betaFeedback

WorksPhysics and Astronomy

Classically Trained Quantum Generative Models with Shared Training and Fine Tuning for Quantum Matter

Abstract

Ground-state measurement statistics characterize quantum matter. Selected correlations can guide the classical training of a quantum sampler even when sampling the full distribution is difficult. We ask whether training across related ground-state distributions helps the sampler reproduce their correlations more accurately. We train a hypernetwork to map Hamiltonian parameters to circuit angles, sharing information across the phase diagram before local fine-tuning with additional reference data for quantum deployment. For classical training, we derive an unbiased Monte Carlo estimator of Fourier coefficients for circuits built from general Pauli rotations, with a cost that depends on the circuit generators. We also introduce a CDF-KL loss for comparing selected marginals. We study the Haldane Hamiltonian and transverse-field Ising model, using an automated search to select reference calculations near phase transitions. On the twelve-site chain, the adaptively pretrained hypernetwork matches the independent fits' final mean total variation distance in fewer local updates and continues to a lower error. On a 1,024-site chain and a 4 by 256 lattice, the fine-tuned circuits have lower squared maximum mean discrepancy from the reference distributions than the independent circuits at the evaluated point in each geometry. These results motivate shared training for preparing quantum samplers across a phase diagram and local fine-tuning.

Introduction

Quantum computers are natural samplers, measurements draw outcomes from the Born distribution. Quantum circuit Born machines (QCBMs) learn target distributions by training circuit angles on reference data [1, 2, 3, 82]. Their potential advantage requires distributions that are hard to sample classically, yet can be trained to reproduce relevant data.

Complexity results for instantaneous quantum polynomial-time (IQP) circuits and random circuit sampling support classical sampling hardness under specified assumptions and error criteria [4, 5, 6]. Both have been implemented experimentally [7, 8], but sampling hardness does not establish efficient training on a specified dataset.

Two complementary approaches address training. The first develops practical objectives, gradient estimators and initialization strategies, including training Ising Born machines [3]. IQP models can be trained classically using maximum mean discrepancy (MMD) and deployed as quantum samplers [9, 10]. Classically accessible correlators provide the loss and its gradients without requiring efficient sampling of the full model distribution [10, 74, 83, 84].

The second studies trainability and classical simulation. Barren plateaus have exponentially small gradient variances [11] and depend on the circuit, loss and expressibility [12, 13]. Many provably plateau-free models admit classical simulation, sometimes after quantum data acquisition [14]. Certain anticoncentrated generative ensembles exhibit loss concentration, while restrictions that improve trainability can permit classical sampling [15]. Trainability does not necessarily imply dequantizability [16].

For IQP models, MMD trainability depends on circuit generators and kernel spectrum [17], while data-informed initialization can improve convergence [18]. Different QCBMs can agree on the correlations used for training and differ on other features of the distribution [19].

Neural representations can learn many-body ground states and their measurement statistics [20, 21]. This motivates the following question.

Can learning from related ground-state distributions improve a classically trained quantum sampler?

We use reference ground-state data to learn quantum samplers across the phase diagram and fine-tune them at

selected points. To this end, we jointly train a hypernetwork and shared QCBM parameters across a family of Hamiltonians. The hypernetwork maps Hamiltonian parameters to circuit angles through a scale-and-shift construction [22], while each update combines losses from several reference distributions. Learning across Hamiltonians also appears in neural quantum states [23] and in provable prediction of ground-state properties within a quantum phase [24, 25]. A related approach introduces the external parameters directly into the Born-machine circuit through input gates [26]. We test whether shared training helps recover correlations absent from the loss.

Obtaining reference ground states can be costly, so we also use the learned family to choose where to request additional data. Rapid changes in its measurement distribution guide this search towards possible phase boundaries, without requiring a predefined order parameter.

Our starting point is likelihood training, equivalent to minimizing an empirical Kullback–Leibler (KL) divergence objective across the reference distributions. When evaluating model probabilities becomes costly, we use selected Fourier coefficients instead. As an alternative to MMD, we use KL divergences to compare the cumulative distribution functions (CDFs) of selected marginals. We call this loss CDF-KL. We also derive an unbiased Monte Carlo (MC) estimator for the model coefficients of ordered, generally noncommuting Pauli circuits. Its cost depends on the number of amplitudes evolved in each MC sample, which is determined by the selected gates. The choice of generators therefore determines both the cost of classical training and the circuit’s inductive bias, including its preserved Pauli symmetries.

Using Monte Carlo coefficients also affects how we measure agreement with the target distribution. Sampling fluctuations can bias a nonlinear loss even when each coefficient estimate is unbiased. We correct this bias for squared MMD [10, 27], while our sampled CDF-KL loss remains biased. The sampled MMD curves in Figure 8 and Figure 9 use the variance correction derived in Appendix D 3.

Figure 8. Shared training and local refinement for an open 1,024-site Haldane chain. (a) Peaks in the learned marginal response, with nine initial references (gray), 41 acquired references (orange) and the published bulk boundaries (dashed) [29]. The ring marks the local fitting coordinate. (b) Bias-corrected squared MMD against one million held-out samples from the reference MPS (bond cap 128). The black curve compares the 20,000 local training samples with the held-out samples. Bars show standard errors across 40 groups.

Figure 9. Shared training and local refinement for an open 4 by 256 spin lattice. (a) Peaks in the learned marginal response, with nine initial references (gray) and 41 acquired references (orange). The ring marks the local fitting coordinate. (b) Bias-corrected squared MMD against one million held-out samples from the reference MPS (bond cap 128). The black curve compares the 20,000 local training samples with the held-out samples. Bars show standard errors across 20 groups.

Together, shared training, classical coefficient estimation and local fine-tuning provide a framework for learning quantum samplers across a phase diagram and testing whether sharing information across Hamiltonians improves the fit to individual target distributions.

The paper is organized as follows. Sec. II introduces the target distributions, hypernetwork and circuit generators. Sec. III develops classical training with selected Fourier coefficients, including the Monte Carlo estimator and the CDF-KL and MMD losses. In § 2, we examine distribution and response accuracy in small transverse-field Ising and Haldane systems, then restrict the reference data, coefficient information and allowed gates. In § 2.2, we adapt the autonomous search of Ref. [28] to concentrate costly reference calculations near phase boundaries. Sec. VI compares shared and independent training at common boundary points and examines local fine-tuning. Sec. VII describes the large-system protocol using selected-marginal responses. Sec. VIII discusses the findings and limitations. Figure 1 summarizes the training and deployment procedure. The appendices provide derivations and numerical settings.

Learning across a phase diagram to sample at selected Hamiltonian points

Figure 1. Learning across a phase diagram to sample at selected Hamiltonian points λ\lambda. The hypernetwork maps λ\lambda to scale and shift vectors that set the circuit angles with shared base angles, Eq. (4). Both are trained jointly across reference distributions. Autonomous discovery adds reference-state calculations near rapid distribution changes, without phase labels or a dense sweep. Local fine-tuning uses additional reference data before quantum deployment.

Learning a family of measurement distributions

Targets and reference data

Let H(λ)H(\lambda) be a Hamiltonian with parameter vector λ∈Λ⊂Rdλ\lambda\in\Lambda\subset\mathbb{R}^{d_\lambda}, where Λ\Lambda is the parameter domain considered and dλd_\lambda is the number of Hamiltonian parameters. For a specified ground state ∣ψ0(λ)⟩|\psi_0(\lambda)\rangle, the target is

pλ(x)=∣⟨x∣ψ0(λ)⟩∣2,x∈{0,1}n.(1)p_\lambda(\mathbf{x}) = \left|\langle\mathbf{x}|\psi_0(\lambda)\rangle\right|^2,\qquad\mathbf{x}\in\{0,1\}^n . \tag*{(1)}

This fixes the measurement basis. Matching pλp_\lambda does not in general recover the phases of the wavefunction or establish quantum-state fidelity.

We use two main targets. The transverse-field Ising model (TFIM) is

HI=−J∑⟨i,j⟩ZiZj−g∑iXi,(2)H_{\mathrm{I}}=-J\sum_{\langle i,j\rangle}Z_iZ_j-g\sum_iX_i, \tag*{(2)}

where the numerical benchmark uses a periodic 3×33\times3 lattice. The second target is the open, cluster-type Haldane chain

HH=−J∑i=1n−2ZiXi+1Zi+2−h1∑i=1nXi−h2∑i=1n−1XiXi+1,(3)H_{\mathrm{H}}=-J\sum_{i=1}^{n-2}Z_iX_{i+1}Z_{i+2}-h_1\sum_{i=1}^{n}X_i-h_2\sum_{i=1}^{n-1}X_iX_{i+1}, \tag*{(3)}

with λ=(h1/J,h2/J)\lambda=(h_1/J,h_2/J). This is the spin model studied in Ref. [29], rather than the usual fermionic honeycomb Haldane model [30]. We fit a separate model for each target family. Information is shared across its Hamiltonian parameters.

At each reference point λj\lambda_j, we obtain a ground state and draw a dataset Dj\mathcal{D}_j of computational-basis outcomes. Small systems use exact diagonalization. Larger chains use matrix product states (MPSs) and density-matrix renormalization group (DMRG) [31, 32]. The latter supply approximate reference states whose accuracy depends on the solver and bond dimension. Training receives samples or selected correlators, whereas full probabilities are retained only for small-system evaluation. Appendix A specifies the boundary conditions and how reference states are selected when the ground state is degenerate.

A hypernetwork generates the circuit angles

A hypernetwork is a neural network that supplies parameters of another model [33, 34]. Related quantum approaches use neural networks to map Hamiltonian parameters to circuit angles [35] or tensor-train hypernetworks to generate variational circuit parameters [36]. Following the ground-state tomography construction of Ref. [22], we map normalized Hamiltonian parameters to a scale and shift of a shared base vector,

θ(λ)=[1+γϕ(λ)]⊙θ0+βϕ(λ).(4)\boldsymbol{\theta}(\boldsymbol{\lambda}) = [1+\boldsymbol{\gamma}_{\boldsymbol{\phi}}(\boldsymbol{\lambda})]\odot\boldsymbol{\theta}_0+\boldsymbol{\beta}_{\boldsymbol{\phi}}(\boldsymbol{\lambda}). \tag*{(4)}

Here ⊙\odot denotes componentwise multiplication, (a⊙b)j=ajbj(\mathbf{a}\odot\mathbf{b})_j=a_jb_j, and 1\mathbf{1} is the vector of ones. The scale-and-shift operation is also known as feature-wise linear modulation [37]. We train the base circuit angles θ0\boldsymbol{\theta}_0 and hypernetwork weights ϕ\boldsymbol{\phi} jointly, writing η=(θ0,ϕ)\boldsymbol{\eta}=(\boldsymbol{\theta}_0,\boldsymbol{\phi}). At each Hamiltonian point, the hypernetwork receives the normalized coordinates of λ\boldsymbol{\lambda}, such as the two coupling ratios specifying the Haldane Hamiltonian, and outputs the vectors γϕ(λ)\boldsymbol{\gamma}_{\boldsymbol{\phi}}(\boldsymbol{\lambda}) and βϕ(λ)\boldsymbol{\beta}_{\boldsymbol{\phi}}(\boldsymbol{\lambda}). Reference bitstrings and coefficients estimated from them are used to calculate the training loss, rather than being supplied as inputs to the hypernetwork. Each training update averages this loss over the available reference points and uses its gradients to update both θ0\boldsymbol{\theta}_0 and ϕ\boldsymbol{\phi}, as written explicitly in Appendix E 1. The reference data therefore guide the learned map from Hamiltonian parameters to circuit angles through these updates. The circuit architecture remains fixed as its angles change with λ\boldsymbol{\lambda}.

Let ncn_c be the total number of circuit qubits, including the nn measured qubits and nc−nn_c-n unmeasured ancillas. Write a∈{0,1}nc−n\mathbf{a}\in\{0,1\}^{n_c-n} for the ancilla outcomes. The model distribution is

qη,λ(x)=∑a∣⟨x,a∣Uθ(λ)∣ψin⟩∣2.(5)q_{\boldsymbol{\eta},\boldsymbol{\lambda}}(\mathbf{x})=\sum_{\mathbf{a}}\left|\langle\mathbf{x},\mathbf{a}|U_{\boldsymbol{\theta}(\boldsymbol{\lambda})}|\psi_{\mathrm{in}}\rangle\right|^2. \tag*{(5)}

We call the combined hypernetwork and QCBM model HyperQCBM.

We include ancillas to enlarge the family of distributions on the measured qubits. This is motivated by the universality of IQP Born machines with sufficiently many hidden qubits [38], although that result does not establish universality for our restricted circuits. Ancillas are marginalized in both training and evaluation. Once trained, the hypernetwork supplies circuit angles at a new parameter value without further optimization or reference data at that point. Quantum deployment consists of preparing ∣ψin⟩|\psi_{\mathrm{in}}\rangle, applying the generated circuit and reading out the nn measured qubits. Before deployment, we can fine-tune the circuit using additional reference samples from the chosen Hamiltonian point. It starts from information learned across the diagram instead of fitting an independent circuit from its initial parameters.

Circuit generators and conserved symmetries

We use an ordered product of Pauli rotations,

Uθ=e−iθdPd⋯e−iθ1P1,(6)U_{\boldsymbol{\theta}}=e^{-i\theta_dP_d}\cdots e^{-i\theta_1P_1}, \tag*{(6)}

where each Hermitian Pauli string satisfies Pj2=IP_j^2=I. Lie algebras provide a way to select generators with specified conservation laws [39, 40]. This choice supplies an inductive bias by restricting the transformations available at every Hamiltonian point, independently of the fitted angles. If a Pauli symmetry commutes with every generator, the circuit preserves it throughout training and deployment. Its restriction on the output distribution also depends on the input state. We compare Haldane, Heisenberg and quadratic fermionic families with commuting XX-string IQP circuits. The algebra determines which Pauli generators are allowed. The circuit model is defined by the generators we select and the order in which we apply their gates. A fixed gate sequence need not realize every unitary generated by the algebra.

For the Haldane family, permitted strings have an even number of Y/ZY/Z factors on each of the two circuit sublattices. They commute with the corresponding products of XX operators. Heisenberg generators preserve global XX and ZZ parity. Quadratic fermionic generators are Majorana bilinears, and IQP generators contain only XX and identity. Appendix B gives compact algebra expressions, gate selection and parameter counts. Gate weight and pair distance are distinct restrictions.

Preparation matters as well as the generators. For the main Haldane and IQP studies we prepare the Greenberger–Horne–Zeilinger (GHZ) state ∣GHZ⟩=12(∣0⟩⊗nc+∣1⟩⊗nc)|\mathrm{GHZ}\rangle=\frac{1}{\sqrt{2}}(|0\rangle^{\otimes n_c}+|1\rangle^{\otimes n_c}). If every gate commutes with global XX, this preparation gives q(x)=q(x‾)q(\mathbf{x})=q(\overline{\mathbf{x}}), including after ancilla marginalization. All odd-order ZZ correlators on measured qubits then vanish. The use of an input symmetry vector follows the classical-training construction of Ref. [10]. A computational zero input would generally give a different family of distributions. GHZ preparation requires one Hadamard and nc−1n_c-1 controlled-NOT gates in a connected tree, with depth determined by connectivity. Other preparations used in the numerical comparisons are stated explicitly.

Classical training with selected correlations

From likelihood to Fourier coefficients

Our initial joint training of the hypernetwork and base circuit angles uses a KL-divergence objective. With reference samples and accessible model probabilities, this is implemented by minimizing the empirical negative log likelihood (NLL)

LNLL(η)=−1K∑j=1K1Nj∑x∈Djlog⁡qη,λj(x).(7)\mathcal{L}_{\mathrm{NLL}}(\boldsymbol{\eta})=-\frac{1}{K}\sum_{j=1}^{K}\frac{1}{N_j}\sum_{\mathbf{x}\in D_j}\log q_{\boldsymbol{\eta},\boldsymbol{\lambda}_j}(\mathbf{x}). \tag*{(7)}

All reference points receive equal weight. The difference between this loss and the mean KL divergence from the empirical targets is their entropy, which is independent of η\boldsymbol{\eta}. This is a choice of training objective, not a requirement of the hypernetwork architecture. Its present implementation calculates the model probability of every possible outcome and is used as a small-system baseline. To avoid calculating the probability of every possible outcome, write a distribution in its Walsh–Hadamard basis,

cS(p)=⟨ZS⟩p=Ex∼p(−1)∑i∈Sxi,(8)c_S(p)=\langle Z_S\rangle_p=\mathbb{E}_{\mathbf{x}\sim p}(-1)^{\sum_{i\in S}x_i}, \tag*{(8)}
p(x)=2−n∑S⊆[n]cS(p)(−1)∑i∈Sxi.(9)p(\mathbf{x})=2^{-n}\sum_{S\subseteq[n]}c_S(p)(-1)^{\sum_{i\in S}x_i}. \tag*{(9)}

Here ZS=∏i∈SZiZ_S=\prod_{i\in S}Z_i, with identity on the remaining qubits and ancillas. The support of a coefficient is the set SS of qubits on which ZSZ_S acts nontrivially. The order of a coefficient is the number of qubits in its support, ∣S∣|S|, and specifies how many measured bits enter the parity correlation. The empty support gives c∅=1c_{\varnothing}=1. Target coefficients are estimated by sample averages of the same parity.

Selecting a set Ω\Omega lets us train on a subset of the 2n2^n Fourier coefficients. Several distributions can agree on Ω\Omega [19]. Shared pretraining uses information beyond what an MMD fit at a single Hamiltonian point can provide. The initial angles for local fine-tuning reflect how the selected correlations vary across related Hamiltonians. We test whether this also helps recover correlations beyond the training order. Classical shadows also use selected measurements to estimate properties [41]. Here the reference data come from a fixed measurement basis, and the estimator computes observables of the model circuit.

Monte Carlo estimates for ordered Pauli circuits

We now turn to calculating the model coefficients without constructing the full output distribution. For IQP circuits built from commuting XX strings, coefficients can be estimated by averaging phase functions over random bitstrings [10, 42]. We extend this approach to ordered Pauli rotations so that classical coefficient-based training can also use noncommuting generator families, including those chosen to preserve the symmetries discussed in Sec. II C. We apply the gates in the order given in Eq. (6), without assuming that they commute.

The idea is to write a coefficient cS=⟨ψin∣Uθ†ZSUθ∣ψin⟩c_S=\langle\psi_{\mathrm{in}}|U_\theta^\dagger Z_S U_\theta|\psi_{\mathrm{in}}\rangle as the expectation of a quantity calculated from a random bitstring. Each bitstring selects two normalized vectors that we evolve within invariant subspaces of the circuit. The real part of their overlap gives one Monte Carlo sample, and averaging independent samples gives an unbiased estimate of cSc_S. These random bitstrings are inputs to a classical calculation, not samples from the model’s output distribution.

The cost depends both on how many samples we need and on how much work each sample takes. The variance controls the number of samples needed for a specified additive accuracy. We show that the sampled overlap is bounded, giving a variance bound independent of system size.

Evaluating each overlap can nevertheless be expensive because the vectors may have exponentially many entries. We therefore examine how the gate generators determine their size and which restrictions make the calculation efficient.

We first take ∣ψin⟩=∣0⟩⊗nc|\psi_{\mathrm{in}}\rangle=|0\rangle^{\otimes n_c}, then extend the construction to the GHZ input used in our Haldane circuits. Appendices C1 and C2 give the gate updates, estimator derivation and error bounds.

Gate action in the XX basis. We work in the XX basis because XX-string rotations only change phases there. We can then identify which basis states the remaining gates mix and how many amplitudes each Monte Carlo sample requires. The output is still measured in the computational basis.

We label each XX-basis state of the complete circuit register by a bitstring b\mathbf{b} and write

∣b~⟩=H⊗nc∣b⟩,|\tilde{\mathbf{b}}\rangle=H^{\otimes n_c}|\mathbf{b}\rangle,

where HH is the single-qubit Hadamard gate. A bit bi=0b_i=0 denotes ∣+⟩i|+\rangle_i, while bi=1b_i=1 denotes ∣−⟩i|-\rangle_i. Each label covers all ncn_c qubits, including the ancillas. Ancilla measurement outcomes are still denoted by a\mathbf{a}.

To describe the gates, we use binary vectors in F2nc\mathbb{F}_2^{n_c}, where F2={0,1}\mathbb{F}_2=\{0,1\} with arithmetic modulo two. We write ⊕\oplus for bitwise addition and evaluate binary dot products modulo two. Each Pauli word is specified by two such vectors, xj\mathbf{x}_j and zj\mathbf{z}_j, through

Pj=iνjXxjZzj,νj=∑i=1nc(xj)i(zj)i.(10)P_j=i^{\nu_j}X^{\mathbf{x}_j}Z^{\mathbf{z}_j},\qquad\nu_j=\sum_{i=1}^{n_c}(\mathbf{x}_j)_i(\mathbf{z}_j)_i. \tag*{(10)}

Here Xxj=⨂iXi(xj)iX^{\mathbf{x}_j}=\bigotimes_i X_i^{(\mathbf{x}_j)_i} and Zzj=⨂iZi(zj)iZ^{\mathbf{z}_j}=\bigotimes_i Z_i^{(\mathbf{z}_j)_i}. A zero exponent contributes the identity in the corresponding tensor product. A YY factor has both entries equal to one, with the required phase supplied by iνji^{\nu_j}. The sum defining νj\nu_j is an ordinary integer sum, so it counts the YY factors rather than their parity. Any overall minus sign in PjP_j can be absorbed into the real rotation angle θj\theta_j.

