings. If the authors add a partition/ordering scan and a parameter-dependence check, the central conclusion will be much better supported. The scope of the study (LO only, linear hydrogen chains, STO-3G) should also be stated more prominently as a limitation rather than as a universal impossibility proof."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: a straightforward, mostly correct overhead analysis of circuit cutting for five electronic structure ansätze. The LUCJ result is a genuinely useful pointer; the negative claims about the other ansätze are plausible but not proven, because every reported cut count comes from one fixed bipartition with no search over qubit orderings or partition choices.\n\nThe paper's arithmetic is right: for CX/CZ, each cut multiplies shots by 9; for CP(θ), by (1+2|sin(θ/2)|)^2. The H2 example is telling—UCCSD/JW with 18 cuts is hopeless at 9^18, while LUCJ with two small-angle CP gates costs about 1.25x. The scaling plots for longer chains show LUCJ staying small while the others blow past double precision. If I were picking an ansatz to run on distributed hardware with a circuit-cutting compiler, I'd take this as a real hint to look at LUCJ.\n\nThe main weakness is the fixed-cut methodology. Every Ncuts is computed for one specific cut (the middle of the qubit register, or the alpha/beta boundary). The number of crossing gates is highly compiler- and ordering-dependent. The paper claims 'these algorithms cannot be efficiently parallelized using LO,' but what it actually shows is that, with the default package mappings and a naive partitioning, the overhead is prohibitive. That's still useful, but it's not an algorithm-level statement. The exponential sensitivity makes this matter: for H2 UCCSD/JW, dropping from 18 cuts to 6 cuts changes the overhead from ~1.5e17 to ~5e5, a completely different regime. A small partition search or at least a discussion of how the cut was chosen would be necessary before I'd trust the stronger conclusion.\n\nSmaller issues: the Tequila parameters are set to zero with a claim that this doesn't affect gate counts—probably true for CX/CZ counts, but it should be verified. Labeling overheads beyond double precision as 'infinite' is fine as a computational cutoff but reads as overstatement. And reference [25] is not the correct source for UpCCD/UpCCGSD; that's a quantum-router paper. The right citation is likely the 2018 Lee et al. JCTC paper, which is actually already [39] in the bibliography.\n\nBottom line: a decent empirical note, worth a referee if the partition sensitivity is addressed or the claims are softened. The LUCJ positive result is solid enough to cite cautiously; the negative results need the partition caveat. I'd send it to review, but the authors should either do a modest partition/ordering search or rewrite the conclusion to match the evidence.","headline":"Useful LUCJ pointer and correct overhead math, but the infeasibility verdict overreaches from a single fixed partition.","tokens_in":10807,"tokens_out":4078,"would_cite":true,"duration_ms":47328,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","81V55"],"pacs":["03.67.Ac","03.67.Lx"],"model":"deepseek-v4-flash","headline":"Most standard electronic-structure ansätze cannot be efficiently parallelized with local-operations circuit cutting, because their shot overhead scales as 9 to the number of cut controlled-X or controlled-Z gates, while LUCJ stays small.","keywords":["electronic structure theory","distributed quantum computing","circuit cutting","local operations","sampling overhead","unitary coupled cluster","LUCJ ansatz","quantum chemistry"],"falsifier":"Search all two-way partitions of the qubit register for UCCSD with the Jordan-Wigner encoding on $\\mathrm{H}_2$ (and on larger $\\mathrm{H}_{2n}$ chains) and recompute the LO sampling overhead from the minimum number of crossing controlled-X and controlled-Z gates; if any partition yields a cut count whose total overhead grows polynomially rather than as $9^{N_{\\mathrm{cuts}}}$, the paper's infeasibility conclusion for that ansatz would be overturned.","tokens_in":9807,"feed_emoji":"⚛️","tokens_out":10085,"duration_ms":101696,"temperature":0.7,"pith_summary":"This paper asks whether five standard quantum-chemistry trial wave functions can be split across multiple small quantum processors by circuit cutting with local operations (LO), the decomposition of two-qubit gates into mixtures of single-qubit operations without classical communication. It reports that four of the five ansätze—UCCSD, UpCCD, UpCCGSD, and SPA+GS—suffer sampling overheads equal to $9$ raised to the number of cut controlled-X or controlled-Z gates, so even the smallest molecule, $\\mathrm{H}_2$ in a minimal basis, demands $9^{18} \\approx 1.5\\times10^{17}$ circuit evaluations for UCCSD with the Jordan-Wigner encoding. The fifth ansatz, the local unitary