{"id":"fea4df6e-93d5-471b-a2dc-1c6c3917285b","arxiv_id":"2610.38193","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Independent zero-mean branches in a single-pass LCU cannot erase a branch's gradient variance below a coefficient-scaled floor, while interleaved circuits concentrate only under a joint random-sampling assumption.","lead":"Quantum machine learning circuits can be combined coherently, and this paper determines when that combination preserves or kills the gradient signal needed for training. The results give exact variance formulas and a control-algebra criterion that separates harmless coefficient tuning from genuine exponential concentration.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Proposition 4.1's gradient floor is correctly proved, but only for exactly zero branch means and exact independence; the paper's approximate-moment machinery does not extend it, so near-identity initializations are left outside the guarantee.","rationale":"I treated Proposition 4.1 as the load-bearing claim and checked the proof line by line: Eq. (25) is the exact decomposition, conditioning on (U_l, U̇_l) makes each R_k zero-mean and mutually independent, and the covariance with g_branch vanishes, giving Var(∂θL) = |α_l|⁴Var(g_branch) + Σ E[R_k²]. Nothing in the argument uses Haar measure, a particular gate position, or a special form for U_l, so the proposition and its 'any gate position' remark hold as stated. I also checked claims the headline rests on: the L = 2 case of Theorem 4.10 reproduces the Dirichlet moments (E[c⁴] = 1/5, E[c₁²c₂²] = 1/30) and reduces to Var[m] at L = 1; the 3V and 7V thresholds in §8.2 follow from Eq. (124); r* ≈ 7.41 and w* ≈ 0.5437 in Theorem I.1(iv) satisfy the stated stationarity relation; and the matchgate gradient benchmark 4/(2n−1) agrees with the independent adjoint-orbit computation E[(RΓRᵀ)²_ab] = ‖Γ‖²_F/(2n(2n−1)) = 1/(2n−1). The reader's weakest_assumption names the right premise, so I agree on the locus of concern. My difference is one of emphasis: the issue is not that the premise is unstated (it is stated in §3.6 and b232) but that no stability version exists, and the paper's approximate-moment results do not supply one, because their error scale is not tied to the branch variance. That limits how far the abstract can be pushed, but it does not impugn the theorem, and the paper's scoping statements are accurate, so the verdict should not change. One secondary internal check I could not settle by hand: whether the sector formula (28), applied with a single so(2n) sector, reproduces the 1/(2n−1) branch variance used in Figure 2(a) for n ≥ 3. At n = 4 my hand evaluation gives (1 − 2/16)·16/28 = 1/2 versus the figure's 1/7, which suggests (28) is being used where the two half-spinor blocks require blockwise projections; this would affect the 'resolve by (28)' remark in Theorem 7.1(i), not the floor itself.","tokens_in":55782,"tokens_out":62462,"duration_ms":513529,"concrete_test":"Take the paper's two-branch measurement-adjacent benchmark (Eq. 57, Figure 2(b)) and dilute the undifferentiated branch to the mixture E_p = (1−p)δ_{U=I} + p·Haar(e^{so(2n)}), so E[U₂] = (1−p)I, keeping U₁ = e^{iθX₁X₂}V₁ with V₁ Haar-random. Matchgate circuits are classically simulable in poly(n), so Var(∂θL)/λ⁴ can be estimated to high precision at n = 4, 6, 8 for p = 0.9, 0.99, 0.999 and compared with |α₁|⁴Var(g_branch) = w₁⁴·4/(2n−1). If the ratio falls below 1 already at p = 0.999, the zero-mean hypothesis is doing decisive work rather than being a technical convenience, and the abstract should carry that qualifier.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The floor's proof (Eqs. 25–26) is a conditional-expectation argument: holding (U_l, U̇_l) fixed, each interference term R_k = 2Re(α_l ᾱ_k Tr[O U̇_l ρ U_k†]) has zero conditional mean because E[U_k†] = 0, so Cov(g_branch, Σ_k R_k) = 0 and only nonnegative squares survive. That algebra is exact and correct. The load-bearing point is that the cancellation is an exact-first-moment statement. If E[U_k] = δ ≠ 0, each R_k acquires the deterministic piece 2Re(α_l ᾱ_k Tr[O U̇_l ρ δ†]), which depends on the same random U̇_l that determines g_branch; the covariance is then generically nonzero and may be negative. Writing v = Var(g_branch) and ρ = Var(Σ_k R_k), one only has Var(∂θL) ≥ |α_l|⁴v + ρ − 2|α_l|²√(ρv), which vanishes at ρ = |α_l|⁴v, so no positive floor survives in the worst case. Section 3.6 acknowledges that shared-parameter and near-identity ensembles 'require checking these hypotheses separately' (b232), so this is not concealed. What is missing is a stability statement: Prop. H.2 bounds |Var_E[L] − Var_Haar[L]| ≤ 