In this basis, XiX_i multiplies a basis state by (−1)bi(-1)^{b_i}, while ZiZ_i flips the bit bib_i. Applying PjP_j therefore changes the basis label from b\mathbf{b} to b⊕zj\mathbf{b}\oplus\mathbf{z}_j and multiplies the state by a phase. Appendix C1 gives that phase explicitly. The rotation e−iθjPje^{-i\theta_jP_j} can consequently mix only the states labelled by b\mathbf{b} and b⊕zj\mathbf{b}\oplus\mathbf{z}_j. If zj=0\mathbf{z}_j=0, each basis state acquires only a phase.

Successive gates can change the initial bitstring only by adding their vectors zj\mathbf{z}_j modulo two. We collect all possible binary sums of these vectors in

V=span⁡F2{z1,…,zd},r=dim⁡V.(11)V=\operatorname{span}_{\mathbb{F}_2}\{\mathbf{z}_1,\ldots,\mathbf{z}_d\},\qquad r=\dim V. \tag*{(11)}

The dimension rr is the number of linearly independent vectors needed to generate these sums. Starting from ∣b~⟩|\tilde{\mathbf{b}}\rangle, the circuit therefore remains within the span of basis states labelled by b+V={b⊕v:v∈V}\mathbf{b}+V=\{\mathbf{b}\oplus\mathbf{v}:\mathbf{v}\in V\}. This set is called a coset and contains 2r2^r bitstrings. We simulate the evolution within one such subspace at a time, using a normalized vector with 2r2^r entries.

Calculating a Monte Carlo sample. To calculate one Monte Carlo sample, we use a matrix B∈F2nc×rB \in\mathbb{F}_2^{n_c \times r} whose columns form a basis of VV, chosen once for all samples. For a sampled bitstring b\mathbf{b}, the vector entry indexed by t∈F2r\mathbf{t} \in\mathbb{F}_2^r corresponds to the XX-basis state specified by b⊕Bt\mathbf{b} \oplus B\mathbf{t}. For zero input, every entry initially equals 2−r/22^{-r/2}. Apply gates 1,…,d1,\ldots,d in order to obtain the normalized vector wθ(b)(t)w_{\theta}^{(\mathbf{b})}(\mathbf{t}).

Let s∈F2nc\mathbf{s} \in\mathbb{F}_2^{n_c} have a one at the position of each measured qubit in SS and zeros elsewhere, including the ancillas. The observable ZSZ_S maps the basis state specified by b\mathbf{b} to the one specified by b⊕s\mathbf{b} \oplus\mathbf{s}. Evolve a second vector using bitstring b⊕s\mathbf{b} \oplus\mathbf{s}, the same basis BB and the same angles. Pairing entries with the same t\mathbf{t} gives

GS(b;θ)=Re⁡∑t∈F2rwθ(b⊕s)(t)∗wθ(b)(t),G_S(\mathbf{b};\theta)=\operatorname{Re}\sum_{\mathbf{t}\in\mathbb{F}_2^r}w_{\theta}^{(\mathbf{b}\oplus\mathbf{s})}(\mathbf{t})^*w_{\theta}^{(\mathbf{b})}(\mathbf{t}),
c^S=1NMC∑m=1NMCGS(bm;θ).(12)\widehat{c}_S=\frac{1}{N_{\mathrm{MC}}}\sum_{m=1}^{N_{\mathrm{MC}}}G_S(\mathbf{b}_m;\theta). \tag*{(12)}

Draw the NMCN_{\mathrm{MC}} bitstrings independently and uniformly from F2nc\mathbb{F}_2^{n_c}. Zero input has equal weight in every coset, and every coset contains the same number of bitstrings. Averaging GSG_S therefore gives the desired coefficient. Since it is the real overlap of two normalized vectors, ∣GS∣≤1\lvert G_S\rvert\le1, and for fixed angles

Ec^S=cS,\mathbb{E}\widehat{c}_S=c_S,
Var⁡(c^S)=Var⁡(GS)NMC≤1NMC.(13)\operatorname{Var}(\widehat{c}_S)=\frac{\operatorname{Var}(G_S)}{N_{\mathrm{MC}}}\le\frac{1}{N_{\mathrm{MC}}}. \tag*{(13)}

These are additive-error guarantees for the coefficient. They do not make a nonlinear loss of c^S\widehat{c}_S unbiased.

Proposition (informal). For zero input and fixed circuit angles, Eq. (12) gives unbiased Fourier coefficient estimates, including for noncommuting circuits. With NMCN_{\mathrm{MC}} independent uniform draws, the root mean squared sampling error is at most NMC−1/2N_{\mathrm{MC}}^{-1/2}, independently of the number of qubits. Each draw evolves two vectors with 2r2^r amplitudes, with a cost of O(d2r)O(d2^r) for the dd gate updates after preprocessing. We restrict the gates to keep these vectors small, allowing us to estimate the correlations needed for training without storing the full quantum state. Proposition 1 in Appendix C gives the full error and computational bounds.

GHZ input. For GHZ input, the XX-basis amplitudes are uniform on states specified by even-parity bitstrings, meaning bitstrings with an even number of ones, so 1⋅b=0\mathbf{1}\cdot\mathbf{b}=0, where 1\mathbf{1} is the all-ones vector. If every gate vector zj\mathbf{z}_j has even parity, the circuit preserves this support. For even ∣S∣\lvert S\rvert, sample bitstrings uniformly from this support and use the same estimator. For odd ∣S∣\lvert S\rvert, the coefficient is exactly zero. Corollary 1 proves that the same estimation bounds apply under this condition, which holds for our Haldane and IQP gates. Other input states or parity-changing gates require a separate treatment. For Haldane gates, restricting the active Y/ZY/Z sites lowers rr but also introduces additional conserved XX products. The resulting expressivity change is tested in Figure 6.

Figure 6. Effects of restricted training for the nine-site Haldane target. All models are Haldane QCBMs. (a.i,ii) Susceptibility maps after likelihood training on 10 × 10 and 7 × 7 grids, with the colour scale of Figure 3. (b) Fixed-grid MMD with exact or MC coefficients through order four. (c.i,ii) Fixed and adaptive 49-point fits using CDF-KL and MMD, respectively, with the same order limit and MC budget. (d) Adaptive CDF-KL with restricted gates giving r=2r = 2. Turquoise curves mark adaptive learned peaks, gray curves exact finite-size peaks, gray points the nine initial references, and orange points the 40 acquired references. Panels retain 298 circuit angle parameters and 9,581 hypernetwork weights. All plotted responses are evaluated from full model probabilities.

(i) Recovering the IQP estimator. For Pj=XxjP_j=X^{\mathbf{x}_j}, every zj\mathbf{z}_j is zero and each vector has a single entry. Writing Φθ(b)=∑jθj(−1)xj⋅b\Phi_{\theta}(\mathbf{b})=\sum_j\theta_j(-1)^{\mathbf{x}_j\cdot\mathbf{b}} gives wθ(b)=e−iΦθ(b)w_{\theta}^{(\mathbf{b})}=e^{-i\Phi_{\theta}(\mathbf{b})}. The overlap in Eq. (12) becomes

c^SIQP=1NMC∑m=1NMCcos⁡[2∑j:xj⋅s=1θj(−1)xj⋅bm].(14)\widehat{c}_S^{\mathrm{IQP}}=\frac{1}{N_{\mathrm{MC}}}\sum_{m=1}^{N_{\mathrm{MC}}}\cos\left[2\sum_{j:\mathbf{x}_j\cdot\mathbf{s}=1}\theta_j(-1)^{\mathbf{x}_j\cdot\mathbf{b}_m}\right]. \tag*{(14)}

The general overlap estimator thus reduces to the IQP phase estimator. For zero input, sample b\mathbf{b} uniformly from all ncn_c-bit strings. For GHZ input and even-order coefficients, sample uniformly from bitstrings with an even number of ones. Appendix C.3 derives the phase cancellation leading to Eq. (14), and Appendix C.4 derives the GHZ sampling rule.

(ii) Cost of the full Haldane gate set. We now determine rr for the Haldane gate set in which the allowed Y/ZY/Z pairs connect all qubits within each circuit sublattice. This gives the number of amplitudes, 2r2^r, that each vector must contain.

Number the circuit qubits from 11 to ncn_c, including the ancillas at their assigned positions. The even and odd positions define the two sublattices, E={j∈{1,…,nc}:j even}E=\{j\in\{1,\ldots,n_c\}:j\text{ even}\} and O={j∈{1,…,nc}:j odd}O=\{j\in\{1,\ldots,n_c\}:j\text{ odd}\}. These sets refer to the circuit qubit register, including ancillas. A gate with YY or ZZ on two sites in the same sublattice flips those two bits of the XX-basis label. It therefore preserves the parity of that label on each sublattice. Any XX factors change phases without flipping bits.

When the permitted pairs connect each sublattice, possibly through intermediate sites, their binary sums generate every flip pattern with even parity on both sublattices. The span is therefore

VH={z:z⋅1E=z⋅1O=0},r=nc−2.(15)V_H=\{\mathbf{z}:\mathbf{z}\cdot\mathbf{1}_E=\mathbf{z}\cdot\mathbf{1}_O=0\},\qquad r=n_c-2. \tag*{(15)}

Here 1E\mathbf{1}_E and 1O\mathbf{1}_O have ones on their respective sublattices and zeros elsewhere. Each nonempty sublattice imposes one independent parity constraint, leaving nc−2n_c-2 independent binary directions. Appendix C.5 proves this by constructing a basis within each connected sublattice. In the twelve-qubit circuit of Figure 3, this gives r=10r=10.

Figure 3. Learned distribution susceptibility for the nine-site open Haldane chain. Colours show χ22\chi_{22} from Eq. (26), the response to h2/Jh_{2}/J at fixed h1/Jh_{1}/J. Panels compare (a) the exact target with (b) the Haldane QCBM, (c) RBM, (d) IQP, (e) Heisenberg QCBM and (f) Fermionic QCBM. All panels use a common colour scale and evaluation grid. The inset shows the 10×1010 \times10 training grid, coloured by target susceptibility. Each model uses 5,000 samples per training point and one initialization. Responses are calculated from all outcome probabilities after training.

One Monte Carlo sample is still the real overlap of two normalized vectors, so the variance bound is unchanged. However, each vector now contains 2nc−22^{n_c-2} amplitudes. The cost of evaluating one sample with this method therefore grows exponentially with ncn_c. The restriction below reduces this cost by limiting the sites on which Y/ZY/Z pairs can act.

(iii) Restricting the active Haldane sites. We select nonempty qubit sets AE⊆EA_E\subseteq E and AO⊆OA_O\subseteq O within the two circuit sublattices (even and odd positions, respectively). We permit Y/ZY/Z pair gates only within these selected sets, while retaining XX-type gates throughout the circuit. We call the selected qubits active sites. Connecting each selected set through the permitted pair gates gives

r=∣AE∣+∣AO∣−2.(16)r = |A_E| + |A_O| - 2. \tag*{(16)}

Two active sites per sublattice give r=2r = 2 and four entries per vector. Three give r=4r = 4 and sixteen entries. Choosing O(log⁡nc)O(\log n_c) active sites gives polynomial evaluation cost for a polynomial gate count. The gate restriction also changes the distributions the circuit can represent. Appendix C5 derives Eq. (16), including the case of disconnected selected sets.

To see how these updates work, consider XX-type gates together with a single ZuZvZ_uZ_v generator, where uu and vv belong to the same sublattice. The only possible flip changes both bits uu and vv, so r=1r = 1 and each vector has two entries. For zero input, we initialize w=(1,1)T/2w = (1,1)^T/\sqrt{2}. A rotation e−iαZuZve^{-i\alpha Z_uZ_v} then acts as

w⟵(cos⁡α−isin⁡α−isin⁡αcos⁡α)w.(17)w \longleftarrow \begin{pmatrix} \cos\alpha& -i\sin\alpha\\ -i\sin\alpha& \cos\alpha \end{pmatrix}w. \tag*{(17)}

Here the basis matrix has one column, B=z=eu⊕evB = z = e_u \oplus e_v, where eue_u is one at site uu and zero elsewhere. An XjxjX_j^{x_j} rotation multiplies entry t∈{0,1}t \in\{0,1\} by e−iθj(−1)xj⋅(b⊕tz)e^{-i\theta_j(-1)^{x_j\cdot(b\oplus tz)}}. After these updates in circuit order, one MC sample is

GS(b)=Re⁡[w(b⊕s)(0)∗w(b)(0)+w(b⊕s)(1)∗w(b)(1)].(18)G_S(\mathbf b) = \operatorname{Re}\left[w^{(\mathbf b\oplus\mathbf s)}(0)^*w^{(\mathbf b)}(0) + w^{(\mathbf b\oplus\mathbf s)}(1)^*w^{(\mathbf b)}(1)\right]. \tag*{(18)}

Thus the IQP scalar phase is replaced by a two-entry calculation in this noncommuting example. Appendix C6 derives both update matrices and works through two- and three-gate examples using the overlap estimator and Pauli conjugation.

CDF-KL on selected marginals

CDF-KL extends the KL-based training formulation to selected correlations by comparing cumulative probabilities of small marginals. It provides an alternative to MMD when the full likelihood is unavailable. A set of coefficients can define a small marginal distribution without defining the full nn-bit distribution. For A⊂[n]A \subset[n] with ∣A∣=ℓ|A| = \ell, all coefficients supported in AA determine

pA(z)=2−ℓ∑S⊆AcS(p)(−1)∑i∈Szi.(19)p_A(z) = 2^{-\ell}\sum_{S\subseteq A}c_S(p)(-1)^{\sum_{i\in S}z_i}. \tag*{(19)}

We select a collection A\mathcal{A} of such subsets. We need every coefficient within each selected marginal, but need not use every ℓ\ell-site subset of the system.

For each marginal, order its bit strings by their fixed binary integer value. Let Fp,A(t)=∑z≤tpA(z)F_{p,A}(t) = \sum_{z\leq t}p_A(z), with t=0,…,2ℓ−2t = 0,\ldots,2^\ell-2. We compare these cumulative probabilities using

dB(u,v):=ulog⁡uv+(1−u)log⁡1−u1−v,(20)d_B(u,v) := u\log\frac{u}{v} + (1-u)\log\frac{1-u}{1-v}, \tag*{(20)}
LCDF:=∑A∈A∑t=02ℓ−2dB(Fp,A(t),Fq,A(t))∣A∣(2ℓ−1).(21)\mathcal{L}_{\mathrm{CDF}} := \frac{\sum_{A\in\mathcal{A}}\sum_{t=0}^{2^\ell-2}d_B\left(F_{p,A}(t),F_{q,A}(t)\right)}{|\mathcal{A}|(2^\ell-1)}. \tag*{(21)}

Choosing the sizes and collection of marginals indirectly controls the number of distinct coefficients that must be estimated, accounting for overlap between marginals and coefficients fixed by symmetry. In our restricted-MMD comparisons, MMD uses this same fixed set, with weights determined by its bandwidth σ\sigma. In Gaussian-sampled MMD, σ\sigma instead determines the probabilities of drawing coefficient supports. We can match the number of coefficient estimates in that comparison, but the sampled sets and orders need not coincide.

This is an average of Bernoulli KL divergences for threshold events. It changes which probabilities are compared, rather than approximating the full-distribution KL by an unbiased estimator. It is also distinct from the KL divergence of the marginal probability vectors. The bit ordering is part of its definition. We average this marginal loss across reference points to update the hypernetwork and base circuit angles, as in the likelihood formulation.

Noisy coefficients can reconstruct negative probabilities. Before forming each CDF, we project its reconstructed vector onto the probability simplex and mix in a small uniform component, applying the same procedure to the target marginal. This uniform mixing follows label smoothing [43], applied here to both marginals before calculating CDF-KL. Here it keeps every compared cumulative probability strictly between zero and one, making the CDF-KL terms and their derivatives with respect to those probabilities finite. This allows gradient-based training without logarithmic singularities at zero-probability events. Smoothing changes the loss, and simplex projection remains only piecewise differentiable. An unbiased coefficient estimator therefore does not give an unbiased CDF-KL loss or gradient. We compare exact and MC coefficient training directly. Appendix D states the projection, smoothing and effect of retaining only selected marginals.

MMD and a combined objective

We compare CDF-KL with MMD, which compares distributions through a positive definite kernel [27]. Both losses can use the same selected coefficients, but compare different functions of them. For the Gaussian kernel on bit strings, kσ(x,y)=exp⁡[−∥x−y∥22/(2σ2)]k_\sigma(\mathbf{x},\mathbf{y})=\exp[-\lVert\mathbf{x}-\mathbf{y}\rVert_2^2/(2\sigma^2)], its squared value is

MMD⁡2(p,q)=∑S⊆[n]wσ(S)[cS(p)−cS(q)]2.(22)\operatorname{MMD}^2(p,q)=\sum_{S\subseteq[n]}w_\sigma(S)[c_S(p)-c_S(q)]^2. \tag*{(22)}

where wσ(S)=ρ∣S∣(1−ρ)n−∣S∣w_\sigma(S)=\rho^{|S|}(1-\rho)^{n-|S|} and ρ=(1−e−1/(2σ2))/2\rho=(1-e^{-1/(2\sigma^2)})/2. The bandwidth therefore sets a distribution over coefficient orders. This representation is used in classical IQP optimization [9, 10].

The kernel weights define the probabilities of choosing the supports. Larger σ\sigma favors smaller supports and hence lower-order coefficients. Smaller σ\sigma favors larger supports. In particular, the mean support size is ES∼wσ∣S∣=nρ\mathbb{E}_{S\sim w_\sigma}|S|=n\rho. This bandwidth choice also affects trainability. For product circuits and suitable shallow circuits, choosing σ2=Θ(n)\sigma^2 = \Theta(n) avoids exponential loss concentration under the circuit and data assumptions of the corresponding trainability results [44]. Their kernel uses 2σ2\sigma in the denominator, so their linear bandwidth scaling corresponds to σ=Θ(n)\sigma= \Theta(\sqrt{n}) in our convention. This is a sufficient regime under those assumptions, rather than a universal requirement for successful training.

Two independent MC estimates give an unbiased product (c^S(1)−cS(p))(c^S(2)−cS(p))(\widehat{c}_{S}^{(1)}-c_{S}(p))(\widehat{c}_{S}^{(2)}-c_{S}(p)) for a fixed empirical target coefficient. Squaring one noisy estimate would add its variance. In the matched comparisons, the two factors split the stated MC budget equally.

We also train with the combined loss LCDF+αMMDLMMD\mathcal{L}_{\mathrm{CDF}}+\alpha_{\mathrm{MMD}}\mathcal{L}_{\mathrm{MMD}}, where αMMD≥0\alpha_{\mathrm{MMD}}\geq0 sets the weight of the MMD term relative to CDF-KL. MMD weights squared coefficient differences through its Gaussian kernel, while CDF-KL compares cumulative probabilities within separately chosen, equally weighted marginals. Appendix D 2 gives the kernel expansion and MC estimators, and Appendix G reports the additional fixed-grid comparisons.

Distribution learning and phase signatures

We first test the distributions represented by the classically trained circuits where exact small-system evaluation is available. The comparisons address sampling accuracy, the effect of the circuit architecture, and the recovery of changes across the phase diagram. The small system sizes nn considered in this section allow us to evaluate these quantities by summing over all 2n2^n measurement bitstrings.

Evaluation metrics and response functions

For the small systems, we calculate distribution infidelity and total variation distance (TVD) [45] from all outcomes,

I(p,q)=1−(∑xp(x)q(x))2(23)\mathcal{I}(p,q)=1-\left(\sum_x\sqrt{p(x)q(x)}\right)^2 \tag*{(23)}
TVD(p,q)=12∑x∣p(x)−q(x)∣.(24)\mathrm{TVD}(p,q)=\frac{1}{2}\sum_x\lvert p(x)-q(x)\rvert. \tag*{(24)}

They obey [46]

1−1−I≤TVD≤I.(25)1-\sqrt{1-\mathcal{I}}\leq\mathrm{TVD}\leq\sqrt{\mathcal{I}}. \tag*{(25)}

We use infidelity as our main measure of distribution error because its local variation defines the distribution susceptibility used below.

For a differentiable distribution family we define

χμν(qλ)=14∑x:qλ(x)>0∂μqλ(x)∂νqλ(x)qλ(x).(26)\chi_{\mu\nu}(q_\lambda)=\frac{1}{4}\sum_{x:q_\lambda(x)>0}\frac{\partial_\mu q_\lambda(x)\partial_\nu q_\lambda(x)}{q_\lambda(x)}. \tag*{(26)}

It is one quarter of the classical Fisher information matrix. Locally, I(qλ,qλ+dλ)=∑μνχμνdλμdλν+o(∥dλ∥2)\mathcal{I}(q_\lambda,q_{\lambda+d\lambda})=\sum_{\mu\nu}\chi_{\mu\nu}d\lambda_\mu d\lambda_\nu+o(\lVert d\lambda\rVert^2) under regularity of the support (Appendix A 2). Fidelity responses have long been used to study phase transitions [47, 48]. Here the response concerns a specified measurement distribution, not the full quantum-state fidelity. Derivatives of the learned model hold its trained weights fixed.

Phase transitions are commonly investigated through order parameters, their derivatives and fluctuations, and correlation functions. Magnetic susceptibility and heat capacity are familiar response indicators. Fidelity susceptibility and Fisher information characterize changes in states or measurement distributions without specifying an order parameter [48, 49]. Data-driven approaches also construct indicators from sampled configurations, using principal-component analysis [50], supervised classification and learning by confusion [51, 52], or derivatives of inferred physical parameters [53]. Changes in mean input features provide another interpretable option [54]. Analytical studies express optimal classification and regression indicators in terms of the underlying distributions [55], extend this approach to generative classifiers [56], and bound several normalized indicators by the square root of the classical Fisher information [57]. ClassiFIM further estimates the Fisher information from sampled configurations through an auxiliary classification task [58]. These connections place the distribution susceptibility in Eq. (26) within a broader set of phase-transition indicators.