cluster Jastrow (LUCJ) circuit, cuts small-angle controlled-phase gates and keeps its total overhead at or below $5.5$ for hydrogen chains up to $\\mathrm{H}_{50}$. The authors conclude that most current electronic-structure algorithms cannot be parallelized efficiently with LO circuit cutting and that distributed quantum chemistry needs new ansätze or new cutting protocols.","feed_headline":"Most quantum chemistry circuits can't be split efficiently","feed_subtitle":"Four of five ansätze hit exponential shot overhead under local-operations cutting; LUCJ stays under 6.","key_machinery":"The load-bearing mechanism is LO circuit cutting viewed through quasiprobability simulation: a two-qubit gate $U$ crossing the partition is replaced by a random choice of single-qubit channels $F_i$ with coefficients $a_i$, and the cost is the squared sampling overhead $\\kappa^2 = (\\sum_i |a_i|)^2$, with the total overhead for a circuit equal to the product over all cut gates of $\\gamma_S(U)^2$. For controlled-X and controlled-Z gates this per-gate factor is fixed at $9$, which turns every additional cut into a factor-of-$9$ multiplier; for controlled-phase gates it is $(1+2|\\sin(\\theta/2)|)^2$, which is close to $1$ when the angle is small. The paper's argument consists of counting the controlled-X, controlled-Z, and controlled-phase gates that cross the chosen partition in each ansatz and then evaluating these factors for hydrogen-chain systems.","core_discovery":"On the paper's own terms, the central discovery is a systematic count of what happens when each ansatz is cut in half at a fixed partition under LO. For UCCSD, UpCCD, UpCCGSD, and SPA+GS, every two-qubit gate crossing the cut is a controlled-X or controlled-Z gate, and each such gate contributes a sampling-overhead factor of $\\gamma^2 = 9$; the total overhead is therefore $9^{N_{\\mathrm{cuts}}}$. For $\\mathrm{H}_2$ with UCCSD this is already $1.5\\times10^{17}$ under the Jordan-Wigner encoding and $7.98\\times10^{22}$ under Bravyi-Kitaev, and for every larger hydrogen chain the overhead exceeds the double-precision maximum of $1.7977\\times10^{308}$ and is treated as infinite. LUCJ behaves differently because its cut gates are controlled-phase rotations with small angles $\\theta$, carrying overhead $(1+2|\\sin(\\theta/2)|)^2$ each; the maximum total overhead across $\\mathrm{H}_2$ through $\\mathrm{H}_{50}$ and one to five circuit layers is $5.5$ for all-to-all connectivity and $2.08$ for heavy-hex. The conclusion drawn is that most common ansätze are impractical to distribute with local operations, while LUCJ is a promising exception.","pith_inferences":["The analysis fixes a single partition per circuit; searching over all two-way partitions could reduce the cut count for some ansätze, though it is not obvious it would change the exponential character for CX/CZ-dominated circuits.","Because the overhead is set by the type of cut gate, not by the chemistry, the same verdict would likely carry over to larger molecules and basis sets whenever a middle-register cut crosses many CX/CZ gates.","A testable extension is to allow one-way classical communication ($\\mathrm{LO}{\\rightarrow}\\mathrm{CC}$) or LOCC, whose per-gate overheads are no larger than LO; the paper does not quantify those reductions, so the feasibility boundary for these ansätze under those protocols remains open.","For LUCJ, the small overhead suggests a concrete design rule for future distributed chemistry circuits: keep cut two-qubit gates as small-angle phase rotations and minimize their number; the paper's data support this rule but do not state it as a general principle."],"forward_implications":["For UCCSD, UpCCD, UpCCGSD, and SPA+GS, any LO-distributed implementation would need the uncut circuit's shot count multiplied by $9^{N_{\\mathrm{cuts}}}$, so for $\\mathrm{H}_2$ already the unpartitioned circuit is more efficient.","For hydrogen chains larger than $\\mathrm{H}_2$, the reported overhead for these four ansätze exceeds $1.7977\\times10^{308}$ and is effectively infinite, meaning an LO-distributed run would never finish.","LUCJ remains practical under LO cutting: at most about $5.5$ extra shot overhead for $\\mathrm{H}_{50}$ with all-to-all connectivity and $2.08$ with heavy-hex, across one to five layers.","Fermionic encoding choice (Jordan-Wigner vs Bravyi-Kitaev) changes the number of cuts but does not change the qualitative verdict for the CX/CZ-based ansätze.","New distributed electronic-structure methods, such as distributed unitary selective coupled cluster, are needed if distributed quantum computing is to serve quantum chemistry."],"supporting_citations":[{"why":"Defines circuit knitting via quasiprobability simulation and the LO/LOCC settings that set the sampling-overhead framework.","marker":"[28]"},{"why":"Establishes the per-cut sampling overhead factor $\\gamma^2=9$ for virtual two-qubit