8λ⁴‖O‖² Σ_l w_l ε_l, an error measured at the scale λ⁴‖O‖² rather than at the scale of the branch variance, and it concerns the loss, not the floor. In the relevant regime the moment error needed to certify |α_l|⁴Var(g_branch) is therefore exponentially small whenever Var(g_branch) is inverse-polynomial and λ⁴‖O‖² is only polynomially bounded. Since near-identity initialization violates E[U] = 0 by construction, the abstract's 'cannot erase' phrasing is exact only in a measure-zero regime of initialization schemes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper analyses trainability of coherent linear combinations of parameterized circuits (LCUs). Its central technical result is a single-branch gradient floor (§4.1, Prop. 4.1): if θ enters only branch ℓ, (U_ℓ, ∂_θU_ℓ) is independent of the other branches, and E[U_k]=0 for k≠ℓ, then Var(∂_θL) ≥ |α_ℓ|⁴Var(g^ℓ_branch), at any gate position and with no Haar assumption. Under independent zero-mean branches the paper derives exact loss (Thm 4.3) and measurement-adjacent gradient (Thm 4.9) variances, decomposed into a branch part and two interference channels E|m_ℓk|² and E(m²_ℓk); the second is evaluated through invariant bilinear forms and produces a phase-dependent term, illustrated by an exactly computable four-qubit matchgate value E(m²)=−5/32. For the interleaved branch–mixer architecture, §§5–6 give dimension bounds with an exact absorption criterion (Thm 5.2), an associative-envelope upper bound (Thm 5.4), a saturation theorem under full su(L) ancilla control with an exact centre (Thm 5.8), universality (Cor. 5.9), and — under an explicitly added joint second-moment design hypothesis — a sector variance formula (Thm 6.1) and closed form (55). §7 treats matchgate branches; §8 supplies a five-step audit, coefficient-selection rules, postselection and shot budgets, and moment-error bounds.","tokens_in":56308,"tokens_out":17455,"duration_ms":309785,"significance":"The paper is careful, and in the places I recomputed it is arithmetically exact. Every headline statement carries its hypotheses, Table 2 maps arguments to proof locations, and the authors state plainly that a large control algebra is not a sampling law, that no depth is shown to reach the design hypothesis, and that the normalized postselected gradient is not covered. Strengths worth naming explicitly: the gradient floor is proved by a clean conditional-expectation argument that needs no Haar randomness; the phase-sensitive channel E(m²) is a genuine addition to the second-moment analysis of Khatri et al., and the four-qubit value is checkable by hand; the Dirichlet specialization (Thm 4.10) and the number in Fig. 2 agree with independently computed values I reproduced (Eqs. 38, 57, and both curves of Fig. 2c); and the saturation theorem pins the central directions rather than only a dimension. The result is a useful, falsifiable design guide for LCU architectures, and the paper is honest about what it does not establish.","major_comments":[{"comment":"The floor is exact only at exactly zero branch means, and the manuscript nowhere quantifies how it degrades. §3.6 (b232) acknowledges that shared-parameter and near-identity ensembles 'require checking these hypotheses separately', so the condition is disclosed; what is missing is any statement of the consequence. Prop. H.2 does not supply one, for two reasons the manuscript itself would recognize: it bounds the loss variance rather than the gradient variance, and its error scale is λ⁴‖O‖²Σ_ℓ w_ℓε_ℓ, not the scale |α_ℓ|⁴Var(g^ℓ_branch) at which the floor lives. The failure is not perturbative in the relevant sense. Writing δ_k = E[U_k] ≠ 0, each interference term acquires the piece r_k = 2Re(α_ℓᾱ_k Tr[O U̇_ℓ ρ δ_k†]), which is a function of the same U̇_ℓ that fixes g^ℓ_branch; the covariance with the branch term is generically nonzero and its sign is not controlled, so the only unconditional statement is the Cauchy–Schwarz one, Var(∂_θL) ≥ |α_ℓ|⁴v + Σ_k Var(R_k) − 2|α_ℓ|²√(v Σ_k Var(R_k)), which is vacuous exactly when Σ_k Var(R_k) = |α_ℓ|⁴v. Because the deterministic part r_k is explicit, a first-order correction is available at little cost: conditioning on U̇_ℓ gives Cov(g, Σ_k r_k) in closed form and hence a floor of the shape |α_ℓ|⁴v − C‖δ‖√(v·Var(g)) + O(‖δ‖²). I would like either such a proposition, or at minimum a worked near-identity example (for instance U_ℓ(θ)=e^{iθH} with the undifferentiated branches evaluated where E[U]=I) showing how fast the guaranteed floor collapses. Without one of these, the abstract's 'cannot erase' is exact only for ensembles whose undifferentiated branches are fully randomized, which is also the