A phase indicator calculated from bitstrings can probe correlations beyond those available to the classical training procedure [19]. If the indicator and its parameter response depend only on classically accessible correlators, they can be evaluated directly. For indicators estimated from samples, a comparison of classical and quantum approaches would need to account for the cost of generating those samples as well as training the model and evaluating the indicator.

A one-parameter TFIM benchmark

Figure 2 tests whether classical training across the field of the nine-site periodic TFIM supplies accurate sampling circuits between reference points. Nine training fields span g/J∈[1,4.5]g/J\in[1,4.5], with 5,000 samples per field and eight separate validation fields. We train for 1,500 likelihood updates and select checkpoints by empirical validation NLL.

Figure 2. Distribution learning across the field of the periodic 3×33 \times3 TFIM. (a) Distribution infidelity I(p,q)\mathcal{I}(p,q), (b) magnetic response and (c) distribution susceptibility. Models include hypernetwork IQP with and without three unmeasured ancillas, the adapted RBM, and IQP with conditional rotations with and without a hypernetwork. Black curves show exact-target responses, and the dashed line marks the finite-size magnetic-response peak. Circles and open diamonds mark training and validation fields. Curves and bands show means and one sample standard deviation across three initializations with shared reference data. Appendix E specifies the IQP generators, GHZ preparation, RyR_y encoding and RBM conditioning.

Single-site base angles are drawn uniformly from [0.2,0.35][0.2,0.35] and pair angles from a zero-mean Gaussian with standard deviation 0.03. The hypernetwork initially introduces only small angle modulations. Initialization and circuit architecture both influence the trainability of IQP generative models [17, 18]. Appendix § 2.1.4 2 gives the likelihood-training settings.

We compare hypernetwork IQP with a restricted Boltzmann machine (RBM) adapted to the hypernetwork setting, extending earlier RBM–Born-machine comparisons [59]. Over the 144 test fields and three initializations, their mean distribution infidelities are similar, at 0.0115 and 0.0106, respectively. Three unmeasured ancillas reduce the IQP infidelity to 0.00263, a reduction by a factor of approximately 4 relative to IQP without ancillas and the RBM. The circuit angle count increases from 45 to 78. Appendix E specifies the circuit preparations.

The hypernetwork learns field-dependent IQP angles through a nonlinear map. We also test conditional input rotations, as in Born-machine extensions such as [26], using a fixed affine field encoding through RyR_y on each qubit. The IQP angles are either shared across fields or supplied by a hypernetwork, giving mean test infidelities of 0.206 and 0.0514, respectively. The hypernetwork therefore improves the rotation model, but neither variant improves on hypernetwork IQP without input rotations. This supports learning the Hamiltonian-to-circuit parameter map, which need not be linear. The comparison does not isolate nonlinearity, since the input rotations also change the circuit and can break its global-flip symmetry.

To test whether the model reproduces changes between training fields, we use two types of indicators discussed in § 2 A, an order-parameter response and the distribution susceptibility.

The magnetic response ∣∂g/JEq∣mz∣∣|\partial_{g/J}\mathbb{E}_q|m_z||, where mz=n−1∑i(−1)xim_z = n^{-1}\sum_i(-1)^{x_i}, tracks changes in magnetization as the transverse field suppresses magnetic order. Taking ∣mz∣|m_z| avoids cancellation between opposite magnetizations. The distribution susceptibility probes changes in all outcome probabilities, including those that the mean magnetization magnitude may miss. Low infidelity alone does not ensure accurate responses, so we compare both curves and their peaks with those of the exact nine-site system.

Learning the Haldane phase diagram

Near the Haldane symmetry-protected topological phase boundary, the correlations used to identify the phase become weak, so more measurements are needed to distinguish them from sampling noise [29]. Finite system sizes also smooth the transition [60].

Figure 3 extends the task to two Hamiltonian parameters and compares the inductive biases of different generator families under the same reference data. The nine-site target uses h1/J∈[0,1.6]h_1/J \in[0,1.6] and h2/J∈[−1.6,1.6]h_2/J \in[-1.6,1.6]. A fixed 10×1010 \times10 grid supplies 100 reference datasets, with 5,000 samples each. We use 24 separate validation points and evaluate on a held-out 40×4040 \times40 grid. This comparison

uses one initialization and 1,500 likelihood updates per model.

We use the distribution susceptibility χ22\chi_{22} of Eq. (26), evaluated with respect to h2/Jh_2/J at fixed h1/Jh_1/J. Comparing the learned and exact-target maps tests whether the model reproduces the finite-chain phase-transition signatures.

The circuits act on nine measured qubits and three interleaved ancillas. Each has 298 circuit angle parameters, with generators and repetitions chosen as in Appendix B. We count one angle per parametrized gate, including repeated generators and gates whose angles do not affect the distribution of measured outcomes. Table I separates these angles from hypernetwork weights. Circuits with the same number of angle parameters can represent different sets of distributions. The RBM has approximately matched generator and hypernetwork budgets.

QCBMsRBM
Generator parameters298299
Hypernetwork parameters9,5819,714
Total9,87910,013

Table I. Parameter counts for Figure 3, including all circuit angle parameters and hypernetwork weights. Circuit families share the displayed counts. The RBM generator count includes its interaction matrix and biases.

The Haldane QCBM reproduces the two prominent response bands of the exact target in Figure 3(a,Figure 3b), although their positions and widths do not agree everywhere. The other models also recover broad response features, with differences in their location and contrast. These maps assess the parameter dependence of the learned distributions, beyond their pointwise accuracy.

Table II shows that the Haldane QCBM gives the lowest mean infidelity and TVD, 0.00670 and 0.0436, respectively. The Heisenberg QCBM is next in both metrics, with 0.0126 and 0.0624, followed by IQP, the Fermionic QCBM and the RBM. Thus equal circuit angle counts do not give equal accuracy in this comparison. These observations concern one fit per model with the stated budgets. Additional target Hamiltonians are considered in Appendix G.

ModelMean I(p,q)\mathcal{I}(p,q)Mean TVD
Haldane QCBM0.006700.0436
RBM0.04470.124
IQP0.01820.0682
Heisenberg QCBM0.01260.0624
Fermionic QCBM0.03290.0929

Table II. Mean distribution errors over the held-out Haldane grid in Figure 3. Values refer to one fit per model.

Fourier coefficients by correlation order

Distribution error averages over all outcomes. The Fourier spectrum gives a different view by separating correlations by order. At a fixed parameter point, we calculate the root mean square (RMS) coefficient magnitude

E∣S∣=k[cS2]=[(nk)−1∑∣S∣=kcS2]1/2(27)\sqrt{\mathbb{E}_{\lvert S\rvert=k}[c_S^2]} = \left[\binom{n}{k}^{-1}\sum_{\lvert S\rvert=k} c_S^2\right]^{1/2} \tag*{(27)}

The average is uniform over subsets of size kk. It is not an average over training repetitions. Here and in Figure 7, we calculate the model and target probabilities of every possible outcome to obtain the evaluation coefficients. The plotted RMS and squared errors therefore contain no MC evaluation noise, even when the model was trained with MC estimates.

Figure 7. Shared training and local refinement for the twelve-site Haldane model with three ancillas. (a.i) TVD during independent training, angle-only fine-tuning and hypernetwork fine-tuning. The guides mark when hypernetwork fine-tuning reaches the independent fits’ final mean TVD. (a.ii) Final squared MMD across Gaussian bandwidths, evaluated using all coefficient orders. (b) Fourier RMS and (c) coefficient MSE, defined in Eqs. (32)–(33). Upper panels compare adaptive pretraining with data-informed initialization, and lower panels compare hypernetwork fine-tuning with trained independent models. The separator distinguishes orders used in training (k≤4k \le4, Seen) from higher orders (Unseen). Local fits use the same samples and 1,500 CDF-KL plus MMD updates, excluding pretraining. Curves show means and bands one sample standard deviation across four initializations, with 40 common boundary points averaged within each initialization. Evaluation uses exact model and target probabilities.

Figure 4 compares the Haldane model with its target at four points on h1/J=0.8h_1/J=0.8. None belongs to the fixed training grid. This is a test of interpolation after likelihood training. The selected-coefficient training comparison follows in Sec. VI. Only even orders are shown because the preparation and circuit enforce bit-complement symmetry. The spectrum reveals where the learned correlation scale departs from the target, including at high orders with small exact coefficients. Different sets of coefficients can have the same RMS.

The closest relative agreement across the displayed orders occurs at h2/J=−0.5h_2/J=-0.5, the point with the highest target susceptibility in Figure 4 (χ22≃0.992\chi_{22}\simeq0.992). Here the learned RMS follows the target through order eight. At the lower-susceptibility points h2/J=0.5h_2/J=0.5 and 1.51.5, the model overestimates the RMS at orders six and eight, despite close agreement at lower orders.

Figure 4. Fourier RMS spectra at four untrained points. Solid curves with circular markers show the Haldane QCBM from Figure 3(b). Dashed curves show the exact targets. Both spectra show the RMS defined in Eq. (27), using exact probabilities.

Locating rapid changes in the learned family

Spin correlators can provide interpretable signatures of phase structure. For example, TetriSCNN detects phases from measurement snapshots and expresses order parameters in terms of spin correlators [6]. Here, we use the response of the learned distribution to locate candidate sampling points near phase boundaries, without identifying an order parameter.

To choose where to sample near a transition, we first need to locate rapid changes in the learned distributions. Learning a family gives a differentiable response throughout the diagram. We identify candidate boundaries by locating interior local maxima of χ22\chi_{22} along vertical slices. Adjacent maxima are linked only when their separation is within the grid tolerance specified in Appendix A.3, to avoid joining distant peaks into an artificial boundary. This procedure uses no phase labels and does not force the model peaks onto a reference curve. Figure 5 applies the procedure to the Haldane model already fitted in Figure 3(b). The reconstructed boundaries lie close to the exact nine-site susceptibility peaks along both branches. Published bulk boundaries in [29] supply a separate physical guide. Those curves were obtained numerically from the curvature of the infinite-chain energy density,

e∞(u,v)=lim⁡n→∞E0(n)(u,v)n,(28)e_{\infty}(u,v)=\lim_{n\to\infty}\frac{E_{0}^{(n)}(u,v)}{n}, \tag*{(28)}

where u=h1/Ju=h_1/J and v=h2/Jv=h_2/J. The boundaries are identified by singular or peaked ∣∂v2e∞∣\lvert\partial_v^2 e_{\infty}\rvert along a slice. A finite-size probability-response peak and a bulk energy-curvature boundary are different reference quantities.

Effects of restricting classical training

We now test whether the response can still be recovered when classical training has fewer reference states and access to only selected correlations. Figure 6 first reduces the likelihood-training grid from 10×1010 \times10 to 7×77 \times7. This retains full model probabilities and isolates the smaller reference grid within that comparison. Mean held-out infidelity increases from 0.00670 to 0.0175.

We then train on coefficients through order four. CDF-KL uses all 126 four-bit marginals. Their union contains 255 nonconstant coefficients. Of these, 162 even-order model coefficients require evaluation, while odd coefficients vanish by symmetry. The order-four MMD comparison uses the same coefficient information with its kernel weights. MC versions use 64 draws per nontrivial coefficient per update. All susceptibilities in Figure 6 are evaluated from exact model probabilities, including for models trained with MC estimates.

The fixed-grid CDF-KL infidelity is 0.0743 with exact coefficients and 0.0751 with MC estimates. The corresponding MMD values are 0.0497 and 0.0496. Under these settings, changing the coefficient estimator has a smaller effect on mean infidelity than replacing full likelihood by a restricted loss. The final panel further restricts the allowed gates to give r=2r=2, so each simulated vector has four amplitudes. Its loss still uses order-four coefficients, but the gate restriction changes the set of distributions the circuit can represent.

We describe how the adaptive fits select reference points in Sec. V and apply the procedure to larger systems in Sec. VII. Appendix G gives the full error table and compares gate restrictions.

Adaptive reference acquisition

Choosing where to calculate the next ground state

Each new reference point requires a separate ground-state calculation. For large systems, this cost limits how many distributions can contribute to classical training. Our purpose is to select a sparse set of useful references, with particular interest in regions near phase transitions. A dense grid can spend most of its cost in regions where the distribution changes slowly. Earlier machine-learning studies locate phase changes from labelled or unlabelled data [50, 51, 52]. Latent representations can expose phase structure [62], and symbolic methods can connect these representations to physical observables [63]. Autonomous acquisition can use latent representations to guide differentiable ground-state calculations [28]. We adapt the latent-distance search to use derivatives of our learned correlators, avoiding differentiation through the reference solver.

Let yp(λ)y_p(\lambda) contain a fixed set of target coefficients at acquired points, standardized using those data. We fit an autoencoder with encoder eψe_\psi and decoder dψd_\psi by reconstruction error. We encode the coefficients at each acquired reference point and take their average in latent space,

μt=1∣Dt∣∑λj∈Dteψ(yp(λj)).(29)\mu_t=\frac{1}{\lvert\mathcal{D}_t\rvert}\sum_{\lambda_j\in\mathcal{D}_t}e_\psi\left(y_p(\lambda_j)\right). \tag*{(29)}

For a candidate point, the trained circuit supplies yq(λ)y_q(\lambda). We define

At(λ)=−∥eψ(yq(λ))−μt∥22+δ2dz,(30)A_t(\lambda)=-\frac{\sqrt{\left\lVert e_\psi\left(y_q(\lambda)\right)-\mu_t\right\rVert_2^2+\delta^2}}{\sqrt{d_z}}, \tag*{(30)}

where dzd_z is the latent dimension and δ>0\delta>0 regularizes the norm. After rescaling each Hamiltonian parameter to the interval [0,1][0, 1], we update the resulting vector uu according to

us+1=Π[0,1]dλ[us−α∇uAt(us)∥∇uAt(us)∥2+ϵ].(31)u_{s+1} = \Pi_{[0,1]^{d_\lambda}}\left[u_s - \alpha\frac{\nabla_u A_t(u_s)}{\lVert\nabla_u A_t(u_s)\rVert_2 + \epsilon}\right]. \tag*{(31)}

Here, Π[0,1]dλ\Pi_{[0,1]^{d_\lambda}} projects onto the normalized parameter domain by clipping each coordinate to [0,1][0, 1]. The search moves the encoded model prediction away from this average. We also sample less explored regions and require new points to be sufficiently far from existing ones, so the search does not concentrate on a single region. We then calculate a reference state at the accepted point, add its data, and refit the shared model. Appendix § 2.1.4 3 gives the search and training settings. In the nine-site fits, some acquired points cluster along branches of the learned response in Figure 6. With 49 references, the adaptive CDF-KL and MMD fits have mean infidelities of 0.0819 and 0.0655 over the full grid, compared with 0.0751 and 0.0496 for their fixed-grid counterparts. These averages describe accuracy across the diagram. The selected coordinates show where the search concentrates the reference calculations.

Sharing information and refining boundary samplers

A common set of boundary targets

The low-order loss does not directly compare higher-order coefficients with their targets. These coefficients still depend on the shared parameters, generator choice and preparation. We ask whether fitting related distributions across the diagram helps recover them at a selected boundary point, compared with fitting that point independently.

uses Figure 7 a twelve-site Haldane target, three ancillas and an r=4r = 4 circuit with 316 circuit angle parameters. Fixed-grid and adaptive hypernetworks each train on 49 reference points. The loss uses all 495 four-bit marginals, requiring 561 nonzero even-order coefficients estimated with 64 MC draws each. No coefficient above order four enters this loss.

We freeze 40 coordinates on the adaptive model’s learned susceptibility boundary. At every one of these same coordinates we compute a new target and evaluate both pretrained hypernetworks. We also train an independent QCBM directly on its local data, with the same circuit, local samples and 1,500 updates as fine-tuning. Pretraining uses CDF-KL, while all local fits use CDF-KL plus MMD restricted to orders through four. The pretrained hypernetworks use data from other Hamiltonian points, while the independent models are trained at the selected boundary points.

For each model and initialization, we calculate the RMS coefficient magnitude and mean squared error (MSE) of the coefficients,

RMSk=Eλ∈BE∣S∣=k[cS(qλ)2],(32)\mathrm{RMS}_k = \sqrt{\mathbb{E}_{\lambda\in\mathcal{B}}\mathbb{E}_{|S|=k}\left[c_S(q_\lambda)^2\right],} \tag*{(32)}
MSEk=Eλ∈BE∣S∣=k[(cS(qλ)−cS(pλ))2].(33)\mathrm{MSE}_k = \mathbb{E}_{\lambda\in\mathcal{B}}\mathbb{E}_{|S|=k}\left[(c_S(q_\lambda)-c_S(p_\lambda))^2\right]. \tag*{(33)}

Both averages are uniform. The squared error is calculated before averaging, so it retains coefficient-wise differences that an RMS spectrum can hide. The small-system evaluation uses every coefficient of each displayed order, including those absent from training. Across initializations, we keep the training and validation data, acquisition points and order, MC draws and boundary points fixed. Figure E2 in Appendix § 2.1.4 shows the susceptibility maps, learned boundary points and pretraining loss curves for the four initializations.

Figure E2. Pretraining the twelve-site models used in Figure 7. Susceptibility χ22\chi_{22} after (a) adaptive and (b) fixed-7 × 7 training for four initializations, labeled i–iv. Orange circles mark learned response peaks and dashed curves show the bulk guides. (c) CDF-KL training loss averaged over the available references. Thin curves show individual initializations and thick curves their mean. Adaptive training grows from nine to 49 references, while fixed-grid training uses all 49 throughout.

Before local fitting, the pretrained model already captures much of the target’s low-order correlation strength in Figure 7(b.Figure 7i). At orders two and four, its gap from the target RMS is approximately ten and three times smaller, respectively, than that of the data-informed initialization. The initialization gives nearly zero RMS above order two, while pretraining produces nonzero higher-order correlations.

We evaluate coefficient errors above order four to test how well the learned circuit reproduces correlations outside the training loss. Sample-based coverage tests measure a complementary quantity, the fraction of previously unseen valid outcomes reached within a given sampling budget [64, 65, 66].

Local fine-tuning from the learned family

Interpolation supplies a circuit without a new local fit. When reference data at the selected point are available, fine-tuning provides a second route to preparing it for quantum deployment. For each boundary coordinate, we make a fresh copy of the pretrained hypernetwork and base angles and optimize both on the local data. All copies within a variant start from the same pretrained parameters and fresh optimizer moments. No parameters pass between boundary points. Fine-tuning and independent training use 1,500 local updates with CDF-KL plus MMD and the same local samples. The learning rate decays linearly from 0.003 to 0.00003. Angle-only fine-tuning starts from the pretrained circuit’s 316 output angles and optimizes them directly. Hypernetwork fine-tuning instead updates the hypernetwork and shared base angles, a total of 10,473 parameters. We perform this separately for the fixed-grid and adaptive models. Appendix § 2.1.4 gives the numerical settings.

Figure 7(a.Figure 7i) shows that adaptive hypernetwork fine-tuning improves on the final TVD of independent training. Across the four initializations, its final mean TVD is lower than independent training in each initialization. The improvement also appears in Figure 7(a.ii). Hypernetwork fine-tuning gives lower mean squared MMD than both independent training and angle-only fine-tuning across the displayed bandwidths. At σ=1\sigma= 1, the reduction relative to independent training is approximately 25%. This evaluation uses all coefficient orders, including those absent from the training loss.

Local refinement also brings the pretrained model’s higher-order RMS magnitudes closer to the target in Figure 7(b.ii), even though these orders do not enter the loss. In Figure 7(c.ii), hypernetwork fine-tuning gives lower order-four coefficient error, while independent training gives lower order-eight error. Appendix § 2.1.4 gives a separate CDF-KL-only comparison.

The lower distribution errors and reduced number of local updates make shared pretraining a promising starting point for preparing samplers at several Hamiltonian points. A trained family can supply these initializations before local data are used to refine each circuit.

Larger systems

We now extend shared training and local refinement to a chain and a rectangular lattice with 1,024 measured qubits. Reference samples and classically estimated correlations allow us to train and compare these models without storing their full probability distributions.

Training a 1,024-site chain

The chain model uses a restricted Haldane circuit with 342 unmeasured ancillas and r=2r = 2 (Appendix F2). At each Hamiltonian point, the hypernetwork supplies the 2,739 circuit angles using 87,693 weights.

The loss combines CDF-KL on 32 selected six-site marginals with the full-system Gaussian MMD of Eq. (22), using sampled Fourier supports without an order cutoff. Exact gate cancellation further reduces the work needed for each coefficient while preserving the sampling circuit, as described in Appendix C7. Appendix § 2.1.5 gives the coefficient budgets and loss weights.

The fit uses nine initial reference points and 41 acquired points, with 20,000 reference samples per point and 1,500 updates in total. We evaluate the final checkpoint. The reference states are MPS approximations optimized with a bond-dimension cap of 128. MPS with limited bond dimension can retain useful phase features, as demonstrated for autonomous discovery in an XXZ chain [28], although the finite bond dimension can limit accuracy near transitions [67].

Learned response peaks and local refinement

We locate candidate phase boundaries using the mean susceptibility of the selected marginals,

χˉ22=14∣A∣∑A∈A∑z[∂2qA,ϵ(z)]2qA,ϵ(z)(34)\bar{\chi}_{22} = \frac{1}{4|\mathcal{A}|} \sum_{A \in\mathcal{A}} \sum_z \frac{\left[\partial_{2}q_{A,\epsilon}(\mathbf{z})\right]^2}{q_{A,\epsilon}(\mathbf{z})} \tag*{(34)}

where qA,ϵq_{A,\epsilon} denotes the projected, smoothed marginal reconstructed from MC coefficient estimates. The MC random inputs are held fixed across the evaluation grid before differentiation. This indicator measures changes in the selected model marginals as h2/Jh_2/J varies.

Figure 8(a) recovers portions of both bulk phase-boundary branches from 50 reference points. The learned peaks follow segments of the bulk guides, with gaps and additional peaks elsewhere. This agreement shows that selected marginals can retain phase-boundary signatures at this system size.