gates of the CX/CZ type.","marker":"[43]"},{"why":"Supplies the per-cut overhead formula $(1+2|\\sin(\\theta/2)|)^2$ for controlled-phase gates used in the LUCJ analysis.","marker":"[44]"},{"why":"Introduces the LUCJ ansatz and its all-to-all and heavy-hex circuit layouts that the favorable-scaling result relies on.","marker":"[27]"},{"why":"Provides the Jordan-Wigner mapping whose gate structure determines the crossing CX/CZ cut counts for the circuits studied.","marker":"[41]"},{"why":"Provides the Bravyi-Kitaev encoding, whose different gate count produces the higher overheads reported for those ansätze.","marker":"[42]"},{"why":"Proposes the SPA+GS ansatz, one of the methods whose LO sampling overhead converges to infinity beyond $\\mathrm{H}_2$.","marker":"[26]"},{"why":"Defines the UpCCD and UpCCGSD pair and generalized cluster operators analyzed in the cut-count study.","marker":"[25]"},{"why":"Implements the local-operations circuit-cutting tool that restricts the study to the LO setting.","marker":"[32]"}],"fun_headline_variants":["Four of five chemistry ansätze blow up under circuit cutting","Circuit cutting yields exponential overhead for common ansätze","LUCJ: the one ansatz that survives LO-based distribution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusions rest on cutting each circuit at one fixed partition—the middle of the qubit register, or the alpha/beta boundary for LUCJ—with zero-valued ansatz parameters, so a different partition or parameter-dependent gate layout could change the cut counts and possibly the feasibility verdict.","fun_headline_variants_meta":{"raw":{"variants":["Four of five chemistry ansätze blow up under circuit cutting","Circuit cutting yields exponential overhead for common ansätze","LUCJ: the one ansatz that survives LO-based distribution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1415,"prompt_tokens":963,"completion_tokens":452,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":399}},"tokens_in":579,"tokens_out":452,"duration_ms":5421,"temperature":1.0,"reasoning_tokens":399,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:40:27.600132+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Search all two-way partitions of the qubit register for UCCSD with the Jordan-Wigner encoding on $\\mathrm{H}_2$ (and on larger $\\mathrm{H}_{2n}$ chains) and recompute the LO sampling overhead from the minimum number of crossing controlled-X and controlled-Z gates; if any partition yields a cut count whose total overhead grows polynomially rather than as $9^{N_{\\mathrm{cuts}}}$, the paper's infeasibility conclusion for that ansatz would be overturned.","supporting_citations":[{"cited_title":"Circuit knitting with classical communication,","cited_arxiv_id":null,"evidence_quote":"Defines circuit knitting via quasiprobability simulation and the LO/LOCC settings that set the sampling-overhead framework."},{"cited_title":"Constructing a virtual two-qubit gate by sampling single-qubit operations,","cited_arxiv_id":null,"evidence_quote":"Establishes the per-cut sampling overhead factor $\\gamma^2=9$ for virtual two-qubit gates of the CX/CZ type."},{"cited_title":"Cutting circuits with multiple two-qubit unitaries,","cited_arxiv_id":null,"evidence_quote":"Supplies the per-cut overhead formula $(1+2|\\sin(\\theta/2)|)^2$ for controlled-phase gates used in the LUCJ analysis."},{"cited_title":"Bridging physical intuition and hardware efficiency for correlated electronic states: the local unitary cluster jastrow ansatz for electronic structure,","cited_arxiv_id":null,"evidence_quote":"Introduces the LUCJ ansatz and its all-to-all and heavy-hex circuit layouts that the favorable-scaling result relies on."},{"cited_title":"¨uber das paulische ¨aquivalenzverbot. z phys 47: 631,","cited_arxiv_id":null,"evidence_quote":"Provides the Jordan-Wigner mapping whose gate structure determines the crossing CX/CZ cut counts for the circuits studied."},{"cited_title":"Fermionic quantum computation,","cited_arxiv_id":null,"evidence_quote":"Provides the Bravyi-Kitaev encoding, whose different gate count produces the higher overheads reported for those ansätze."},{"cited_title":"Optimized low-depth quantum circuits for molecular electronic structure using a separable-pair approx- imation,","cited_arxiv_id":null,"evidence_quote":"Proposes the SPA+GS ansatz, one of the methods whose LO sampling overhead converges to infinity beyond $\\mathrm{H}_2$."},{"cited_title":"A quantum router architecture for high-fidelity entanglement flows in quantum networks,","cited_arxiv_id":null,"evidence_quote":"Defines the UpCCD and UpCCGSD pair and generalized cluster operators analyzed in the cut-count study."},{"cited_title":"Qiskit addon: circuit cutting,","cited_arxiv_id":null,"evidence_quote":"Implements the local-operations circuit-cutting tool that restricts the study to the LO setting."}],"review_version":1}