regime in which those branches typically concentrate on their own.","section":"§4.1 (Prop. 4.1) and §8.4 (Prop. H.2)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The abstract contrasts polynomial single-pass LCU variances with exponential concentration of matchgate interleaved circuits, but the second half is conditional on the joint-design hypothesis, and §7 states at b596–b597 that neither analysis locates the crossover nor verifies convergence, 'without numerical verification'. The conditional framing is correct and clearly labelled, and Fig. 2's caption properly separates the ensembles; the concern is that the abstract's pairing invites the reading that interleaving itself, rather than the added sampling assumption, produces concentration. Since matchgate circuits are classically simulable in the Gaussian formalism used in Appendix F, a modest numerical anchor is feasible: for L=2 and n up to about 10, compute the first two conjugation moments of the interleaved ensemble as a function of depth T and report the depth at which they approach the joint Haar values, or show that at accessible depth the discrepancy is large. Either outcome would sharpen the architectural contrast; an explicit statement that the interleaved ensemble at polynomial depth is provably far from a second-moment design over su(L2^{n−1})^{⊕2} would serve equally well. As it stands, Eq. (56) and the Fig. 2(c) curves are correct benchmark values for an assumed sampling law, not evidence about interleaved circuits at any depth.","section":"§7, 'Fixed depth and the joint Haar-random variance bound', and §6.1","authors_acknowledge":"The finite-depth crossover for the interleaved family remains a conditional prediction here, without numerical verification.","authors_acknowledge_p":0.75}],"minor_comments":[{"comment":"The caption says the detailed definitions appear in Section 3.6, but PREP, SELECT and Eq. (13) are defined in §3.4, while §3.6 defines the two circuit models. Please correct the cross-reference.","section":"Figure 1 caption","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The glyph P carries three distinct meanings: the projected purity P_g(X) in Eq. (4), the ancilla projectors P_ℓ in Eq. (12), and the fermion parity operator Z^{⊗n} in §7 and Example 4.7. Similarly κ_ℓ denotes a branch commutant dimension while κ_s denotes the joint one. Distinct symbols for the projected purity and the parity operator would remove a real source of misreading.","section":"Notation, §§3.3, 3.4, 4.7, 7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"PREP_R and PREP_L are defined as separate unitaries, but Eq. (16) names a single coefficient-layer DLA g_coeff, and Definition 5.1 uses only that one algebra. Please state whether the model assumes a single trainable coefficient layer (with PREP_R = PREP_L up to fixed phases) or two, and which algebra enters Eq. (44) in the two-preparation case.","section":"§3.4, Eq. (16)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Four objectives now coexist — L, L/λ², L_BE and L_BE/λ² — with different prefactors, and the text switches between them in a sentence ('all of its variance formulas are obtained by dividing the displayed variances by λ⁴'). A one-row table listing each objective, its definition and the prefactor carried by its variance would eliminate the ambiguity.","section":"Eq. (20) and §6.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence 'No Haar-random sampling law, zero mean for U_ℓ, or restriction on the position of the differentiated gate is required' is correct, but it sits immediately before Theorem 4.9, whose closed form is restricted to measurement-adjacent gates. Please add an explicit pointer in the floor's remark so the two scopes are not conflated by a reader skimming §4.","section":"§4.1, remark after Prop. 4.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The 'Gradient floor' row should list the finite-second-moment hypothesis and should say that the zero-mean condition E[U_k]=0 is entrywise (matrix mean), since §4.2 rightly stresses the distinction between a zero matrix mean and a zero conjugation moment.","section":"Table 2, first row","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The two invariance conditions are typeset on a single line and appear to be a conjunction, whereas the first is the infinitesimal commutant condition and the second the infinitesimal form condition. Please separate them and note that the second is needed only for the E(m²) channel.","section":"§8.1, step A2, Eq. (58)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"'Theorem 4.10 recovers their identical-group formula when the unconjugated overlap moment