At the marked point (h1/J,h2/J)=(1.46,−1.40)(h_1/J,h_2/J)=(1.46,-1.40), Figure 8(b) compares the pretrained circuit, its locally fine-tuned copy and an independently trained circuit against the same MPS reference. For both geometries, fine-tuning and independent training use the same 20,000 local samples and 600 updates. Evaluation uses separate held-out samples, with the black curve comparing the training and held-out sets from the same reference MPS.

The pretrained chain model already gives lower mean squared MMD than independent training at each of the ten tested bandwidths from 5 to 64, and fine-tuning reduces it further. At σ=32\sigma=32, the model errors remain above the train–test comparison. Fine-tuning gives 1.49×10−51.49 \times10^{-5}, compared with 4.48×10−54.48 \times10^{-5} for independent training, an improvement by a factor of approximately three.

A 1,024-site rectangular lattice

We also train on a 4 by 256 spin lattice with open boundaries, using the same circuit and training schedule as the chain. The CDF-KL marginals are now six-site rectangles. The rectangular Hamiltonian and reference calculations are specified in Appendix § 2.1.5. Figure 9(a) shows several branches of peaks in the marginal susceptibility, identifying regions where the learned distributions change rapidly.

At the marked coordinate (h1/J,h2/J)=(1.22,−0.04)(h_1/J,h_2/J)=(1.22,-0.04), local refinement gives a larger improvement than for the chain. The fine-tuned model has the lowest mean squared MMD of the three models at all 11 displayed bandwidths. At σ=32\sigma=32, it reaches 5.36×10−65.36 \times10^{-6}, compared with 3.98×10−53.98 \times10^{-5} for independent training, a reduction by a factor of approximately seven. At this bandwidth, the fine-tuned error is also comparable in magnitude to the train–test comparison.

The pretrained lattice model has higher error than independent training at σ=32\sigma=32, yet becomes the more accurate model after fine-tuning. Shared pretraining therefore provides a useful starting point even when its direct prediction is less accurate. The improvements in both geometries extend the local-refinement results of Figure 7 to 1,024 measured qubits, using reference samples and classically estimated correlations.

Discussion

a. Shared classical training. We have shown that parameterised quantum circuits can learn families of ground-state measurement distributions through classical training. Each update combines losses from several Hamiltonian points, allowing the hypernetwork to learn how the circuit angles vary across the phase diagram. It can then predict angles at a new coordinate. When reference samples are available there, we can fine-tune the circuit for that particular distribution. Across 40 boundary points of the twelve-site chain, fine-tuning gave a lower final mean TVD than independent training.

The nine-site TFIM illustrates how the choice of circuit affects this fit. The hypernetwork IQP model and the RBM gave similar mean infidelities, 0.0115 and 0.0106, respectively. Adding three unmeasured ancillas to the IQP circuit reduced its mean infidelity to 0.00263, about a fourfold improvement. Encoding the field through input rotations gave higher infidelity, even when a hypernetwork supplied the remaining angles. For this family, learning how the angles vary with the field worked better than the fixed input encoding we tested.

b. Scaling to larger systems. The Fourier representation let us train on the 1,024-site chain and the 4 by 256 lattice using selected correlations. We trained across 50 reference points per geometry and compared local fits at one point in each. At σ=32\sigma=32, fine-tuning gave about one third of the independent model’s squared MMD on the chain and one seventh on the lattice. There, the shared model started with a higher error than the independently trained circuit and reached a lower error after local refinement.

c. Choice of objective. On the nine-site Haldane chain, MMD training gave lower mean infidelity across the parameter grid than CDF-KL when both used correlations through order four. The susceptibility of the CDF-KL fit traced portions of the phase boundary using the selected marginals. Infidelity measures agreement with the target at each point. Susceptibility measures how rapidly the learned distribution changes between nearby points. The position of a susceptibility peak can therefore agree with the target even when the probabilities on either side differ. The Fourier coefficients then show which correlations it reproduces and how their errors vary with order.

d. Information sharing and fine-tuning. Figure 7(b.Figure 7i) shows that shared training already brings the low-order correlation magnitudes closer to the target than data-informed initialization, before any local fitting. Initializations that avoid barren plateaus can still lead to different minima [68]. Shared training supplies a starting point informed by related distributions, and the hypernetwork parameterization can influence the minimum reached during local optimization [69]. In Figure 7, the adaptively trained hypernetwork reached the independent fits’ final mean TVD after about 550 local updates, compared with 1,500 for independent training. These counts exclude pretraining. With the same local update budget, hypernetwork fine-tuning gave lower mean squared MMD across the displayed bandwidths, including a reduction of about 25% at σ=1\sigma= 1.

This MMD comparison includes all coefficient orders, although the local loss uses only orders through four. Across the four selected initializations, fine-tuning brought the RMS coefficient magnitudes closer to the target at orders six, eight and ten. At order eight, however, independent training gave lower mean squared error for the individual coefficients. Agreement in the overall strength of the correlations can thus hide errors in the magnitude or sign of individual coefficients.

e. Reference acquisition and accuracy. Adaptive acquisition uses the correlations predicted by the current model to choose where to calculate the next ground state. In the nine-site fits, some of the acquired points cluster along branches of the learned boundary. At the same budget of 49 references, the fixed grid gave lower mean infidelity over the full diagram for both losses. When ground-state calculations are expensive, a practical use of the learned response is to choose a boundary point where a local fit is worth the cost. On the 1,024-site chain, the reference MPS at the evaluation point reached the sweep limit. The held-out samples quantify agreement with this finite-bond approximation to the ground state.

f. Dequantization. We draw reference samples directly from the MPS [70]. For broader circuit families, estimating selected correlations can be easier than constructing a classical sampler for the full distribution [69]. The gate generators determine rr and hence the cost of estimating correlations during training. Even at a fixed rr, which qubits the gates couple matters when seeking circuits whose outputs remain difficult to sample classically. Shared pretraining offers a promising route to avoiding dequantization if information learned across Hamiltonians helps the circuit recover correlations outside the local loss that are difficult for an efficient classical sampler to reproduce. Targets near phase boundaries, where correlations can change rapidly, provide a way to test this possibility by training both samplers on the same references and comparing their accuracy and sampling cost.

Code availability

The code is available on GitHub.

Acknowledgements

MHG proposed applying the train-on-classical, deploy-on-quantum approach [10] to conditional QCBMs [26] during the Spring School on Near-Term Quantum Computing at the Benasque Science Center. The hackathon challenge was proposed by A. Cervera-Lierta, A. Pérez-Salinas, A. Pozas-Kerstjens and J. Arguello-Luengo. There, Guillermo Rubiños-Rodríguez demonstrated the usefulness of this approach through deployment on quantum hardware with error mitigation. MHG also thanks fellow team members Oriol Balló and Daniel Quinn for discussions during the hackathon.

MHG is deeply grateful to Yuxuan Zhang for taking the time to explain automated phase-transition detection just before his train departed.

EK and MD acknowledge support from the Quantum Advantage TurboCharger (QATCH) Programme. MHG was supported by the UK Engineering and Physical Sciences Research Council (EPSRC) through grant EP/W524384/1 and by the University of Edinburgh.

AI Usage

The ideas in this work are original and were developed mainly through scientific discussions. We used GPT-6 Astra to assist with manuscript planning and writing, figure rendering and coding.

References

Targets, reference states and response calculation

Measurement basis and reference-state selection

All probability comparisons use computational-basis outcomes of the measured qubits. The TFIM benchmark counts each undirected bond of the periodic 3×33 \times3 lattice once. The Haldane chain in Eq. (3) has open ends, with n−2n-2 three-site terms and n−1n-1 two-site terms. The normalization is in Pauli matrices, not spin operators Sα=σα/2S^\alpha= \sigma^\alpha/2.

At small-system degeneracies, we project a fixed seeded vector into the ground eigenspace and normalize it, using a second fixed vector if needed. The first vector is invariant under global bit flip. This fixes the reference state without separately fixing both sublattice parities. The choice matters because different states in a degenerate ground space can have different measurement probabilities. We omit undefined ground-state responses at degeneracies.

For a nondegenerate ground state, the component of its derivative orthogonal to the state is

(I−∣ψ0⟩⟨ψ0∣)∣∂μψ0⟩=∑a>0∣ψa⟩⟨ψa∣∂μH∣ψ0⟩E0−Ea(35)\left(I-\lvert\psi_0\rangle\langle\psi_0\rvert\right)\lvert\partial_\mu\psi_0\rangle= \sum_{a>0}\lvert\psi_a\rangle\frac{\langle\psi_a\lvert\partial_\mu H\rvert\psi_0\rangle}{E_0-E_a} \tag*{(35)}

It gives ∂μp(x)=2Re⁡[⟨x∣ψ0⟩∗⟨x∣∂μψ0⟩]\partial_\mu p(x)=2\operatorname{Re}[\langle x\lvert\psi_0\rangle^*\langle x\lvert\partial_\mu\psi_0\rangle]. For learned distributions we differentiate through the circuit and the hypernetwork, holding trained parameters fixed. Input-rotation variants also include the explicit derivative of their encoding. Both calculations use the same dimensionless Hamiltonian coordinates in a comparison.

From infidelity to susceptibility

Assume that qλ(x)q_\lambda(x) is differentiable and its support remains fixed near λ\lambda, with positive probabilities on that support. All sums below are over this support. Normalization gives ∑x∂μq(x)=0\sum_x\partial_\mu q(x)=0. Expanding the overlap of the square-root probability vectors yields

∑xqλ(x)qλ+dλ(x)=1−18∑μν,x∂μq(x)∂νq(x)q(x)dλμdλν+o(∥dλ∥2).(36)\sum_x\sqrt{q_\lambda(x)q_{\lambda+d\lambda}(x)}=1-\frac{1}{8}\sum_{\mu\nu,x}\frac{\partial_\mu q(x)\partial_\nu q(x)}{q(x)}d\lambda_\mu d\lambda_\nu+o(\lVert d\lambda\rVert^2). \tag*{(36)}

Squaring this overlap gives Eq. (26), connecting local distribution infidelity to susceptibility. This relation concerns nearby points within one distribution family. Small infidelity between a model and its target at individual points does not by itself guarantee accurate derivatives. At zero probabilities, the susceptibility expression requires a well-defined limit.

TVD does bound the discrepancy of a fixed bounded observable. If ∣f(x)∣≤1\lvert f(x)\rvert\leq1, then ∣Epf−Eqf∣≤2TVD⁡(p,q)\lvert\mathbb{E}_p f-\mathbb{E}_q f\rvert\leq2\operatorname{TVD}(p,q). Derivative and peak-position errors also depend on how the distributions vary between parameter points and on the shape of the peaks.

Peak extraction and bulk guides

For the Haldane maps we inspect vertical slices at fixed h1/Jh_1/J. Interior local maxima of χ22\chi_{22} are retained when their prominence is at least one tenth of the finite range on the slice. Undefined points break the slice into finite segments. Accepted maxima on adjacent columns are linked by mutual nearest neighbors only when their vertical separation is at most three grid rows. This preserves gaps and multiple branches without fitting the peaks to a reference curve.

The bulk guides in Figure 5 are the numerical boundaries published in Fig. 2(a) of Ref. [29]. They come from an infinite-system ground-state calculation. The exact reference peaks instead describe the finite-chain measurement distribution.

Figure 5. Boundary detection from the Haldane QCBM fit in Figure 3(b). Learned susceptibility peaks are compared with exact finite-size peaks and the bulk guides.

Circuit families and model parameters

The generator family determines the allowed circuit transformations and conserved Pauli products, summarized in Table B1. We specify these families first, then give the finite gate lists and parameter counts used in the experiments.

FamilyPermitted generators or selection ruleConserved products
HaldaneEq. (B1), with even Y/ZY/Z counts on EE and OOSe,SoS_e,S_o
HeisenbergEq. (B2), with even X/YX/Y and Y/ZY/Z countsSX,SZS_X,S_Z
FermionicZjZ_j, AiZi+1⋯Zj−1BjA_iZ_{i+1}\cdots Z_{j-1}B_j, A,B∈{X,Y}A,B\in\{X,Y\}SZS_Z
IQPXSX_S on selected nonempty subsets SSall XX products
Restricted HaldaneXiX_i, allowed XiXjX_iX_j, AiBjA_iB_j within AEA_E or AOA_O, A,B∈{Y,Z}A,B\in\{Y,Z\}also XxX_x for x∈V⊥x\in V^\perp

Table B1. Table B1. Gate families and conserved Pauli products. EE and OO denote even and odd circuit sites, including ancillas. Each circuit uses a fixed ordered selection of the permitted generators. Gate weight counts nonidentity factors, independently of the distance between sites.

Compact Lie-algebra description

Let Pnc={I,X,Y,Z}⊗nc\mathcal{P}_{n_c}=\{I,X,Y,Z\}^{\otimes n_c} and let NJA(P)N_J^A(P) count factors from A⊂{X,Y,Z}A\subset\{X,Y,Z\} on sites JJ. Write Se=∏j evenXjS_e=\prod_{j\ \mathrm{even}}X_j, So=∏j oddXjS_o=\prod_{j\ \mathrm{odd}}X_j, and Sα=∏jσjαS_\alpha=\prod_j\sigma_j^\alpha. The Haldane family is

gH=span⁡R{iP: P∉{I,Se,So,SX},N{Y,Z}E(P),N{Y,Z}O(P) even}.(37)\mathfrak{g}_{\mathrm{H}}=\operatorname{span}_{\mathbb{R}}\left\{iP:\ \begin{array}{l}P\notin\{I,S_e,S_o,S_X\},\\ N_{\{Y,Z\}}^E(P),N_{\{Y,Z\}}^O(P)\ \text{even}\end{array}\right\}. \tag*{(37)}

We use the Pauli family proposed for the locally controlled Haldane chain in [19]. For even nc≥4n_c\ge4, the independently controlled Heisenberg family is

gHeis=span⁡R{iP: P∉{I,SX,SY,SZ},N{X,Y}(P),N{Y,Z}(P) even}.(38)\mathfrak{g}_{\mathrm{Heis}}=\operatorname{span}_{\mathbb{R}}\left\{iP:\ \begin{array}{l}P\notin\{I,S_X,S_Y,S_Z\},\\ N_{\{X,Y\}}(P),N_{\{Y,Z\}}(P)\ \text{even}\end{array}\right\}. \tag*{(38)}

This Pauli characterization is used in [19, 39]. Each commuting symmetry pair has four equally sized sectors. Removing its central directions gives

gH≅gHeis≅su(2nc−2)⊕4,dim⁡g=4nc−1−4,(39)\mathfrak{g}_{\mathrm{H}}\cong\mathfrak{g}_{\mathrm{Heis}}\cong\mathfrak{su}\left(2^{n_c-2}\right)^{\oplus4},\qquad\dim\mathfrak{g}=4^{n_c-1}-4, \tag*{(39)}

with different embeddings for the two families. The Haldane expression also applies for odd ncn_c with both circuit sublattices nonempty. The Heisenberg expression above assumes even ncn_c.

For fermionic circuits, define Jordan–Wigner Majoranas

γ2j−1=(∏k<jZk)Xj,γ2j=(∏k<jZk)Yj.(40)\gamma_{2j-1}=\left(\prod_{k<j}Z_k\right)X_j,\qquad\gamma_{2j}=\left(\prod_{k<j}Z_k\right)Y_j. \tag*{(40)}

The real Lie span of γaγb\gamma_a\gamma_b, a<ba<b, is so(2nc)\mathfrak{so}(2n_c) with dimension nc(2nc−1)n_c(2n_c-1). Up to sign, its Hermitian rotation generators are ZjZ_j and Ai(∏i<k<jZk)BjA_i\left(\prod_{i<k<j}Z_k\right)B_j, with A,B∈{X,Y}A,B\in\{X,Y\}. The intervening ZZ factors are part of the Jordan–Wigner representation. These quadratic generators produce fermionic Gaussian unitaries, also called fermionic linear optics, including when generators are repeated. With Gaussian inputs and computational-basis (occupation-number) measurements, their output distributions can be sampled efficiently on a classical computer [71, 72].

For the fermionic benchmarks, the computational zero state is Gaussian. The GHZ input is a superposition of two Gaussian occupation states whose relative phase can be tracked through the fermionic circuit [73].

IQP generators commute and their Lie span is Abelian. With all nonempty XX strings through weight ww, its dimension is ∑k=1w(nck)\sum_{k=1}^{w}\binom{n_c}{k}. The Lie-algebra dimension describes the available generators. The circuit parameter count depends on the finite gate list chosen for each experiment.

Finite gate lists and restrictions on active sites

For Figure 3, ancillas follow each group of three measured qubits. Haldane and Heisenberg words through weight three are sorted by increasing weight, support span and saved binary ordering. The first 298 distinct permitted words are retained. The fermionic basis has 276 words on twelve wires. Repeated rotations separated by noncommuting gates supply the remaining circuit angle parameters. IQP uses every one-, two- and three-site XX string, also giving 298 angles. Each model retains its fixed saved order. The GHZ preparation is on the complete register.

For the restricted Haldane family, choose nonempty qubit sets AE⊆EA_E\subseteq E and AO⊆OA_O\subseteq O within the two circuit sublattices (even and odd positions, respectively). The selected qubits are called active sites. The Y/ZY/Z pair gates are permitted only within these sets, while the specified X rotations are retained throughout the circuit. If the pairs connect each selected set, their binary incidence vectors give

r=(∣AE∣−1)+(∣AO∣−1).(41)r = (|A_E| - 1) + (|A_O| - 1). \tag*{(41)}

More generally, a graph of allowed pairs with cc connected components, including isolated sites, gives r=nc−cr = n_c - c because each connected component spans the even-parity labels on its vertices. Adding a new pair inside a connected component does not increase rr. All-X words have zero vector z\mathbf{z} and also leave rr unchanged. The unrestricted twelve-wire Haldane list has r=10r = 10. The r=4r = 4 and r=2r = 2 comparisons retain 298 parametrized rotations by allowed repetitions, but their additional conserved XX products change the model. Each adaptive fit also chooses its own reference coordinates, so the comparison changes both the allowed gates and the training data. Figure 7 instead uses fifteen wires, three active sites on each parity and 144 distinct permitted generators, repeated to obtain 316 parametrized rotations. Appendix § 2.1.5 2 specifies the circuit for both 1,024-site geometries.

Hypernetwork and RBM counts

For input dimension dλd_\lambda, a single tanh hidden layer of width ww and an output of length 2d2d give

h(λ)=tanh⁡(W1λ+b1),(42)h(\lambda) = \tanh(W_1\lambda+ b_1), \tag*{(42)}
(γϕ,βϕ)=W2h+b2,(43)(\gamma_\phi, \beta_\phi) = W_2h + b_2, \tag*{(43)}
Nhyper=w(dλ+1)+2d(w+1).(44)N_{\mathrm{hyper}} = w(d_\lambda+ 1) + 2d(w + 1). \tag*{(44)}

The dd base angles are additional trainable parameters. Fixed scaling of the network output does not add weights. This yields 9,581 hypernetwork weights for Figure 3, 10,157 for Figure 7, and 87,693 for Figure 8, each with w=15w = 15 and dλ=2d_\lambda= 2.

The RBM retains a shared visible-hidden interaction matrix and changes the visible and hidden biases using the HyperRBM rule [22]. With visible bits x\mathbf{x} and binary hidden units, write

fλ(x)=exp⁡(aλ⋅x)∏j[1+exp⁡(bλ,j+Wj⋅x)].(45)f_\lambda(\mathbf{x}) = \exp(\mathbf{a}_\lambda\cdot\mathbf{x}) \prod_j [1 + \exp(b_{\lambda,j} + \mathbf{W}_j \cdot\mathbf{x})]. \tag*{(45)}

The symmetrized unnormalized weight is fλ(x)+fλ(xˉ)f_\lambda(\mathbf{x}) + f_\lambda(\bar{\mathbf{x}}). We normalize it by summing the weights of every possible outcome for the reported small systems. The Haldane comparison uses 29 hidden units. Its 261 interaction weights and 38 biases give 299 generator parameters. A width-122 network for the 38 biases has 9,714 weights. The interaction matrix is shared across Hamiltonian points and trained jointly with the bias network.

Derivation and special cases of the Monte Carlo estimator

We derive the coefficient estimator introduced in Sec. III B, establish its variance and computational cost, and give the IQP and GHZ special cases. The derivation retains the order of the Pauli gates and applies to fixed circuit parameters. Appendix D explains how coefficient estimates enter the training losses.

Proposition 1 (Classical estimation for ordered Pauli circuits). Let Uθ=e−iθdPd⋯e−iθ1P1U_\theta= e^{-i\theta_d P_d}\cdots e^{-i\theta_1 P_1} be a circuit of d≥1d \ge1 Pauli rotations with Hermitian generators on ncn_c qubits, with fixed real angles and input ∣0⟩=∣0⟩⊗nc|0\rangle= |0\rangle^{\otimes n_c}. Write Pj=iνjXjxjZjzjP_j = i^{\nu_j}X_j^{x_j}Z_j^{z_j} as in Sec. III B, and let

r=dim⁡F2span⁡{z1,…,zd}.r = \dim_{\mathbb{F}_2}\operatorname{span}\{\mathbf{z}_1,\ldots,\mathbf{z}_d\}.