vanishes' reads as an unconditional claim, while Remark 4.11 shows the Dirichlet-averaged variance acquires the extra term (43) whenever τ>0 and the input does not satisfy Lemma B.1. Please foreshadow that qualification in §2.","section":"§2, paragraph on Khatri et al.","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The caption should say that panel (a)'s 'Haar-random preparation' curve is the L=2, λ=1 specialization of Theorem 4.10 and that panel (b)'s floor is Eq. (24) with Var(g_branch)=4/(2n−1). For Eq. (57), one intermediate line giving E|Γ₁₂|²=2/2^n and the resulting (1/4)2^{−n} would help, since the exponents are easy to misread in the typeset form.","section":"Fig. 2 caption and Eq. (57)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"'Direct multiplication gives' precedes the two contractions (35)–(36); a single intermediate step for Tr[J†₊ O J₊ O^T]=−5/4 would let a reader verify the number without redoing the Majorana algebra, which is the point of including an exactly solvable example.","section":"Example 4.7","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Short version: the derivations hold up and the authors label their own limits, which is rarer than it should be. The single-pass half is the part worth your time; the interleaved half is a control-algebra result wearing a variance result's clothes.\n\nWhat is actually new. Proposition 4.1 is simple and useful: with fixed coefficients, if the other branches have zero entrywise means and are independent of the differentiated branch and its derivative, the interference terms have zero conditional mean, the cross-covariances vanish, and the |α_l|^4 floor is just a sum of squares. Anyone could have written that argument; the point is nobody did. The second real contribution is retaining the unconjugated moment E(m^2) at fixed coefficients (Theorems 4.3 and 4.9). Khatri et al. average it away with Dirichlet weights; here it survives with explicit phase dependence, and the four-qubit matchgate example (E(m^2) = −5/32, Eq. 38) shows it can be nonzero. Lemma 4.6 is the right tool for that and the matchgate counts τ = 2 are checked. Theorem 5.8 is a clean block-by-block answer — su(Ld_i) per Wedderburn block plus a q-dimensional center — and the center bookkeeping in Example 5.7 is exactly where these papers usually slip. The trace-component condition in Corollary 5.9 is a nice catch. Theorem 4.10 recovering Khatri's Eq. (9) term by term is the kind of check that earns trust.\n\nWhere it is soft. Section 3.6 (b232) already says the floor's hypotheses fail for shared-parameter and near-identity ensembles, so the stress-test concern is acknowledged rather than concealed. I would push back on how the stress test framed it: zero-mean branch ensembles are a modeling class, not a measure-zero accident — Haar on a group with no fixed vectors, or any ensemble with a uniform global phase, satisfies it. What is fair is narrower: the abstract's \"cannot erase\" reads stronger than the theorem, and Proposition H.2 bounds moment error at the λ^4‖O‖² scale rather than the branch-variance scale, so approximate moments do not obviously extend the floor. One careful sentence in the abstract would close that gap. Second, the closed-form gradient covers only measurement-adjacent gates; other positions need the edge or independent-factor conditions of C.2, which the text states (b609). Third, all the interleaved concentration statements are conditional on joint-design sampling (b512, b597, b699). That is honest, but it means the second half of the paper is a control paper plus a conditional prediction, and readers should not read it as a blank-plateau theorem for interleaved LCUs.\n\nThe Section 8.1 audit and the weight thresholds in Theorem I.1 are practically useful and I would hand them to a student.\n\nWho it is for: anyone building LCU-based models who needs to know which quantities survive coefficient tuning, and anyone working on DLA-based variance theory. The citation pattern is normal — the borrowed theorems come from published or posted sources and are attributed, including the explicit concession that Theorem 4.3 specializes Alcântara et al.