For any subset SS of the measured qubits, set cS=⟨0∣Uθ†ZSUθ∣0⟩c_S = \langle0|U_\theta^\dagger Z_S U_\theta|0\rangle. The estimator in Eq. (12), using NMCN_{\mathrm{MC}} independent uniform bitstrings, satisfies

Ec^S=cS,Var⁡(c^S)≤1−cS2NMC,Pr⁡(∣c^S−cS∣≥ε)≤2e−NMCε2/2.(46)\mathbb{E}\hat{c}_S = c_S,\qquad\operatorname{Var}(\hat{c}_S) \le\frac{1-c_S^2}{N_{\mathrm{MC}}},\qquad\Pr(|\hat{c}_S-c_S|\ge\varepsilon) \le2e^{-N_{\mathrm{MC}}\varepsilon^2/2}. \tag*{(46)}

Consequently, for 0<ε≤10 < \varepsilon\le1 and 0<δ<10 < \delta< 1, taking

NMC=⌈2ε−2log⁡(2/δ)⌉N_{\mathrm{MC}} = \lceil2\varepsilon^{-2}\log(2/\delta)\rceil

suffices for additive error ε\varepsilon with probability at least 1−δ1-\delta.

Let Wx=∑j=1d∣supp⁡(xj)∣≤dncW_x=\sum_{j=1}^{d}|\operatorname{supp}(x_j)|\le d n_c. Including preprocessing, a direct implementation uses

T=O(dnc2+d(1+r)2r+NMC[nc+Wx+d2r])(47)T=O\left(d n_c^2+d(1+r)2^r+N_{\mathrm{MC}}\left[n_c+W_x+d2^r\right]\right) \tag*{(47)}
M=O(dnc+d(1+r)2r)(48)\mathcal{M}=O\left(d n_c+d(1+r)2^r\right) \tag*{(48)}

operations and storage, respectively. Storage includes the circuit description, precomputed gate tables and working vectors, and is independent of NMCN_{\mathrm{MC}} when samples are processed sequentially. These counts treat numerical arithmetic and array indexing as unit-cost operations, with gate sines and cosines supplied as numerical inputs. Unbiasedness and the statistical bounds assume exact gate updates and do not include numerical roundoff. If d=poly⁡(nc)d=\operatorname{poly}(n_c) and r=O(log⁡nc)r=O(\log n_c), the operation count is polynomial in ncn_c, ε−1\varepsilon^{-1} and log⁡(δ−1)\log(\delta^{-1}), and the storage is polynomial in ncn_c.

The proof of Proposition 1 begins in Appendix C1, where we show that the mean overlap of two vectors with 2r2^r amplitudes equals the Fourier coefficient cSc_S. Appendix C2 completes the proof by bounding the variance, the number of draws needed for a given accuracy, and the time and storage required.

Invariant sets of bitstrings

Proof of Proposition 1. Use ∣b~⟩=H⊗nc∣b⟩|\widetilde{b}\rangle=H^{\otimes n_c}|b\rangle for the XX basis. On one site, Xi∣b~⟩=(−1)bi∣b~⟩X_i|\widetilde{b}\rangle=(-1)^{b_i}|\widetilde{b}\rangle, while ZiZ_i flips bib_i. Thus

Zzj∣b~⟩=∣b⊕zj~⟩(49)Z^{z_j}|\widetilde{b}\rangle=|\widetilde{b\oplus z_j}\rangle \tag*{(49)}
Xxj∣b⊕zj~⟩=(−1)xj⋅(b⊕zj)∣b⊕zj~⟩.(50)X^{x_j}|\widetilde{b\oplus z_j}\rangle=(-1)^{x_j\cdot(b\oplus z_j)}|\widetilde{b\oplus z_j}\rangle. \tag*{(50)}

Applying the rightmost factor first in Pj=iνjXxjZzjP_j=i^{\nu_j}X^{x_j}Z^{z_j} gives

Pj∣b~⟩=iνj(−1)xj⋅(b⊕zj)∣b⊕zj~⟩,(51)P_j|\widetilde{b}\rangle=i^{\nu_j}(-1)^{x_j\cdot(b\oplus z_j)}|\widetilde{b\oplus z_j}\rangle, \tag*{(51)}

Since Pj2=IP_j^2=I, its rotation is e−iθjPj=cos⁡θjI−isin⁡θjPje^{-i\theta_jP_j}=\cos\theta_j I-i\sin\theta_jP_j. It mixes only the two bitstrings separated by zjz_j. Repeating these shifts preserves the span of basis states with bitstrings in b+Vb+V. Use the basis matrix BB from Sec. III B. The binary vector ℓj∈F2r\ell_j\in\mathbb{F}_2^r is defined by Bℓj=zjB\ell_j=z_j. It specifies which vector entries gate jj mixes. The binary vector hj=BTxj∈F2rh_j=B^{\mathsf{T}}x_j\in\mathbb{F}_2^r specifies their relative signs. For zero input, ∣0⟩=2−nc/2∑b∣b~⟩|0\rangle=2^{-n_c/2}\sum_b|\widetilde{b}\rangle. A coset contains 2r2^r terms and has squared norm 2r−nc2^{r-n_c}. Dividing its amplitudes by the square root of that weight gives the normalized initial vector w0(b)(t)=2−r/2w_0^{(b)}(t)=2^{-r/2}, t∈F2rt\in\mathbb{F}_2^r. Its exact gate update is

wj(b)(t)=cos⁡θj wj−1(b)(t)−iνj+1sin⁡θj(−1)xj⋅b+hj⋅twj−1(b)(t⊕ℓj).(52)w_j^{(b)}(t)=\cos\theta_j\,w_{j-1}^{(b)}(t) -i^{\nu_j+1}\sin\theta_j(-1)^{x_j\cdot b+h_j\cdot t}w_{j-1}^{(b)}(t\oplus\ell_j). \tag*{(52)}

To check the phase, the source bitstring for the second term is b⊕Bt⊕zjb\oplus Bt\oplus z_j. Substitution in the Pauli action above gives iνj(−1)xj⋅b+hj⋅ti^{\nu_j}(-1)^{x_j\cdot b+h_j\cdot t}, which is multiplied by −isin⁡θj-i\sin\theta_j. Both terms use the pre-gate vector. The update is unitary within the invariant set, so the vector remains normalized. Bitstring arithmetic is modulo two, while νj\nu_j is an integer phase exponent. The representative for the second vector is b⊕sb\oplus s, and the same coordinate tt pairs bitstrings shifted by ss in Eq. (12). This alignment is required even when the two bitstrings belong to the same coset. Replacing the shifted representative by a canonical one without permuting its entries would change the estimator.

There are KV=2nc−rK_V=2^{n_c-r} distinct sets. The zero input ∣0⟩=∣0⟩⊗nc|0\rangle=|0\rangle^{\otimes n_c} has equal weight in all of them. Its final amplitude at b⊕Btb\oplus Bt is

⟨b⊕Bt∣Uθ∣0⟩=KV−1/2wd(b)(t).(53)\langle b\oplus Bt|U_\theta|0\rangle=K_V^{-1/2}w_d^{(b)}(t). \tag*{(53)}

The observable ZsZ_s shifts XX-basis bitstrings by ss. Hence, if R\mathcal{R} contains one representative of each set,

cS=1KV∑b∈RRe⁡∑twd(b⊕s)(t)∗wd(b)(t).(54)c_S=\frac{1}{K_V}\sum_{b\in\mathcal{R}}\operatorname{Re}\sum_t w_d^{(b\oplus s)}(t)^*w_d^{(b)}(t). \tag*{(54)}

Uniform bitstrings sample these sets uniformly. More explicitly, replacing bb by b⊕Bub\oplus Bu sends each vector entry tt to t⊕ut\oplus u in both vectors. Their overlap is unchanged. Thus every coset contributes the same sample value for each of its 2r2^r representatives, allowing us to sample bitstrings without enumerating the cosets. This proves the expectation identity in Eq. (13).

Variance and computational cost

By Cauchy–Schwarz, ∣GS∣≤1|G_S| \le1. Independent draws therefore obey

Var⁡(c^S)=KV−1∑b∈RGS(b)2−cS2NMC≤1−cS2NMC≤1NMC.(55)\operatorname{Var}(\hat{c}_S)=\frac{K_V^{-1}\sum_{\mathbf{b}\in\mathbb{R}}G_S(\mathbf{b})^2-c_S^2}{N_{\mathrm{MC}}}\le\frac{1-c_S^2}{N_{\mathrm{MC}}}\le\frac{1}{N_{\mathrm{MC}}}. \tag*{(55)}

Hoeffding’s bound gives

Pr⁡(∣c^S−cS∣≥ε)≤2exp⁡(−NMCε2/2).(56)\Pr(|\hat{c}_S-c_S|\ge\varepsilon)\le2\exp(-N_{\mathrm{MC}}\varepsilon^2/2). \tag*{(56)}

For a fixed collection of MM coefficients, a union bound replaces the prefactor by 2M2M. Independence between different coefficients is not needed for that bound. These bounds hold for fixed circuit parameters and independent estimator draws.

To account for preprocessing, binary elimination finds a basis BB and the coordinates ℓj\ell_j in O(dnc2)O(dn_c^2) operations. Forming hj=BTxj\mathbf{h}_j=B^T\mathbf{x}_j and the partner offsets xj⋅s\mathbf{x}_j\cdot\mathbf{s} also fits within this bound. For each gate, tabulate the permutation t↦t⊕ℓj\mathbf{t}\mapsto\mathbf{t}\oplus\ell_j and signs (−1)hj⋅t(-1)^{\mathbf{h}_j\cdot\mathbf{t}} in Eq. (52). Direct construction takes O(d(1+r)2r)O(d(1+r)2^r) binary operations and bits of storage, including the rr-bit permutation entries. The circuit description and binary basis data require O(dnc)O(dn_c) further storage, since r≤min⁡(d,nc)r\le\min(d,n_c).

Each draw uses two vectors of length 2r2^r, each passed through dd gates. Generating its ncn_c random bits costs O(nc)O(n_c), and calculating the parities xj⋅b\mathbf{x}_j\cdot\mathbf{b} costs O(Wx)O(W_x) by visiting the nonzero entries of each xj\mathbf{x}_j. The partner parities follow from xj⋅(b⊕s)=xj⋅b+xj⋅s\mathbf{x}_j\cdot(\mathbf{b}\oplus\mathbf{s})=\mathbf{x}_j\cdot\mathbf{b}+\mathbf{x}_j\cdot\mathbf{s}. With the tables available, each gate update takes O(2r)O(2^r) arithmetic and indexing operations. The final overlap has the same order of cost. Processing the draws sequentially therefore takes O(NMC[nc+Wx+d2r])O(N_{\mathrm{MC}}[n_c+W_x+d2^r]) operations and O(2r)O(2^r) numerical working storage, including temporary vectors. Only a running sum is needed. Adding preprocessing gives Eqs. (47) and (C3). Substituting the stated sample count and r=O(log⁡nc)r=O(\log n_c) gives the polynomial bounds.

The empty circuit can be evaluated directly. Finite-precision evaluation introduces an additional numerical error, separate from the sampling error in Eq. (46). Reverse-mode differentiation can require more memory unless intermediate states are recomputed. The estimator is a finite sum of differentiable functions of the angles, and its sampling distribution is angle-independent. Thus expectation and differentiation can be interchanged for its coefficient gradient. This does not give the same variance bound for that gradient, or remove bias from a subsequent nonlinear loss. At an adaptive training step, the fixed-angle statement applies conditionally when fresh bitstrings are sampled independently of the current parameters. It does not justify evaluating a fitted model on the same random bitstrings used to select it.

Enumerating all supported cosets removes MC variance. For the full Haldane gate list, however, each amplitude vector has exponentially many entries, so even a small number of cosets can be costly. Restricting the active sites reduces this cost, as used in Figure 6–Figure 9. Other simulation methods may remain efficient for gate sets requiring larger vectors in this estimator.

IQP phase estimator

For Pj=XxjP_j=X^{\mathbf{x}_j} every vector zj\mathbf{z}_j vanishes, so r=0r=0 and each vector has one entry, initially equal to one. The gate acts diagonally,

wj(b)=e−iθj(−1)xj⋅bwj−1(b).(57)w_j^{(\mathbf{b})}=e^{-i\theta_j(-1)^{\mathbf{x}_j\cdot\mathbf{b}}}w_{j-1}^{(\mathbf{b})}. \tag*{(57)}

Multiplying these factors gives wθ(b)=e−iΦθ(b)w_{\boldsymbol{\theta}}^{(\mathbf{b})}=e^{-i\Phi_{\boldsymbol{\theta}}(\mathbf{b})}, where Φθ(b)=∑jθj(−1)xj⋅b\Phi_{\boldsymbol{\theta}}(\mathbf{b})=\sum_j\theta_j(-1)^{\mathbf{x}_j\cdot\mathbf{b}}. Substitution in the overlap yields

GS(b;θ)=Re⁡[eiΦθ(b⊕s)e−iΦθ(b)](58)G_S(\mathbf{b};\boldsymbol{\theta})=\operatorname{Re}\left[e^{i\Phi_{\boldsymbol{\theta}}(\mathbf{b}\oplus\mathbf{s})}e^{-i\Phi_{\boldsymbol{\theta}}(\mathbf{b})}\right] \tag*{(58)}
=cos⁡[Φθ(b⊕s)−Φθ(b)].(59)=\cos[\Phi_{\boldsymbol{\theta}}(\mathbf{b}\oplus\mathbf{s})-\Phi_{\boldsymbol{\theta}}(\mathbf{b})]. \tag*{(59)}

Reducing the phase difference term by term gives

Φθ(b⊕s)−Φθ(b)=∑jθj(−1)xj⋅b[(−1)xj⋅s−1](60)\Phi_{\boldsymbol{\theta}}(\mathbf{b}\oplus\mathbf{s})-\Phi_{\boldsymbol{\theta}}(\mathbf{b})=\sum_j\theta_j(-1)^{\mathbf{x}_j\cdot\mathbf{b}}\left[(-1)^{\mathbf{x}_j\cdot\mathbf{s}}-1\right] \tag*{(60)}
=−2∑j:xj⋅s=1θj(−1)xj⋅b.(61)=-2\sum_{j:\mathbf{x}_j\cdot\mathbf{s}=1}\theta_j(-1)^{\mathbf{x}_j\cdot\mathbf{b}}. \tag*{(61)}

Terms with xj⋅s=0x_j \cdot s = 0 cancel. Since cosine is even, the remaining expression gives the sample used in Eq. (14). Averaging it gives

cS=Ebcos⁡[Φθ(b⊕s)−Φθ(b)].(62)c_S = \mathbb{E}_{\mathbf{b}}\cos[\Phi_\theta(\mathbf{b} \oplus\mathbf{s})-\Phi_\theta(\mathbf{b})]. \tag*{(62)}

This is the phase-average form underlying classical IQP optimization [9, 42]. With zero input, b\mathbf{b} is uniform over all bitstrings. Appendix C 4 derives the sampling distribution for GHZ input.

GHZ input and even-parity sampling

Corollary 1 (GHZ input with parity-preserving gates). Replace the zero input in Proposition 1 by ∣GHZ+⟩=(∣0⟩⊗nc+∣1⟩⊗nc)/2|\mathrm{GHZ+}\rangle= (|0\rangle^{\otimes n_c} + |1\rangle^{\otimes n_c})/\sqrt{2}, and suppose 1⋅zj=0\mathbf{1}\cdot\mathbf{z}_j = 0 for every gate. For even ∣S∣|S|, the same estimator, with independent bitstrings drawn uniformly from the even-parity subspace, obeys the same expectation, variance, probability and complexity bounds. For odd ∣S∣|S|, cS=0c_S=0 exactly.

Proof. The two computational-basis components have XX-basis expansions

∣0⟩⊗nc=2−nc/2∑b∣b~⟩,(63)|0\rangle^{\otimes n_c}=2^{-n_c/2}\sum_{\mathbf{b}}|\widetilde{\mathbf{b}}\rangle, \tag*{(63)}
∣1⟩⊗nc=2−nc/2∑b(−1)1⋅b∣b~⟩.(64)|1\rangle^{\otimes n_c}=2^{-n_c/2}\sum_{\mathbf{b}}(-1)^{\mathbf{1}\cdot\mathbf{b}}|\widetilde{\mathbf{b}}\rangle. \tag*{(64)}

Adding them and dividing by 2\sqrt{2} gives amplitude 2−(nc+1)/2[1+(−1)1⋅b]2^{-(n_c+1)/2}[1+(-1)^{\mathbf{1}\cdot\mathbf{b}}]. This vanishes for odd-parity bitstrings and equals 2−(nc−1)/22^{-(n_c-1)/2} for even-parity bitstrings. Therefore

∣GHZ+⟩=2−(nc−1)/2∑b: 1⋅b=0∣b~⟩.(65)|\mathrm{GHZ+}\rangle=2^{-(n_c-1)/2}\sum_{\mathbf{b}:\,\mathbf{1}\cdot\mathbf{b}=0}|\widetilde{\mathbf{b}}\rangle. \tag*{(65)}

When every zj\mathbf{z}_j has even parity, VV lies in the even-parity subspace, and each coset starting at an even-parity bitstring stays inside this support. Dividing the 2nc−12^{n_c-1} supported bitstrings into sets of size 2r2^r gives KGHZ=2nc−r−1K_{\mathrm{GHZ}}=2^{n_c-r-1} supported cosets. Each has weight KGHZ−1K_{\mathrm{GHZ}}^{-1} and the same normalized initial entries 2−r/22^{-r/2} as before. Thus the gate update is unchanged. For even ∣S∣|S|, 1⋅s=0\mathbf{1}\cdot\mathbf{s}=0, so the partner coset at b⊕s\mathbf{b}\oplus\mathbf{s} is also supported, and

cS=1KGHZ∑b∈RevenGS(b;θ),(66)c_S=\frac{1}{K_{\mathrm{GHZ}}}\sum_{\mathbf{b}\in\mathcal{R}_{\mathrm{even}}}G_S(\mathbf{b};\theta), \tag*{(66)}

where Reven\mathcal{R}_{\mathrm{even}} contains one representative of each supported coset. Sampling uniformly among even-parity bitstrings implements this average because every coset has the same size. The normalization and variance proof in Appendix C 2 therefore still apply. For odd ∣S∣|S|, ZSZ_S shifts into the unoccupied odd-parity sector, whose overlap with the state is zero. Hence cS=0c_S=0. Sampling nc−1n_c-1 independent fair bits and choosing the final bit to make the total parity even gives a uniform supported bitstring in O(nc)O(n_c) operations. The gate tables and vector updates are unchanged, so the complexity bounds also apply. □

For the Haldane and IQP circuits with GHZ input used here, only even-order coefficients therefore need MC evaluation. Other inputs or parity-changing gates require their corresponding amplitudes and sampling weights.

Spans of full and restricted Haldane gate vectors

Each Y/ZY/Z pair on sites u,vu,v has binary vector eu⊕ev\mathbf{e}_u\oplus\mathbf{e}_v, independently of whether its factors are YY or ZZ. Here eu\mathbf{e}_u has a one at site uu and zeros elsewhere. An XX-type gate has zero binary vector. Form a graph whose vertices are circuit sites and whose edges are the permitted Y/ZY/Z pairs. An edge vector has even parity within its connected component, and so does every binary sum of edge vectors.

Conversely, consider a connected component CC and fix a root site u∈Cu\in C. For every other site v∈Cv\in C, a path u=v0,v1,…,vm=vu=v_0,v_1,\ldots,v_m=v gives

⨁k=0m−1(evk⊕evk+1)=eu⊕ev,(67)\bigoplus_{k=0}^{m-1}(\mathbf{e}_{v_k}\oplus\mathbf{e}_{v_{k+1}})=\mathbf{e}_u\oplus\mathbf{e}_v, \tag*{(67)}

because each internal vertex occurs twice. The ∣C∣−1|C|-1 root-pair vectors are linearly independent because the coordinate at v≠uv \ne u occurs in just its own vector. They also span every even-parity bitstring zz supported on CC, since

z=⨁v∈C∖{u}zv=1(eu⊕ev).(68)\mathbf{z}=\bigoplus_{\substack{v\in C\setminus\{u\}\\ z_v=1}}(\mathbf{e}_u\oplus\mathbf{e}_v). \tag*{(68)}

The root coordinate agrees because z\mathbf{z} has even parity. Thus the span of these vectors on CC has dimension ∣C∣−1|C|-1. A singleton contributes zero. Components have disjoint supports, so a graph with cc components has total dimension ∑C(∣C∣−1)=nc−c\sum_C(|C|-1)=n_c-c.

For the full Haldane pair set, the even and odd circuit sublattices are each connected and nonempty. There are two components, giving

VH={z:z⋅1E=z⋅1O=0},r=(∣E∣−1)+(∣O∣−1)=nc−2.(69)V_{\mathrm{H}}=\{\mathbf{z}:\mathbf{z}\cdot\mathbf{1}_{E}=\mathbf{z}\cdot\mathbf{1}_{O}=0\},\qquad r=(|E|-1)+(|O|-1)=n_c-2. \tag*{(69)}

Additional Haldane gate vectors with even parity on each sublattice already belong to this span and do not enlarge it. Each simulated amplitude vector has 2nc−22^{n_c-2} entries.

For the restricted circuit, the permitted pairs connect only the nonempty sets AEA_E and AOA_O. Every site outside those sets is an isolated vertex, so

c=2+nc−∣AE∣−∣AO∣,r=nc−c=∣AE∣+∣AO∣−2.(70)c=2+n_c-|A_E|-|A_O|,\qquad r=n_c-c=|A_E|+|A_O|-2. \tag*{(70)}

This proves Eq. (16). Two sites in each selected set give r=2r=2 and four entries. Three give r=4r=4 and sixteen entries. Choosing a total of O(log⁡nc)O(\log n_c) selected sites makes 2r2^r polynomial in ncn_c. The complexity statement also requires the gate count and requested accuracy to scale as specified in Appendix C 2. If a selected set is disconnected, count its separate components instead of applying the connected-set formula.

Two-entry updates and a noncommuting Haldane example

Let i,ji,j be distinct sites on the same circuit sublattice, and retain ZiZjZ_iZ_j together with X-type generators. The only nonzero gate vector is z=ei⊕ej\mathbf{z}=\mathbf{e}_i\oplus\mathbf{e}_j, so V={0,z}V=\{0,\mathbf{z}\} and r=1r=1. Choose B=zB=\mathbf{z}. The two entries, indexed by t=0,1t=0,1, correspond to bitstrings b\mathbf{b} and b⊕z\mathbf{b}\oplus\mathbf{z}. On these basis states, ZiZjZ_iZ_j interchanges the entries. Expanding its exponential therefore gives

R(β)=e−iβ(0110)=(cos⁡β−isin⁡β−isin⁡βcos⁡β).(71)R(\beta)=e^{-i\beta\left(\begin{smallmatrix}0&1\\1&0\end{smallmatrix}\right)} =\begin{pmatrix}\cos\beta&-i\sin\beta\\-i\sin\beta&\cos\beta\end{pmatrix}. \tag*{(71)}

An XxX^x rotation is diagonal, with entry tt multiplied by e−iα(−1)x⋅(b⊕tz)e^{-i\alpha(-1)^{\mathbf{x}\cdot(\mathbf{b}\oplus t\mathbf{z})}}. In particular, put ξ=(−1)bi\xi=(-1)^{b_i}. The two eigenvalues of XiX_i are ξ\xi and −ξ-\xi, giving

D(α;ξ)=(e−iαξ00eiαξ).(72)D(\alpha;\xi)=\begin{pmatrix}e^{-i\alpha\xi}&0\\0&e^{i\alpha\xi}\end{pmatrix}. \tag*{(72)}

These are the two-entry updates stated in the main text. Starting from w0=(1,1)T/2\mathbf{w}_0=(1,1)^{\mathsf{T}}/\sqrt{2}, apply them in circuit order. For U=e−iβZiZje−iαXiU=e^{-i\beta Z_iZ_j}e^{-i\alpha X_i},

w1=D(α;ξ)w0=12(e−iαξeiαξ),w2=R(β)w1.(73)\mathbf{w}_1=D(\alpha;\xi)\mathbf{w}_0=\frac{1}{\sqrt{2}}\begin{pmatrix}e^{-i\alpha\xi}\\e^{i\alpha\xi}\end{pmatrix},\qquad\mathbf{w}_2=R(\beta)\mathbf{w}_1. \tag*{(73)}

Multiplication gives

w(b)=12(cos⁡β e−iαξ−isin⁡β eiαξcos⁡β eiαξ−isin⁡β e−iαξ).(74)\mathbf{w}^{(\mathbf{b})}=\frac{1}{\sqrt{2}}\begin{pmatrix}\cos\beta\,e^{-i\alpha\xi}-i\sin\beta\,e^{i\alpha\xi}\\\cos\beta\,e^{i\alpha\xi}-i\sin\beta\,e^{-i\alpha\xi}\end{pmatrix}. \tag*{(74)}

A correlator sample is the real overlap of this vector with the one at b⊕s\mathbf{b}\oplus\mathbf{s}, pairing the same tt in both vectors as in Eq. (12). The final ZiZjZ_iZ_j rotation commutes with every measured ZsZ_s, so

U†ZsU=eiαXieiβZiZjZse−iβZiZje−iαXi=eiαXiZse−iαXi.(75)U^\dagger Z_sU=e^{i\alpha X_i}e^{i\beta Z_iZ_j}Z_se^{-i\beta Z_iZ_j}e^{-i\alpha X_i}=e^{i\alpha X_i}Z_se^{-i\alpha X_i}. \tag*{(75)}

Thus all these correlators are independent of β\beta, despite the noncommuting gates. To obtain dependence on the intervening rotation, take zero input and append an XiX_i rotation,

U=e−iγXie−iβZiZje−iαXi.(76)U=e^{-i\gamma X_i}e^{-i\beta Z_iZ_j}e^{-i\alpha X_i}. \tag*{(76)}

Its gate vectors still span a one-dimensional binary space, so r=1r=1. The final vector is w3=D(γ;ξ)R(β)D(α;ξ)w0w_3=D(\gamma;\xi)R(\beta)D(\alpha;\xi)w_0. Write it as w3=2−1/2(f0(ξ),f1(ξ))Tw_3=2^{-1/2}(f_0(\xi),f_1(\xi))^{\mathsf T}, where

f0(ξ)=e−iγξ(cos⁡β e−iαξ−isin⁡β eiαξ),(77)f_0(\xi)=e^{-i\gamma\xi}\left(\cos\beta\,e^{-i\alpha\xi}-i\sin\beta\,e^{i\alpha\xi}\right), \tag*{(77)}
f1(ξ)=eiγξ(cos⁡β eiαξ−isin⁡β e−iαξ).(78)f_1(\xi)=e^{i\gamma\xi}\left(\cos\beta\,e^{i\alpha\xi}-i\sin\beta\,e^{-i\alpha\xi}\right). \tag*{(78)}

For ZiZ_i, the partner bitstring is b⊕ei\mathbf b\oplus\mathbf e_i, which replaces ξ\xi by −ξ-\xi and interchanges f0f_0 and f1f_1. The sample is consequently

G{i}(b)=12Re⁡[f1(ξ)∗f0(ξ)+f0(ξ)∗f1(ξ)]=Re⁡[f0(ξ)∗f1(ξ)].(79)G_{\{i\}}(\mathbf b)=\frac{1}{2}\operatorname{Re}\left[f_1(\xi)^*f_0(\xi)+f_0(\xi)^*f_1(\xi)\right]=\operatorname{Re}\left[f_0(\xi)^*f_1(\xi)\right]. \tag*{(79)}

The product simplifies to

f0(ξ)∗f1(ξ)=e2iγξ(cos⁡2β e2iαξ+sin⁡2β e−2iαξ)(80)f_0(\xi)^*f_1(\xi)=e^{2i\gamma\xi}\left(\cos^2\beta\,e^{2i\alpha\xi}+\sin^2\beta\,e^{-2i\alpha\xi}\right) \tag*{(80)}
=e2iγξ[cos⁡(2α)+iξcos⁡(2β)sin⁡(2α)].(81)=e^{2i\gamma\xi}\left[\cos(2\alpha)+i\xi\cos(2\beta)\sin(2\alpha)\right]. \tag*{(81)}

Taking the real part and using ξ2=1\xi^2=1 gives a value independent of the sampled bitstring. Its average is therefore

⟨Zi⟩=cos⁡(2γ)cos⁡(2α)−sin⁡(2γ)sin⁡(2α)cos⁡(2β).(82)\langle Z_i\rangle=\cos(2\gamma)\cos(2\alpha) -\sin(2\gamma)\sin(2\alpha)\cos(2\beta). \tag*{(82)}

The same result follows by conjugating the observable, starting with the last applied gate. For anticommuting Pauli words, eiθPQe−iθP=Qcos⁡(2θ)+iPQsin⁡(2θ)e^{i\theta P}Qe^{-i\theta P}=Q\cos(2\theta)+iPQ\sin(2\theta). Here iXiZi=YiiX_iZ_i=Y_i, iZiZjYi=XiZjiZ_iZ_jY_i=X_iZ_j, and iXiYi=−ZiiX_iY_i=-Z_i. The successive observables are

Q1=eiγXiZie−iγXi=Zicos⁡(2γ)+Yisin⁡(2γ),(83)Q_1=e^{i\gamma X_i}Z_ie^{-i\gamma X_i}=Z_i\cos(2\gamma)+Y_i\sin(2\gamma), \tag*{(83)}
Q2=eiβZiZjQ1e−iβZiZj=Zicos⁡(2γ)+Yisin⁡(2γ)cos⁡(2β)+XiZjsin⁡(2γ)sin⁡(2β),(84)\begin{aligned} Q_2&=e^{i\beta Z_iZ_j}Q_1e^{-i\beta Z_iZ_j}\\ &=Z_i\cos(2\gamma)+Y_i\sin(2\gamma)\cos(2\beta)+X_iZ_j\sin(2\gamma)\sin(2\beta), \tag*{(84)} \end{aligned}
Q3=eiαXiQ2e−iαXi=Zi[cos⁡(2γ)cos⁡(2α)−sin⁡(2γ)cos⁡(2β)sin⁡(2α)]+Yi[cos⁡(2γ)sin⁡(2α)+sin⁡(2γ)cos⁡(2β)cos⁡(2α)]+XiZjsin⁡(2γ)sin⁡(2β).(85)\begin{aligned} Q_3&=e^{i\alpha X_i}Q_2e^{-i\alpha X_i}\\ &=Z_i\left[\cos(2\gamma)\cos(2\alpha)-\sin(2\gamma)\cos(2\beta)\sin(2\alpha)\right]\\ &\quad+Y_i\left[\cos(2\gamma)\sin(2\alpha)+\sin(2\gamma)\cos(2\beta)\cos(2\alpha)\right]\\ &\quad+X_iZ_j\sin(2\gamma)\sin(2\beta). \tag*{(85)} \end{aligned}

On ∣0⟩⊗nc\lvert0\rangle^{\otimes n_c}, the expectations of ZiZ_i, YiY_i, XiZjX_iZ_j are 11, 00, 00, respectively. The coefficient of ZiZ_i in Q3=U†ZiUQ_3=U^\dagger Z_iU reproduces the overlap result. The example shows how noncommuting gates can affect the measured coefficient while requiring only two amplitudes per vector. For the parity-preserving GHZ input, this odd-order coefficient vanishes by Appendix C 4.

Exact gate cancellation for selected observables

Some gates cancel in U†ZsUU^\dagger Z_sU because they commute with the complete backward-evolved operator at their location. For a Pauli term QQ,

eiθPQe−iθP={Q,[P,Q]=0,Qcos⁡2θ+iPQsin⁡2θ,{P,Q}=0.(86)e^{i\theta P}Qe^{-i\theta P}= \begin{cases} Q, & [P,Q]=0,\\ Q\cos2\theta+iPQ\sin2\theta, & \{P,Q\}=0. \end{cases} \tag*{(86)}

The implementation keeps a binary span containing the Pauli labels of the backward-evolved observable. A gate commuting with every basis element of that span commutes with the complete observable and can be removed. When a gate anticommutes with any basis element, it is retained and its Pauli label is added to the span. This rule removes gates without changing the selected coefficient, although it may retain some gates whose effects cancel at particular angles. The estimator uses the remaining gates in their original order. The full circuit, including gates omitted from this coefficient calculation, is used for sampling.

Losses from the available coefficients

This appendix gives the marginal reconstruction used by CDF-KL and the independent coefficient estimates used by MMD. Both losses are evaluated at each reference point and averaged to train the shared model.

Marginal reconstruction and valid cumulative probabilities

Let q^A\widehat{q}_{A} be the vector obtained from Eq. (19) using estimated coefficients. Its entries sum to one if c∅=1c_{\varnothing}=1, but may be negative. We compute its Euclidean simplex projection

ΠΔ(v)=arg min⁡ui≥0, ∑iui=1 ∥u−v∥22.(87)\Pi_{\Delta}(v)=\underset{u_i\geq0,\ \sum_i u_i=1}{\operatorname{arg\,min}}\ \lVert u-v\rVert_2^2. \tag*{(87)}

Explicitly, ui=max⁡(vi−τ,0)u_i=\max(v_i-\tau,0), with τ\tau chosen to make the sum one. Sorting the entries determines the active set and τ\tau. The probability used for cumulative sums is

qA,ϵ=(1−ϵ)ΠΔ(q^A)+ϵ2−∣A∣1.(88)q_{A,\epsilon}=(1-\epsilon)\Pi_{\Delta}(\widehat{q}_{A})+\epsilon2^{-\lvert A\rvert}\mathbf{1}. \tag*{(88)}

The same operation is applied to the empirical target marginal. Projection leaves an exact normalized nonnegative vector unchanged before smoothing. For 0<ϵ<10<\epsilon<1, every outcome has probability at least ϵ2−ℓ\epsilon2^{-\ell}, where ℓ=∣A∣\ell=\lvert A\rvert. At each retained threshold, the cumulative probability therefore lies in [ϵ2−ℓ,1−ϵ2−ℓ][\epsilon2^{-\ell},1-\epsilon2^{-\ell}]. The Bernoulli KL terms and their derivatives with respect to the cumulative probabilities are then finite. The projection itself remains piecewise differentiable. We use ϵ=10−3\epsilon=10^{-3} in the reported restricted-loss experiments. The CDF omits the last threshold, whose probability is identically one.

With all thresholds and fixed ordering, zero CDF-KL implies equality of the smoothed marginal vectors because consecutive CDF differences recover their individual probabilities. Equality of selected marginals can still leave the full distributions different. Each ℓ\ell-site marginal requires all 2ℓ2^{\ell} coefficients on that subset, even when only some subsets are used. The small-system four-site studies use the union of all such subsets. The large-system study selects only 32 windows.

The bit order in each marginal is fixed throughout training. Changing it changes the cumulative events and can change CDF-KL. The quadratic expansion

dB(u,u+δ)=δ22u(1−u)+O(δ3)(89)d_{\mathrm{B}}(u,u+\delta)=\frac{\delta^2}{2u(1-u)}+O(\delta^3) \tag*{(89)}

shows why the cumulative events are weighted differently from direct squared coefficient errors. Projection and logarithms generally bias the MC loss and its gradient relative to CDF-KL evaluated with exact coefficients and the same smoothing. The smoothing parameter separately controls how much each marginal is mixed with the uniform distribution.

Each projected marginal is a valid probability vector, but independently projected marginals can disagree on overlapping sites. CDF-KL compares them separately and does not reconstruct a global distribution.

Gaussian kernels and MC products

For binary coordinates,

e−(xi−yi)2/(2σ2)=(1−ρ)+ρ(−1)xi+yi.(90)e^{-(x_i-y_i)^2/(2\sigma^2)}=(1-\rho)+\rho(-1)^{x_i+y_i}. \tag*{(90)}

Multiplying over sites expands the kernel as kΩ(x,y)=∑SwS(S)(−1)∑i∈S(xi+yi)k_{\Omega}(\mathbf{x},\mathbf{y})=\sum_{S}w_S(S)(-1)^{\sum_{i\in S}(x_i+y_i)}. Substitution into the kernel mean discrepancy gives Eq. (22). Setting weights outside Ω\Omega to zero defines a positive-semidefinite restricted kernel. Distributions agreeing on the retained coefficients have zero restricted MMD even when they differ in omitted coefficients. This motivates the higher-order evaluation in Figure 7.

With a fixed empirical target cˉS\bar{c}_{S}, use independent batches of sampled bitstrings in

L^=∑S∈ΩwS(c^S(1)−cˉS)(c^S(2)−cˉS).(91)\widehat{L}=\sum_{S\in\Omega}w_S(\widehat{c}_{S}^{(1)}-\bar{c}_{S})(\widehat{c}_{S}^{(2)}-\bar{c}_{S}). \tag*{(91)}

The two model estimates use independent draws, making L^\widehat{L} unbiased for the restricted squared discrepancy to the fixed empirical target. Squaring a single noisy coefficient error would instead add its variance. Individual estimates can be negative. Finite target samples still contribute error.

In the Gaussian-sampled comparison in Table G1, 162 nonempty supports of even size are drawn with replacement at each update, with probabilities proportional to the full Gaussian weights. The sample average is multiplied by their total kernel mass. Repeated supports count as repeated draws, so 162 draws do not imply 162 distinct coefficients. Odd-order model coefficients are zero. Their empirical target residuals add a parameter-independent term, which is omitted from the optimization. At σ=1\sigma= 1, this sampling includes even orders above four. It therefore matches the coefficient-evaluation draw budget of the CDF study, not its exact information set.

ObjectivePoint selectionPointsrrCoefficientsMean T(p,q)\mathcal{T}(p,q)
LikelihoodFixed grid10010Full probabilities0.00670
LikelihoodFixed grid4910Full probabilities0.0175
CDF-KLFixed grid4910Exact, k≤4k \le40.0743
CDF-KLFixed grid4910MC, k≤4k \le40.0751
MMDFixed grid4910Exact, k≤4k \le40.0497
MMDFixed grid4910MC, k≤4k \le40.0496
CDF-KLAdaptive4910MC, k≤4k \le40.0819
MMDAdaptive4910MC, k≤4k \le40.0655
CDF-KLAdaptive494MC, k≤4k \le40.107
CDF-KLAdaptive492MC, k≤4k \le40.107
MMDAdaptive494MC, k≤4k \le40.0841
MMDAdaptive492MC, k≤4k \le40.0906
Gaussian-sampled MMDFixed grid4910Exact, 162 draws0.0483
Gaussian-sampled MMDFixed grid4910MC, 162 draws0.0484
CDF-KL + MMDFixed grid4910Exact, k≤4k \le40.0561
CDF-KL + MMDFixed grid4910MC, k≤4k \le40.0549

Table G1. Mean Haldane distribution infidelity on the 40 × 40 evaluation grid. Each row is one fit. Exact and MC specify coefficient evaluation during training, with 64 MC draws per nonzero coefficient.

The combined objective uses equal coefficients for CDF-KL and order-four MMD in the reported nine-site fits. The MMD weights are not renormalized after restricting orders. These conventions fix the relative contribution of the two losses.

Bias-corrected MMD evaluation

The sampled MMD in Figure 8 and Figure 9 requires correcting both model and target squared coefficients. For independent, identically distributed real estimator draws Y1,…,YNY_1,\ldots,Y_N with mean cSc_S, the diagonal-subtraction construction used for IQP MMD [10] gives

cS2^=(∑j=1NYj)2−∑j=1NYj2N(N−1)=Y‾ 2−sY2N,N>1,(92)\widehat{c_S^2} = \frac{\left(\sum_{j=1}^{N}Y_j\right)^2-\sum_{j=1}^{N}Y_j^2}{N(N-1)} = \overline{Y}^{\,2}-\frac{s_Y^2}{N}, \qquad N>1, \tag*{(92)}

where sY2s_Y^2 is the sample variance with denominator N−1N-1. This identity also applies to our Pauli coefficient estimator. If model and target coefficients use independent evaluation samples, their squared difference has the unbiased estimate

(cS(q)−cS(p))2=(Y‾−T‾)2−sY2N−sT2M.(93)\left(c_S(q)-c_S(p)\right)^2 = \left(\overline{Y}-\overline{T}\right)^2-\frac{s_Y^2}{N}-\frac{s_T^2}{M}. \tag*{(93)}

Here T1,…,TMT_1,\ldots,T_M are target parity samples, with M>1M>1. Evaluation draws are independent of model fitting and selection. Averaging the corrected squared differences over supports drawn from the Gaussian kernel weights estimates Eq. (22). Appendix § 2.1.5 gives the sample budgets and grouping used for the error bars. The coefficient plots in Figure 4 and Figure 7 use exact probabilities and require no MC variance correction.

Training, acquisition and local refinement

Joint parameter updates across reference distributions

The trainable parameters η=(θ0,ϕ)\eta=(\theta_0,\phi) comprise the shared base circuit angles and the hypernetwork weights. For fixed reference data, let ℓj\ell_j denote the loss at λj\lambda_j. A shared update differentiates the mean loss,

L(η)=1K∑j=1Kℓj(θη(λj)),(94)\mathcal{L}(\eta)=\frac{1}{K}\sum_{j=1}^{K}\ell_j(\theta_\eta(\lambda_j)), \tag*{(94)}
∇ηL=1K∑j=1K[∂θη(λj)∂η]T∇θℓj(θη(λj)).(95)\nabla_\eta\mathcal{L}=\frac{1}{K}\sum_{j=1}^{K}\left[\frac{\partial\theta_\eta(\lambda_j)}{\partial\eta}\right]^{\mathsf{T}}\nabla_\theta\ell_j(\theta_\eta(\lambda_j)). \tag*{(95)}

The hypernetwork takes λj\lambda_j as input, while the corresponding reference samples or coefficients define ℓj\ell_j. In the small-system phase-map fits, all reference points form one batch. For larger systems, we accumulate their gradients sequentially before applying one optimizer update. In both cases, each update learns from several distributions. For MC objectives, the random inputs remain fixed during differentiation in Eq. (95).

For likelihood training, let p^j\widehat{p}_j be the empirical distribution of the reference samples at λj\lambda_j. Then

LNLL=1K∑j=1K[DKL(p^j∥qη,λj)+H(p^j)],(96)\mathcal{L}_{\mathrm{NLL}}=\frac{1}{K}\sum_{j=1}^{K}\left[D_{\mathrm{KL}}(\widehat{p}_j\Vert q_{\eta,\lambda_j})+H(\widehat{p}_j)\right], \tag*{(96)}

where H(p^j)H(\widehat{p}_j) is the Shannon entropy. This term is fixed by the data, so NLL and empirical KL give the same parameter gradients. The CDF-KL and MMD fits use the same shared update with their respective losses.

Likelihood benchmarks and initialization

The small-system likelihood fits use full batches with equal reference-point weights and the adaptive moment estimation (Adam) optimizer [75]. Circuit probabilities are obtained by summing over ancilla outcomes. We select checkpoints by empirical validation NLL. Three TFIM initializations share the same sampled datasets. The Haldane map comparison uses one initialization.

The RBM follows the construction of Ref. [22], with likelihood training using all outcome probabilities.

The TFIM hypernetwork IQP model uses all single- and two-site XX strings on nine qubits, giving 45 angles. Adding three ancillas gives 78 angles. Six ancilla-only rotations leave the measured distribution unchanged because they commute through the remaining IQP gates. Both models use GHZ preparation on the complete register. The conditional-rotation variants encode the field affinely through RyR_y gates before the first Hadamard layer of the diagonal-ZZ representation. These rotations can break the global-flip symmetry. We compare them with shared IQP angles and with hypernetwork-generated angles.

The evaluation grid contains 161 fields, including 144 outside the training and validation sets. Figure E1 follows three diagnostic fits of each architecture using the benchmark settings. NLL and TVD are averaged equally over the nine training fields. Response peaks are the global maxima on the evaluation grid, including its endpoints, with ties resolved by choosing the smallest field.