\n\nRecommendation: send it out. Ask the referees to press on the abstract wording for the floor and on whether the near-identity regime can be handled at all, and to check the saturation proof's step 3 independently.","headline":"The math is right and honestly scoped; the single-pass gradient floor and the fixed-coefficient phase-sensitive moments are real contributions, while the interleaved half is a control result plus a variance claim that is conditional on a sampling assumption nobody has shown to hold.","tokens_in":56846,"tokens_out":2355,"duration_ms":18591},"pith_extraction":{"msc":["81P68","81Q93","81R05"],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"The paper argues that a coherent sum of quantum circuits cannot, under stated independence and zero-mean conditions, erase the training signal of an individual branch, and that interference contributes through two separable moment…","keywords":["barren plateaus","linear combination of unitaries","gradient variance","dynamical Lie algebra","matchgate circuits","invariant bilinear forms","quantum circuit trainability","postselection"],"falsifier":"An explicit two-branch single-pass linear combination with independent, zero-mean branches in which the total gradient variance falls below $|\\alpha_\\ell|^4$ times the differentiated branch's own gradient variance would refute Proposition 4.1; conversely, a numerical evaluation of the four-qubit matchgate example that fails to show a phase-dependent interference term vanishing at $\\gamma_1=\\gamma_2$ and maximal near $\\pi/2$ would undercut the invariant-form channel.","tokens_in":57278,"feed_emoji":"⚛","tokens_out":4927,"duration_ms":33524,"temperature":0.7,"pith_summary":"This paper asks when a coherent sum of quantum circuits — a linear combination of unitaries — destroys the training signal that one of its constituent circuits would have had on its own. Its central claim is a floor: if the differentiated branch is independent of the others and every other branch has zero average as a matrix, interference cannot push the total gradient variance below that branch's own gradient variance times the fourth power of its coefficient magnitude. The result needs no Haar-randomness and no particular gate position, which makes it a design rule rather than a statement about one sampling convention. Beyond the floor, the paper gives exact bookkeeping for the two kinds of interference term — one set by the squared magnitude of the overlap between two branch outputs, one set by the complex square of that overlap — and shows the second is the only place relative coefficient phases enter. It then turns to circuits that alternate branch operations with ancilla mixing, computes what control those circuits generate, and uses matchgate examples to show that polynomial single-pass variance and exponential concentration under joint Haar-random sampling can coexist in the same family of branches.","feed_headline":"Coherent sums cannot erase a branch's gradient, bound shows","feed_subtitle":"If branches are independent and zero-mean, the summed circuit keeps at least the fourth power of the branch coefficient times its own…","key_machinery":"Two objects carry the argument. The first is the gradient floor of Proposition 4.1, which is a conditional-expectation argument rather than a random-matrix one. The second is the split of every pairwise interference moment into two invariant channels. The conjugated moment $\\mathbb{E}|m_{\\ell k}|^2$ is governed by the commutant of the branch group — the matrices $Y$ with $UYU^\\dagger = Y$ — evaluated through Hilbert–Schmidt-orthonormal commutant bases. The unconjugated moment $\\mathbb{E}(m_{\\ell k}^2)$ is governed instead by the invariant bilinear forms $J$ with $UJU^{\\mathsf T} = J$; the transpose rather than the adjoint is what separates this channel, and its dimension $\\tau(G)$ vanishes exactly when the two-copy ensemble has zero mean, which is why an independent uniform global phase or a group such as $U(d)$ removes the phase-dependent term. For the interleaved architecture the organising object is the associative envelope of the branch generators: ancilla commutators turn symmetric products into Lie brackets, so the generated Lie algebra can saturate the envelope block by block even when the joint branch Lie algebra is small, with the only shortfall an at-most-$q$-dimensional centre of block phases.","core_discovery":"The load-bearing claim is Proposition 4.1, a single-branch gradient floor. Fix the coefficients, let $\\theta$ occur only in branch $\\ell$, and let $g^{\\ell}_{\\mathrm{branch}}$ be that branch's own gradient. If the pair $(U_\\ell, \\dot U_\\ell)$ is independent of all other branch unitaries, and each other branch has zero entrywise mean, then $\\mathrm{Var}(\\partial_\\theta L) \\ge |\\alpha_\\ell|^4\\, \\mathrm{Var}(g^{\\ell}_{\\mathrm{branch}})$. The proof conditions on the differentiated branch, notes that each interference term then has zero conditional mean, and keeps only the nonnegative average squares that remain. Nothing about Haar measure, the position of the differentiated gate, or the mean of $U_\\ell$ itself is used. The corollary the authors