Figure E1. Training and response accuracy for the nine-site TFIM. (a) Training NLL and (b) training-grid TVD for the five architectures in Figure 2. (c,d) Absolute errors in the locations of the magnetic-response and susceptibility maxima. Thin curves show three diagnostic fits per model. Thick curves and bands show their mean and one sample standard deviation, with peak errors calculated before averaging. Dotted lines mark the field-grid spacing.

In Figure E1, the three-ancilla IQP model reaches a training-grid TVD of about 0.035, roughly half that of the RBM and IQP without ancillas. This agrees with the ordering of held-out infidelities in Figure 2. Conditional rotations give higher TVD, with or without a hypernetwork. Peak-position errors continue to fluctuate after the training loss has largely flattened, and the ancilla model does not consistently give the closest peaks. Lower distribution error therefore need not imply a more accurate peak position.

Adaptive reference selection

Acquisition follows Eqs. (30) and (31), adapted from Ref. [28]. The small-system feature vector contains every nonconstant coefficient through order four. For larger systems, it combines the nonconstant coefficients of the 32 selected windows. We standardize each feature using its mean and standard deviation over acquired references, with the standard deviation floored at 0.05.

The encoder has two tanh hidden layers of width 32 and an eight-dimensional latent vector. The decoder reconstructs the feature vector. At each acquisition stage, the autoencoder takes 200 Adam updates from its previous weights with fresh optimizer moments. The average μt\mu_t includes all acquired references. We keep the model’s MC draws fixed within each search stage when calculating gradients with respect to the Hamiltonian parameters.

A search starts from an acquired point and takes eight steps of size 0.1 along the normalized gradient, with each Hamiltonian parameter rescaled to [0,1][0,1]. With probability 0.2, we instead choose the point farthest from the acquired references among 2,048 uniformly drawn candidates. Accepted points must have Euclidean separation at least 0.006 in these rescaled parameters. Small-system validation and evaluation points are excluded from acquisition. The 49-point protocol starts with a 3×33 \times3 grid and adds ten references in each of four rounds. Each of the five training stages takes 300 updates, weighting the available reference points equally.

We calculate reference states only at accepted points. The acquisition gradient passes through the model coefficients and encoder, so it requires no derivative of the reference solver.

Combined-loss local refinement

Figure 7 uses CDF-KL pretraining and local CDF-KL plus MMD with equal loss weights. The local MMD kernel has bandwidth σ=1\sigma= 1 and is truncated through order four. We estimate each of the 561 nonzero even-order coefficients with two independent groups of 32 MC draws. CDF-KL uses their mean and MMD uses their cross product. Independent fits use data-informed initialization.

At the 40 boundary points, we store the exact-diagonalization targets as untruncated MPSs and use 5,000 target samples for each local fit. For each initialization and pretraining method, we select a checkpoint by validation NLL. Angle-only and hypernetwork fine-tuning start from the same pretrained circuit angles. At each boundary point, we reset Adam moments and decrease the learning rate linearly from 0.003 to 0.00003 over 1,500 updates. Local evaluation uses the final update.

For the bandwidth evaluation, the Gaussian kernel is Kσ(x,y)=exp⁡[−dH(x,y)/(2σ2)]K_{\sigma}(x,y) = \exp[-d_H(x,y)/(2\sigma^2)], where dHd_H is Hamming distance. We sum over all 4,096 coefficients with weights ρσ∣S∣(1−ρσ)12−∣S∣\rho_{\sigma}^{|S|}(1-\rho_{\sigma})^{12-|S|}, where ρσ=(1−e−1/(2σ2))/2\rho_{\sigma} = (1-e^{-1/(2\sigma^2)})/2. This evaluates the full kernel from exact probabilities, including coefficient orders absent from the training objective.

Figure E2 shows the four initializations used in Figure 7, with panels i–iv corresponding to seeds 0, 1, 2 and 7. Both fixed-grid and adaptive training reach CDF-KL losses of about 2×10−42 \times10^{-4}. The adaptive loss rises when new references are added and then decreases within each stage. Despite similar final losses, the learned peak locations vary across initializations. The maps therefore provide information about the learned response beyond the training loss.

CDF-KL-only comparison

Table E1 uses CDF-KL alone for local training, in contrast to the combined loss in Figure 7. It averages three initializations with the same data and MC draws. Independent QCBMs start from the angles produced by an

kkAFIndependentA+FTF+FT
27.215.950.2430.2240.230
42.753.011.671.271.27
61.291.530.8920.6810.682
80.7621.040.6150.4360.508
100.5771.070.6250.3610.623
120.5222.041.520.3871.16

caption: tbl:E1. Coefficient errors after CDF-KL-only training for the twelve-site Haldane target. Mean squared errors at order kk, in units of 10−310^{-3}, averaged over the same 40 boundary points and three initializations. A and F denote adaptive and fixed-grid pretraining, and FT denotes local fine-tuning.

:::

untrained hypernetwork at each coordinate.

Fine-tuning uses 1,500 Adam updates at learning rate 3×10−43 \times10^{-4} and optimizes both the hypernetwork and base circuit angles. Independent QCBMs use learning rate 0.003. Higher-order coefficient errors are used only for evaluation.

Adaptive fine-tuning has the lowest mean coefficient error at every displayed order. The two fine-tuned models have similar errors through order six, but fine-tuning after adaptive pretraining gives lower errors at orders eight to twelve. At order twelve, its error is about four times smaller than that of independent training. The improvement therefore extends to coefficients absent from the local loss.

Large-system reference calculations and diagnostics

This appendix gives the reference calculations, training budgets and held-out evaluation used in Figure 8 and Figure 9. For both geometries, we sample DMRG reference states in the computational basis and use separate training and evaluation samples at the local fitting point.

MPS references for the chain

An MPS represents amplitudes as contractions of site tensors with finite bond dimensions [76, 77]. We use Julia/ITensor DMRG [32]. Its spin-1/21/2 operators require coefficients 8J8J, 2h12h_1 and 4h24h_2 for the three-, one- and two-site terms to reproduce Eq. (3). The chain has open boundaries.

Each chain reference calculation uses two initial states and retains the lower-energy result. The default bond-dimension cap is 128, the truncation cutoff is 10−1010^{-10}, and the total-energy stopping tolerance is 10−810^{-8}. We allow three to twelve sweeps. The local eigensolver uses tolerance 10−1210^{-12} and Krylov dimension ten. The stopping tolerance measures energy changes between sweeps.

For the same spin-1/21/2 cluster-Ising Hamiltonian, Ref. [29] located bulk phase boundaries using infinite-system DMRG with bond dimension at most 150. Its 135-site calculations used bond dimension 130, with no visible change in the reported figures upon increasing it to 150.

The chain references are optimized in the global-X=+1X = +1 sector, represented as a conserved parity after local Hadamards. We transform the MPS back to the physical computational basis before sampling.

Sequential MPS sampling generates outcomes directly from the retained state without Markov-chain equilibration [70]. Each chain reference supplies 20,000 outcomes for the six-bit histograms and training coefficients. The sampled distribution retains the MPS approximation, which can be more restrictive near critical regions [67].

At the evaluation coordinate (h1/J,h2/J)=(1.46,−1.40)(h_1/J,h_2/J)= (1.46,-1.40), the MMD comparison uses held-out samples from the same cap-128 MPS used to generate the training data.

Sparse circuit and response evaluation

Both geometries use the same circuit on 1,024 measured qubits and 342 unmeasured ancillas. We use single-site XX rotations, nearest-neighbor XXXX rotations in circuit order and paired Y/ZY/Z rotations. These gates give r=2r=2, so each of the two evolved vectors has four amplitudes. Choosing gates to keep rr small requires care, since some choices also yield circuits that are easy to sample classically.

CDF-KL uses 32 disjoint six-site windows. Each marginal has 63 nonconstant coefficients, of which 31 have even order. For MMD, we draw Fourier supports over all 1,024 measured qubits, with 683 draws per bandwidth and update at σ≃3.82\sigma\simeq3.82, 8.35, 11.2. We average the squared MMD over these bandwidths and add it to CDF-KL with weight αMMD=0.893\alpha_{\mathrm{MMD}} = 0.893 for the chain and 0.905 for the lattice, rounded to three significant figures. These weights are calibrated using nine references with 5,000 samples each. Shared training and both local fits use the same loss within each geometry.

For both geometries, training uses 20,000 target bitstrings per reference and 1,024 MC draws per estimated model coefficient. The training set starts with nine points and adds 8, 8, 8, 8 and 9 points in successive stages. The first five stages take 200 updates each and the final stage takes 500. Each update averages gradients over the available reference points and uses fresh Fourier supports and model MC draws. We evaluate the final model after 1,500 updates.

After exact gate cancellation, each observable calculation uses a proper subset of the circuit sites. Uniform even-parity bitstrings on the full register have a uniform marginal on any such subset, since an omitted bit can complete either parity. We can therefore draw uniform bitstrings on the remaining sites while preserving the GHZ expectation.

For the response maps in Figure 8(a) and Figure 9(a), we evaluate the 992 nonconstant even-order coefficients on a 40×4040 \times40 grid, using 1,000 MC draws per coefficient at each point. We keep these draws fixed across the grid and differentiate through the coefficient estimates, marginal projection and smoothing to obtain the selected-marginal susceptibility in Eq. (34).

Local refinement and held-out evaluation

The local evaluation points are (h1/J,h2/J)=(1.46,−1.40)(h_1/J, h_2/J) = (1.46, -1.40) for the chain and (1.22,−0.04)(1.22, -0.04) for the lattice. At each point, both local fits use the same 20,000 bitstrings from the reference MPS and take 600 Adam updates. Fine-tuning starts from the final shared model and updates its base angles and hypernetwork at learning rate 0.0003. Independent training optimizes the circuit angles at learning rate 0.003.

For panel (b) of each figure, we draw one million held-out bitstrings from the MPS that supplied the local training samples. We divide them into 40 groups of 25,000 for the chain and 20 groups of 50,000 for the lattice. Each bandwidth uses 400,000 Gaussian-weighted Fourier support draws in total. Each lattice group has 20,000 draws. The chain assigns 10,016 draws to half the groups and 9,984 to the rest, preserving batches of 32 supports. Each sampled model coefficient uses 100,000 auxiliary MC draws, shared within each batch and across the three models. Squared coefficients use distinct pairs of draws, and the cross term uses independent model and target samples.

The black curve compares the 20,000 local training samples with the held-out samples by calculating the Gaussian kernel directly on the bitstrings. We split the training sample into groups of 500 for the chain and 1,000 for the lattice, pair each with its corresponding held-out group and exclude self-pairs from the within-sample sums. All curves show the unweighted mean over groups. Error bars give the standard error, the sample standard deviation across groups divided by the square root of the number of groups. They measure evaluation uncertainty for fixed trained models.

Rectangular-lattice training and evaluation

Figure 9 uses a 4 by 256 rectangular lattice with open boundaries. The Hamiltonian contains −h1Xi-h_1X_i at each site and −h2XiXj-h_2X_iX_j on each nearest-neighbor edge. Sites are ordered along the four-site direction first. For each edge (i,j)(i,j) with i<ji < j, the Hamiltonian also contains −JZiZjZk-JZ_iZ_jZ_k for every neighbor kk of jj at graph distance two from ii. We use this same site ordering for the MPS, whose contraction cost depends on entanglement and the range of interactions [77].

Each reference uses two initial states in the global-X=+1X = +1 sector and retains the lower-energy DMRG result. The bond cap is 128, with four to twelve sweeps, cutoff 10−810^{-8}, total-energy stopping tolerance 10−510^{-5} and local eigensolver tolerance 10−810^{-8}. We transform back to the physical computational basis before sampling.

The lattice uses the same circuit and training schedule as the chain. Its 32 six-site marginals are 2×32 \times3 rectangles distributed along the long direction of the lattice. The largest detected response peak on the displayed map is at (1.26,0.04)(1.26, 0.04).

Additional numerical comparisons

Comparisons of loss and gate restrictions

Table G1 extends the comparison in Figure 6 to additional training objectives. All infidelities are evaluated from exact model probabilities on the same nine-site grid. Gaussian-sampled MMD gives a small reduction in mean infidelity relative to order-four MMD, while the combined CDF-KL and MMD loss gives a larger error. The Gaussian-sampled objective includes higher even orders at a matched number of coefficient draws. Exact and MC training give similar errors for each objective.

Figure G1. Haldane susceptibility peaks with different gate restrictions. Adaptive CDF-KL fits use (a) r=10r = 10, (b) r=4r = 4 and (c) r=2r = 2, with the same angle count, MC budget and order-four loss. Learned peaks are compared with exact nine-site peaks and the bulk guides [29]. Points show each fit’s nine initial and 40 acquired references.

Figure G1 compares the learned susceptibility peaks for three gate restrictions. All three fits recover parts of the two response branches. The r=4r = 4 fit develops an upper branch that rises away from the reference at larger h1/Jh_1/J, while the r=2r = 2 fit follows the descending reference branch more closely. The two restricted fits have nearly equal mean infidelity, about 0.107, despite these differences in their peaks. Each fit selects its own acquisition points, so the comparison includes changes in both the circuit and the selected data.

Other nine-site targets

We apply the fixed 10 × 10 protocol of Figure 3 to the alternating Heisenberg chain, the axial next-nearest-neighbor Ising (ANNNI) chain, and a cluster chain with Ising interactions and a transverse field. These Hamiltonian families also appear in the phase-classification benchmarks of [78]. A separate hypernetwork is trained for each target.

With XjX_j, YjY_j, ZjZ_j denoting Pauli operators on site jj, the Hamiltonians are

HHeis=14∑j=1n−1Jj(XjXj+1+YjYj+1+ZjZj+1),(97)H_{\mathrm{Heis}}=\frac{1}{4}\sum_{j=1}^{n-1}J_j(X_jX_{j+1}+Y_jY_{j+1}+Z_jZ_{j+1}), \tag*{(97)}
HANNNI=−∑j=1n−1XjXj+1+κ∑j=1n−2XjXj+2−h∑j=1nZj,(98)H_{\mathrm{ANNNI}}=-\sum_{j=1}^{n-1}X_jX_{j+1}+\kappa\sum_{j=1}^{n-2}X_jX_{j+2}-h\sum_{j=1}^{n}Z_j, \tag*{(98)}
HC=∑j=1n(Zj−J1XjXj+1−J2XjZj+1Xj+2).(99)H_{\mathrm{C}}=\sum_{j=1}^{n}(Z_j-J_1X_jX_{j+1}-J_2X_jZ_{j+1}X_{j+2}). \tag*{(99)}

In Eq. (97), Jj=JoddJ_j=J_{\mathrm{odd}} on odd bonds and Jj=JevenJ_j=J_{\mathrm{even}} on even bonds. Both couplings are antiferromagnetic, and the factor 1/4 converts Pauli products to spin-1/2 products. In Eq. (98), the nearest-neighbor ferromagnetic coupling sets the energy unit, κ≥0\kappa\ge0 is the competing antiferromagnetic next-nearest-neighbor coupling, and hh is the transverse field in the same units. In Eq. (99), the transverse-field coefficient is fixed to +1+1, J1J_1 controls the two-site Ising interaction, and J2J_2 controls the three-site cluster interaction XjZj+1Xj+2X_jZ_{j+1}X_{j+2}. The ANNNI parameters κ,h\kappa,h and the cluster parameters J1,J2J_1,J_2 are dimensionless.

Heisenberg and ANNNI have open boundaries, while cluster indices are taken modulo nn. The parameter domains are (Jodd,Jeven)∈[0.25,1]2(J_{\mathrm{odd}},J_{\mathrm{even}})\in[0.25,1]^2, (κ,h)∈[0,1.5]×[0,2](\kappa,h)\in[0,1.5]\times[0,2] and (J1,J2)∈[−2,2]2(J_1,J_2)\in[-2,2]^2. The Heisenberg reference is selected in the global-X=+1X=+1 sector. At degeneracies we use a fixed seeded ground-state selection and omit undefined responses.

Dashed gray curves in Figure G2–Figure G7 show bulk phase boundaries, distinct from the nine-site susceptibility peaks. For the alternating Heisenberg chain, the bulk critical line is Jeven=JoddJ_{\mathrm{even}}=J_{\mathrm{odd}}. The ANNNI guides follow the phase diagram used in Refs. [78, 79]. We use the approximate Ising boundary

Figure G2. Learned susceptibility for the nine-site open alternating Heisenberg chain. Panels show (a) the exact target, (b) Haldane QCBM, (c) RBM, (d) IQP, (e) Heisenberg QCBM and (f) Fermionic QCBM. The response is χ22\chi_{22} with respect to JevenJ_{\mathrm{even}}. The dashed line marks Jeven=JoddJ_{\mathrm{even}} = J_{\mathrm{odd}}, and the inset shows the training grid.

Figure G7. Restricted training for the periodic cluster target. Panels and reference markers follow Figure G3. The response is taken with respect to J2J_2, and coefficient losses include orders one through four. Learned peaks are compared with exact nine-site peaks and the bulk boundaries.

hI(κ)≃2−4κ1+(1−3κ+4κ2)/(1−κ),(100)h_{\mathrm{I}}(\kappa)\simeq\frac{2-4\kappa}{1+\sqrt{(1-3\kappa+4\kappa^2)/(1-\kappa)}}, \tag*{(100)}

for 0≤κ≤1/20\le\kappa\le1/2, and the fitted Kosterlitz–Thouless (KT) and Pokrovsky–Talapov (PT) boundaries for κ>1/2\kappa>1/2 [80],

hKT(κ)≃1.05(κ−12)(κ−0.1),(101)h_{\mathrm{KT}}(\kappa)\simeq1.05\sqrt{\left(\kappa-\frac{1}{2}\right)(\kappa-0.1)}, \tag*{(101)}
hPT(κ)≃1.05(κ−12).(102)h_{\mathrm{PT}}(\kappa)\simeq1.05\left(\kappa-\frac{1}{2}\right). \tag*{(102)}

The KT and PT curves bound the floating phase above and below, respectively.

For the cluster Hamiltonian, the bulk single-particle excitation energy can be written as ε(k)=2∣1+J1eik+J2e2ik∣\varepsilon(k)=2\lvert1+J_1e^{ik}+J_2e^{2ik}\rvert, where kk is crystal momentum and the lattice spacing is one. Its zeros give

J2=−1±J1,J2=1 with ∣J1∣≤2.(103)J_2=-1\pm J_1,\qquad J_2=1\text{ with }\lvert J_1\rvert\le2. \tag*{(103)}

These are the boundaries used in Ref. [78], consistent with the free-fermion solution after accounting for the transverse-field sign [81].

For each target, the architecture comparison uses 100 training points with 5,000 samples each, 24 separate validation points and a 40×4040\times40 evaluation grid. Each model uses one initialization, 1,500 likelihood updates and checkpoint selection by validation likelihood. Model budgets and ancilla locations follow Figure 3. Circuit preparation is GHZ for Heisenberg and the computational zero state for ANNNI and cluster. The RBM enforces bit-complement symmetry, which generally does not hold for the ANNNI and cluster target distributions in this basis. All susceptibilities are evaluated from exact model probabilities, with derivatives taken along the vertical parameter axis.

The restriction studies use Hadamard-free QCBMs for all three targets and follow Figure 6. They compare 100- and 49-point likelihood fits, exact and MC coefficient evaluation, fixed and adaptive 49-point training, and gates restricted to r=2r=2. The coefficient losses use orders through four, retaining only even orders for Heisenberg and odd and even orders for ANNNI and cluster. All fits use 1,500 updates, with 64 MC draws per nontrivial coefficient when required. The circuits retain three unmeasured ancillas, 298 angles and 9,581 hypernetwork weights. Each adaptive fit selects 40 new references from nine initial points.

Alternating Heisenberg chain

In Figure G2, the Haldane and Heisenberg QCBMs closely reproduce the broad susceptibility band of the exact target. Both have mean infidelity about 0.0020, compared with 0.0154 for the Fermionic QCBM, 0.158 for IQP and 0.208 for the RBM. IQP places a much sharper band near the diagonal, while the RBM introduces a pronounced vertical feature absent from the target.

Figure G3 shows that reducing the likelihood grid from 100 to 49 points leaves the response map nearly unchanged. Order-four training also follows the main sequence of finite-size peaks, with close agreement between exact and MC MMD fits. Fixed and adaptive CDF-KL have similar mean infidelities, 0.0202 and 0.0208. The additional gate restriction to r=2r = 2 produces extra peak branches and increases the error to 0.310.

Figure G3. Restricted training for the alternating Heisenberg target. Haldane QCBMs use (a.i,ii) likelihood on 10×1010 \times10 and 7×77 \times7 grids, (b) order-four MMD with exact or MC coefficients, (c.i,ii) fixed or adaptive CDF-KL and MMD, and (d) adaptive CDF-KL with r=2r = 2. Curves compare learned and exact nine-site susceptibility peaks with the bulk guide. Points mark initial and acquired references.

ANNNI chain

The four circuit models in Figure G4 give similar mean infidelities, between 0.0423 and 0.0431. They reproduce the broad response at larger fields, but replace the target’s sharper low-field features by a nearly horizontal band. The RBM error is much larger, 0.641, with the symmetry restriction described above.

Figure G4. Learned susceptibility for the nine-site open ANNNI chain. Panels follow Figure G2, with the response taken with respect to hh. The inset shows the training grid. Dashed curves are the approximate bulk boundaries in Eqs. (100)–(102).

In Figure G5, exact and MC MMD again give nearly identical peak locations. The fitted peaks follow the descending reference branch at small κ\kappa. Beyond κ=0.5\kappa= 0.5, the fits recover only parts of the reference structure, and gaps in the two branches persist after adaptive training. Reducing the likelihood grid raises mean infidelity from 0.0430 to 0.0586, while order-four MMD gives 0.221. Here the loss restriction has a larger effect than replacing exact coefficients by MC estimates.