draw is that an inverse-polynomial branch variance plus an inverse-polynomial $|\\alpha_\\ell|$ is enough to keep the unnormalized gradient variance inverse-polynomial, so the coherent sum does not by itself manufacture a barren plateau. The rest of the single-pass analysis makes the accounting exact: the loss variance splits into branch variances weighted by $|\\alpha_\\ell|^4$ and pairwise interference terms built from $\\mathbb{E}|m_{\\ell k}|^2$ and $\\mathbb{E}(m_{\\ell k}^2)$, where $m_{\\ell k} = \\mathrm{Tr}[O\\,U_\\ell\\rho U_k^\\dagger]$; only the second carries the relative phases $\\gamma_\\ell-\\gamma_k$. A four-qubit matchgate example gives $\\mathbb{E}(m_{12}^2) = -5/32$ and an interference contribution $\\tfrac{5}{16}\\lambda^4 w_1^2 w_2^2[1-\\cos 2(\\gamma_1-\\gamma_2)]$ that vanishes at equal phases and peaks at relative phase $\\pi/2$.","pith_inferences":["The fourth-power coefficient factor is a concrete design tax on wide coherent sums: at $L$ balanced branches the hardware-scaled variance carries $L^{-4}$ before any interference effect, so adding branches to an LCU is not free even in the floor's best case.","Because only the invariant-form channel carries relative phases, the phase-dependent part of the variance is a measurable fingerprint of whether a branch group is self-dual; a four-qubit experiment with matchgate branches, a GHZ input, and two chosen phases would test the predicted null at equal phases and peak at $\\pi/2$.","The gap between the joint Lie closure and the associative envelope suggests that dynamical-Lie-algebra dimension is a weak proxy for trainability in interleaved architectures, since the same small branch algebra can sit under either a small envelope or a maximal one.","Shared-parameter architectures that draw all branches from one circuit with token-dependent angles violate the independence premise outright, so measuring cross-branch correlations in such models is the natural next test of whether the floor applies."],"forward_implications":["An inverse-polynomial branch gradient variance together with an inverse-polynomial coefficient magnitude suffices for an inverse-polynomial unnormalized gradient variance, at any gate position and without assuming Haar-random branches.","Relative coefficient phases matter only through the invariant-form channel, so phase tuning changes the loss variance and the added gradient variance in opposite directions; for two branches, balanced weights maximize the loss variance only when the interference term is at least seven times the branch variance.","With a full special-unitary ancilla mixer, the interleaved algebra saturates the associative envelope up to central block phases, so branches whose joint Lie closure is small — such as independent Pauli-$X$ and Pauli-$Z$ branches of dimension $3n$ — can still generate universal joint control.","For matchgate branches the single-pass LCU keeps inverse-polynomial loss and gradient variance, while the interleaved circuit under joint Haar-random sampling concentrates as $1/(L^2 4^{n-1}-1)$, and fermion parity survives in both regimes as a superselection rule.","Normalized postselected losses need separate treatment: the unnormalized variance transfers only when success-factor fluctuations are small relative to loss fluctuations, and an accepted shot costs $\\lambda^2/p_{\\mathrm{acc}}$ circuit repetitions on average."],"supporting_citations":[{"why":"Supplies the block-encoding construction of a linear combination of unitaries that defines the single-pass circuit and its postselection probability.","marker":"[5]"},{"why":"Provides the second-moment expansion and Dirichlet-averaged LCU loss-variance formula that Theorem 4.10 reproduces term by term and the fixed-coefficient analysis extends.","marker":"[6]"},{"why":"Supplies the dynamical Lie algebra as the control object connecting gate generators to trainability.","marker":"[7]"},{"why":"Supplies the sector-sum variance formula in Lemma A.2 that the joint-space reduction relies on.","marker":"[8]"},{"why":"Supplies the general operator-space decomposition that Theorem 4.3 specializes to the LCU setting once the membership condition fails.","marker":"[11]"},{"why":"Supplies the matchgate variance analysis beyond the dynamical Lie algebra and the Majorana representation used in the worked examples.","marker":"[10]"},{"why":"Supplies the Haar-measure integration identities on unitary, orthogonal, and symplectic groups behind the invariant-form pairings.","marker":"[14]"},{"why":"Supplies the fixed-depth stacked-LCU variance lower bound the interleaved matchgate regime is compared against.","marker":"[17]"},{"why":"Supplies the exact four-point parameter-shift rule used in the shot-budget proposition.","marker":"[39]"}],"fun_headline_variants":["Coherent sums keep a branch gradient floor without Haar randomness","Single-pass LCU gradient variance stays above branch variance floor","Interference adds exact terms but cannot erase one branch's gradient","Repeated ancilla mixing predicts concentration only with extra sampling assumptions","Matchgate example gives nonzero phase-dependent LCU interference term"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The floor rests on the differentiated branch and its derivative being independent of every other branch, and on those other branches averaging to zero as matrices; shared parameters or correlated branch ensembles break this and leave interference free to offset the branch gradient in the mean.","fun_headline_variants_meta":{"raw":{"variants":["Coherent sums keep a branch gradient floor without Haar randomness","Single-pass LCU gradient variance stays above branch variance floor","Interference adds exact terms but cannot erase one branch's gradient","Repeated ancilla mixing predicts concentration only with extra sampling assumptions","Matchgate example gives nonzero phase-dependent LCU interference term"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"cost":0.00023205,"is_byok":false,"cost_details":{"upstream_inference_cost":0.00023205,"upstream_inference_prompt_cost":0.000109746,"upstream_inference_completions_cost":0.000122304},"total_tokens":1878,"prompt_tokens":1206,"completion_tokens":672,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":603}},"tokens_in":1206,"tokens_out":672,"cost_source":"openrouter_usage","duration_ms":6326,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":603,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-31T08:50:46.622592+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"An explicit two-branch single-pass linear combination with independent, zero-mean branches in which the total gradient variance falls below $|\\alpha_\\ell|^4$ times the differentiated branch's own gradient variance would refute Proposition 4.1; conversely, a numerical evaluation of the four-qubit matchgate example that fails to show a phase-dependent interference term vanishing at $\\gamma_1=\\gamma_2$ and maximal near $\\pi/2$ would undercut the invariant-form channel.","supporting_citations":[{"cited_title":"Hamiltonian Simulation Using Linear Combinations of Unitary Operations","cited_arxiv_id":null,"evidence_quote":"Supplies the block-encoding construction of a linear combination of unitaries that defines the single-pass circuit and its postselection probability."},{"cited_title":"Trainability of Parametrised Linear Combinations of Unitaries","cited_arxiv_id":"2506.22310","evidence_quote":"Provides the second-moment expansion and Dirichlet-averaged LCU loss-variance formula that Theorem 4.10 reproduces term by term and the fixed-coefficient analysis extends."},{"cited_title":"Coles, and Marco Cerezo","cited_arxiv_id":null,"evidence_quote":"Supplies the dynamical Lie algebra as the control object connecting gate generators to trainability."},{"cited_title":"A Lie algebraic theory of barren plateaus for deep parameterized quantum circuits","cited_arxiv_id":null,"evidence_quote":"Supplies the sector-sum variance formula in Lemma A.2 that the joint-space reduction relies on."},{"cited_title":"A Representation-Theoretic Framework for Characterizing Barren Plateaus","cited_arxiv_id":"2609.04462","evidence_quote":"Supplies the general operator-space decomposition that Theorem 4.3 specializes to the LCU setting once the membership condition fails."},{"cited_title":"Showcasing a Barren Plateau Theory Beyond the Dynamical Lie Algebra","cited_arxiv_id":"2310.11505","evidence_quote":"Supplies the matchgate variance analysis beyond the dynamical Lie algebra and the Majorana representation used in the worked examples."},{"cited_title":"Integration with Respect to the Haar Measure on Unitary, Orthogonal and Symplectic Group","cited_arxiv_id":null,"evidence_quote":"Supplies the Haar-measure integration identities on unitary, orthogonal, and symplectic groups behind the invariant-form pairings."},{"cited_title":"Stacking the Deck: Tunable Trainability in Stacked LCUs","cited_arxiv_id":"2607.24686","evidence_quote":"Supplies the fixed-depth stacked-LCU variance lower bound the interleaved matchgate regime is compared against."},{"cited_title":"General parameter-shift rules for quantum gradients","cited_arxiv_id":null,"evidence_quote":"Supplies the exact four-point parameter-shift rule used in the shot-budget proposition."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}