Figure G5. Restricted training for the ANNNI target. Panels and reference markers follow Figure G3. The response is taken with respect to hh, and coefficient losses include orders one through four. Learned peaks are compared with exact nine-site peaks and the approximate bulk boundaries.

Periodic cluster chain

For the cluster target in Figure G6, the Haldane, Heisenberg and Fermionic QCBMs have mean infidelities between 0.0823 and 0.0903, compared with 0.154 for IQP and 0.663 for the symmetry-restricted RBM. The circuit maps resolve the broad response near J2=1J_2 = 1. The exact target has additional narrow features along the upper branch, while a horizontal band near J2=−0.8J_2 = -0.8 appears only in the circuit maps.

Figure G6. Learned susceptibility for the nine-site periodic cluster chain. Panels follow Figure G2, with the response taken with respect to J2J_2. The inset shows the training grid. Dashed lines mark the bulk boundaries in Eq. (103).

Figure G7 shows little change between the two likelihood grids. Order-four MMD recovers parts of the upper sequence of peaks and the ascending branch at positive J1J_1, with similar results from exact and MC coefficients. Its mean infidelity is nevertheless about 0.437, compared with 0.0882 for the 49-point likelihood fit. Adaptive and restricted-gate fits show larger changes in peak locations.

References

  1. [1]J.-G. Liu and L. Wang, Differentiable Learning of Quantum Circuit Born Machine, Physical Review A 98, 062324 (2018), arXiv:1804.04168 [quant-ph].DOI
  2. [2]M. Benedetti, D. Garcia-Pintos, O. Perdomo, V. Leyton-Ortega, Y. Nam, and A. Perdomo-Ortiz, A generative modeling approach for benchmarking and training shallow quantum circuits, npj Quantum Information 5, 45 (2019).DOI
  3. [3]B. Coyle, D. Mills, V. Danos, and E. Kashefi, The Born Supremacy: Quantum Advantage and Training of an Ising Born Machine, npj Quantum Information 6, 60 (2020), arXiv:1904.02214 [quant-ph].DOI
  4. [4]M. J. Bremner, A. Montanaro, and D. J. Shepherd, Average-Case Complexity Versus Approximate Simulation of Commuting Quantum Computations, Physical Review Letters 117, 080501 (2016).DOI
  5. [5]A. Bouland, B. Fefferman, C. Nirkhe, and U. Vazirani, On the complexity and verification of quantum random circuit sampling, Nature Physics 15, 159 (2019).DOI
  6. [6]M. J. Bremner, A. Montanaro, and D. J. Shepherd, Achieving quantum supremacy with sparse and noisy commuting quantum computations, Quantum 1, 8 (2017), arXiv:1610.01808 [quant-ph].DOI
  7. [7]F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, R. Barends, R. Biswas, S. Boixo, F. G. S. L. Brandao, D. A. Buell, B. Burkett, Y. Chen, Z. Chen, B. Chiaro, R. Collins, W. Courtney, A. Dunsworth, E. Farhi, B. Foxen, A. Fowler, C. Gidney, M. Giustina, R. Graff, K. Guerin, S. Habegger, M. P. Harrigan, M. J. Hartmann, A. Ho, M. Hoffmann, T. Huang, T. S. Humble, S. V. Isakov, E. Jeffrey, Z. Jiang, D. Kafri, K. Kechedzhi, J. Kelly, P. V. Klimov, S. Knysh, A. Korotkov, K. Kostritsa, D. Landhuis, M. Lindmark, E. Lucero, D. Lykov, S. Mandra, J. R. McClean, M. McEwen, A. Megrant, X. Mi, K. Michielsen, M. Mohseni, J. Mutus, O. Naaman, M. Neeley, C. Neill, M. Y. Niu, E. Ostby, A. Petukhov, J. C. Platt, C. Quintana, E. G. Rieffel, P. Roushan, N. C. Rubin, D. Sank, K. J. Satzinger, V. Smelyanskiy, K. J. Sung, M. D. Trevithick, A. Vainsencher, B. Villalonga, T. White, Z. J. Yao, P. Yeh, A. Zalcman, H. Neven, and J. M. Martinis, Quantum supremacy using a programmable superconducting processor, Nature 574, 505 (2019).DOI
  8. [8]D. Bluvstein, S. J. Evered, A. A. Geim, S. H. Li, H. Zhou, T. Manovitz, S. Ebadi, M. Cain, M. Kali-nowksi, D. Hangleiter, J. P. Bonilla Ataides, N. Maskara, I. Cong, X. Gao, P. Sales Rodriguez, T. Karolyshyn, G. Semeghini, M. J. Gullans, M. Greiner, V. Vuletic, and M. D. Lukin, Logical quantum processor based on reconfigurable atom arrays, Nature 626, 58 (2024).
  9. [9]E. Armengol and J. Bowles, IQPop: Fast optimization of instantaneous quantum polynomial circuits in JAX (2026), arXiv:2501.04776 [quant-ph].arxiv.org/abs/2501.04776
  10. [10]E. Recio-Armengol, S. Ahmed, and J. Bowles, Train on classical, deploy on quantum: Scaling generative quantum machine learning to a thousand qubits (2026), arXiv:2503.02934 [quant-ph].
  11. [11]J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Barren plateaus in quantum neural network training landscapes, Nature Communications 9, 4812 (2018).DOI
  12. [12]M. Cerezo, A. Sone, T. Volkoff, L. Cincio, and P. J. Coles, Cost Function Dependent Barren Plateaus in Shallow Parametrized Quantum Circuits, Nature Communications 12, 1791 (2021), arXiv:2001.00550 [quant-ph].DOI
  13. [13]Z. Holmes, K. Sharma, M. Cerezo, and P. J. Coles, Connecting ansatz expressibility to gradient magnitudes and barren plateaus, PRX Quantum 3, 010313 (2022), arXiv:2101.02138 [quant-ph].DOI
  14. [14]M. Cerezo, M. Larocca, D. García-Martín, N. L. Diaz, P. Braccia, E. Fontana, M. S. Rudolph, P. Bermejo, A. Ijaz, S. Thanaslip, E. R. Anschuetz, and Z. Holmes, Does provable absence of barren plateaus imply classical simulability?, Nature Communications 16, 7907 (2025).DOI
  15. [15]S. Herbst, I. Brandić, and A. Pérez-Salinas, Limits of quantum generative models with classical sampling hardness (2025), arXiv:2512.24801 [quant-ph].arxiv.org/abs/2512.24801
  16. [16]E. Gil-Fuster, C. Gyurik, A. Pérez-Salinas, and V. Dunjko, On the relation between trainability and dequantization of variational quantum learning models (2025), arXiv:2406.07072 [quant-ph].
  17. [17]K. Shen, S. Pielawa, V. Dunjko, and H. Wang, Characterizing the Trainability of Instantaneous Quantum Polynomial Circuit Born Machines (2026), arXiv:2602.11042 [quant-ph].arxiv.org/abs/2602.11042
  18. [18]S. Lerch, J. Bowles, R. Puig, E. Armengol, Z. Holmes, and S. Thanaslip, IQP Born Machines under Data-dependent and Agnostic Initialization Strategies (2026), arXiv:2603.14576 [quant-ph].arxiv.org/abs/2603.14576
  19. [19]M. Herrero-Gonzalez, B. Coyle, K. McDowall, R. Grassie, S. Beentjes, A. Khamsi, and E. Kashefi, The Born Ultimatum: Conditions for Classical Surrogation of Quantum Generative Models with Correlators (2025), arXiv:2511.01845 [quant-ph].
  20. [20]G. Carleo and M. Troyer, Solving the Quantum Many-Body Problem with Artificial Neural Networks, Science 355, 602 (2017), arXiv:1606.02318 [cond-mat.dis-nn].DOI
  21. [21]G. Torlai, G. Mazzola, J. Carrasquilla, M. Troyer, R. Melko, and G. Carleo, Many-body quantum state tomography with neural networks, Nature Physics 14, 447 (2018), arXiv:1703.05334 [cond-mat.dis-nn].DOI
  22. [22]S. Tonner, V. T. Tran, and R. Kueng, Parametric Quantum State Tomography with HyperRBMs (2026), arXiv:2601.20950 [quant-ph].arxiv.org/abs/2601.20950
  23. [23]R. Rende, L. L. Viteritti, F. Becca, A. Scardicchio, A. Laio, and G. Carleo, Foundation Neural-Networks Quantum States as a Unified Ansatz for Multiple Hamiltonians, Nature Communications 16, 7213 (2025), arXiv:2502.09488 [quant-ph].DOI
  24. [24]H.-Y. Huang, R. Kueng, G. Torlai, V. V. Albert, and J. Preskill, Provably efficient machine learning for quantum many-body problems, Science 377, eabk3333 (2022), arXiv:2106.12627 [quant-ph].DOI
  25. [25]L. Lu, H.-Y. Huang, V. T. Tran, S. Lehner, R. Kueng, and J. Preskill, Improved machine learning algorithm for predicting ground state properties, Nature Communications 15, 895 (2024).
  26. [26]O. Kiss, M. Grossi, E. Kajomovitz, and S. Vallecorsa, Conditional Born machine for Monte Carlo event generation, Physical Review A 106, 022612 (2022), arXiv:2205.07674 [quant-ph].DOI
  27. [27]A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. Smola, A Kernel Two-Sample Test, Journal of Machine Learning Research 13, 723 (2012).
  28. [28]S. Zhou, Y. Zhang, S. Wetzel, R. Melko, and X.-Z. Luo, Autonomous phase discovery (2026), arXiv:2609.33802 [quant-ph].arxiv.org/abs/2609.33802
  29. [29]I. Cong, S. Choi, and M. D. Lukin, Quantum Convolutional Neural Networks, Nature Physics 15, 1273 (2019), arXiv:1810.03787 [quant-ph].DOI
  30. [30]F. D. M. Haldane, Model for a Quantum Hall Effect without Landau Levels: Condensed-Matter Realization of the "Parity Anomaly", Physical Review Letters 61, 2015 (1988).pubmed.ncbi.nlm.nih.gov/10038961
  31. [31]U. Schollwock, The density-matrix renormalization group in the age of matrix product states, Annals of Physics 326, 96 (2011), arXiv:1008.3477 [cond-mat.str-el].
  32. [32]M. Fishman, S. R. White, and E. M. Stoudenmire, The ITensor Software Library for Tensor Network Calculations (2021), arXiv:2007.14822 [cs.MS].DOI
  33. [33]D. Ha, A. Dai, and Q. V. Le, HyperNetworks (2016), arXiv:1609.09106 [cs.LG].arxiv.org/abs/1609.09106
  34. [34]V. K. Chauhan, J. Zhou, P. Lu, S. Molaei, and D. A. Clifton, A brief review of hypernetworks in deep learning, Artificial Intelligence Review 57, 250 (2024).DOI
  35. [35]J. Miao, C.-Y. Hsieh, and S.-X. Zhang, Neural-network-encoded variational quantum algorithms, Physical Review Applied 21, 014053 (2024).DOI
  36. [36]J. Qi, C.-H. H. Yang, P.-Y. Chen, and M.-H. Hsieh, TensorHyper-VQC: A tensor-trained guided hypernetwork for robust and scalable variational quantum computing, npj Quantum Information 12, 70 (2026).DOI
  37. [37]E. Perez, F. Strub, H. de Vries, V. Dumoulin, and A. Courville, FiLM: Visual Reasoning with a General Conditioning Layer (2017), arXiv:1709.07871 [cs.CV].arxiv.org/abs/1709.07871
  38. [38]A. Kurkin, K. Shen, S. Pielawa, H. Wang, and V. Dunjko, Universal and kernel-adaptive training for classically trained, quantum-deployed generative models (2025), arXiv:2510.08476 [quant-ph].arxiv.org/abs/2510.08476
  39. [39]E. Kökcü, T. Steckmann, Y. Wang, J. K. Freericks, E. F. Dumitrescu, and A. F. Kemper, Fixed Depth Hamiltonian Simulation via Cartan Decomposition, Physical Review Letters 129, 070501 (2022), arXiv:2104.00728 [quant-ph].DOI
  40. [40]R. Wiersema, E. Kökcü, A. F. Kemper, and B. N. Bakalov, Classification of dynamical Lie algebras of 2-local spin systems on linear, circular and fully connected topologies, npj Quantum Information 10, 110 (2024).
  41. [41]H.-Y. Huang, R. Kueng, and J. Preskill, Predicting Many Properties of a Quantum System from Very Few Measurements, Nature Physics 16, 1050 (2020), arXiv:2002.08953 [quant-ph].DOI
  42. [42]M. V. den Nest, Simulating quantum computers with probabilistic methods (2010), arXiv:0911.1624 [quant-ph].arxiv.org/abs/0911.1624
  43. [43]C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, Rethinking the Inception Architecture for Computer Vision (2015), arXiv:1512.00567 [cs.CV].arxiv.org/abs/1512.00567
  44. [44]M. S. Rudolph, L. Schär, S. Thanasilp, O. Shaya, S. Vallecorsa, M. Grossi, and Z. Holmes, Trainability barriers and opportunities in quantum generative modeling, npj Quantum Information 10, 116 (2024).DOI
  45. [45]L. Tao, S. Xu, C.-H. Wang, N. Suh, and G. Cheng, Discriminative Estimation of Total Variation Distance: A Fidelity Auditor for Generative Data (2024), arXiv:2405.15337 [stat.ML].arxiv.org/abs/2405.15337
  46. [46]C. A. Fuchs and J. van de Graaf, Cryptographic Distinguishability Measures for Quantum Mechanical States (1998), arXiv:quant-ph/9712042.DOI
  47. [47]P. Zanardi and N. Paunković, Ground state overlap and quantum phase transitions, Physical Review E 74, 031123 (2006), arXiv:quant-ph/0512249.DOI
  48. [48]S.-J. Gu, Fidelity approach to quantum phase transitions, International Journal of Modern Physics B 24, 4371 (2010), arXiv:0811.3127 [quant-ph].DOI
  49. [49]M. Prokopenko, J. T. Lizier, O. Obst, and X. R. Wang, Relating Fisher information to order parameters, Physical Review E 84, 041116 (2011).pubmed.ncbi.nlm.nih.gov/22181096
  50. [50]L. Wang, Discovering Phase Transitions with Unsupervised Learning, Physical Review B 94, 195105 (2016), arXiv:1606.00318 [cond-mat.stat-mech].DOI
  51. [51]J. Carrasquilla and R. G. Melko, Machine learning phases of matter, Nature Physics 13, 431 (2017), arXiv:1605.01735 [cond-mat.str-el].DOI
  52. [52]E. P. L. van Nieuwenburg, Y.-H. Liu, and S. D. Huber, Learning phase transitions by confusion, Nature Physics 13, 435 (2017).DOI
  53. [53]F. Schäfer and N. Lörch, Vector field divergence of predictive model output as indication of phase transitions, Physical Review E 99, 062107 (2019).DOI
  54. [54]J. Arnold, F. Schäfer, M. Żonda, and A. U. J. Lode, Interpretable and unsupervised phase classification, Physical Review Research 3, 033052 (2021).DOI
  55. [55]J. Arnold and F. Schäfer, Replacing neural networks by optimal analytical predictors for the detection of phase transitions, Physical Review X 12, 031044 (2022), arXiv:2203.06084 [cond-mat.dis-nn].DOI
  56. [56]J. Arnold, F. Schäfer, A. Edelman, and C. Bruder, Mapping out phase diagrams with generative classifiers, Physical Review Letters 132, 207301 (2024), arXiv:2306.14894 [quant-ph].DOI
  57. [57]J. Arnold, N. Lörch, F. Holtorf, and F. Schäfer, Machine learning phase transitions: Connections to the Fisher information (2023), arXiv:2311.10710 [cond-mat.dis-nn].arxiv.org/abs/2311.10710
  58. [58]V. Kasatkin, E. Mozgunov, N. Ezzell, U. Mishra, I. Hen, and D. Lidar, ClassiFIM: An Unsupervised Method To Detect Phase Transitions (2024), arXiv:2408.03323 [cs.LG].arxiv.org/abs/2408.03323
  59. [59]B. Coyle, M. Henderson, J. C. J. Le, N. Kumar, M. Paini, and E. Kashefi, Quantum versus Classical Generative Modelling in Finance (2020), arXiv:2008.00691 [quant-ph].arxiv.org/abs/2008.00691
  60. [60]J. Herrmann, S. M. L. Lima, A. Remm, P. Zapletal, N. A. McMahon, C. Scarato, F. Swiadek, C. K. Andersen, C. Hellings, S. Krinner, N. Lacroix, S. Lazar, M. Kerschbaum, D. C. Zanuz, G. J. Norris, M. J. Hartmann, A. Wallraff, and C. Eichler, Realizing quantum convolutional neural networks on a superconducting quantum processor to recognize quantum phases, Nature Communications 13, 4144 (2022).DOI
  61. [62]S. J. Wetzel, Unsupervised learning of phase transitions: From principal component analysis to variational autoencoders, Physical Review E 96, 022140 (2017), arXiv:1703.02435 [cond-mat.stat-mech].DOI
  62. [63]P. de Schoulepnikoff, H. P. Nautrup, H. J. Briegel, and G. Muñoz-Gil, Discovering quantum phenomena with Interpretable Machine Learning (2026), arXiv:2604.16015 [quant-ph].arxiv.org/abs/2604.16015
  63. [64]K. Gili, M. Mauri, and A. Perdomo-Ortiz, Generalization Metrics for Practical Quantum Advantage in Generative Models (2023), arXiv:2201.08770 [cs.LG].arxiv.org/abs/2201.08770
  64. [65]K. Gili, M. Hibat-Allah, M. Mauri, C. Ballance, and A. Perdomo-Ortiz, Do Quantum Circuit Born Machines Generalize? (2023), arXiv:2207.13645 [quant-ph].arxiv.org/abs/2207.13645
  65. [66]S. Raj, N. Mathur, and A. Perdomo-Ortiz, "Train classical, deploy quantum" requires rethinking generalization (2026), arXiv:2608.31117 [quant-ph].arxiv.org/abs/2608.31117
  66. [67]L. Tagliazzo, T. R. de Oliveira, S. Ibsdir, and J. I. Latorre, Scaling of entanglement support for Matrix Product States, Physical Review B 78, 024410 (2008), arXiv:0712.1976 [cond-mat.staech].
  67. [68]A. Kulshrestha, R. Puig, D. García-Martín, L. Cincio, I. Safro, Z. Holmes, and M. Cerezo, Exponentially many initializations to avoid barren plateaus (2026), arXiv:2606.18515 [quant-ph].arxiv.org/abs/2606.18515
  68. [69]A. Huang, W. Maxwell, V. Belis, E. Peters, J. Pye, S. Jahangiri, and J. Bowles, Spectral Born machines: Classically trainable quantum generative models for discrete data (2026), arXiv:2607.06675 [quant-ph].arxiv.org/abs/2607.06675
  69. [70]A. J. Ferris and G. Vidal, Perfect Sampling with Unitary Tensor Networks, Physical Review B 85, 165146 (2012), arXiv:1201.3974 [cond-mat.str-el].DOI
  70. [71]B. M. Terhal and D. P. DiVincenzo, Classical simulation of noninteracting-fermion quantum circuits, Physical Review A 65, 032325 (2002), arXiv:quant-ph/0108010.DOI
  71. [72]R. Jozsa and A. Miyake, Matchgates and classical simulation of quantum circuits, Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 464, 3089 (2008), arXiv:0804.4050 [quant-ph].DOI
  72. [73]B. Dias and R. Koenig, Classical simulation of non-Gaussian fermionic circuits, Quantum 8, 1350 (2024), arXiv:2307.12912 [quant-ph].DOI
  73. [74]B. Bakó, Z. Kolarovszki, and Z. Zimbóras, Fermionic Born Machines: Classical training of quantum generative models based on Fermion Sampling (2025), arXiv:2511.13844 [quant-ph].arxiv.org/abs/2511.13844
  74. [75]D. P. Kingma and J. Ba, Adam: A Method for Stochastic Optimization (2017), arXiv:1412.6980 [cs.LG].
  75. [76]D. Perez-Garcia, F. Verstraete, M. M. Wolf, and J. I. Cirac, Matrix Product State Representations (2007), arXiv:quant-ph/0608197.
  76. [77]R. Orus, A Practical Introduction to Tensor Networks: Matrix Product States and Projected Entangled Pair States, Annals of Physics 349, 117 (2014), arXiv:1306.2164 [cond-mat.str-el].DOI
  77. [78]P. Bermejo, P. Braccia, M. S. Rudolph, Z. Holmes, L. Cincio, and M. Cerezo, Quantum Convolutional Neural Networks are Effectively Classically Simulable, PRX Quantum 7, 020304 (2026).arxiv.org/abs/2408.12739
  78. [79]M. Cea, M. Grossi, S. Monaco, E. Rico, L. Taglia- cozzo, and S. Vallecorsa, Exploring the Phase Diagram of the quantum one-dimensional ANNNI model (2026), arXiv:2402.10122 [cond-mat.str-el].
  79. [80]M. Beccaria, M. Campostrini, and A. Feo, Evidence for a floating phase of the transverse ANNNI model at high frustration, Physical Review B 76, 094410 (2007).DOI
  80. [81]Y. Niu, S. B. Chung, C.-H. Hsu, I. Mandal, S. Raghu, and S. Chakravarty, Majorana zero modes in a quantum Ising chain with longer-ranged interactions, Physical Review B 85, 035110 (2012).DOI
  81. [82]B. Bakó, D. T. R. Nagy, P. Hága, Zs. Kallus, and Z. Zimborás, Problem-informed graphical quantum generative learning, Quantum Science and Technology 11, 035012 (2026), arXiv:2405.14072 [quant-ph].DOI
  82. [83]Z. Kolarovszki, B. Bakó, M. Oszmaniec, C. Oh, and Z. Zimborás, Generative modeling with Gaussian Boson Sampling: classically trainable Bosonic Born Machines (2026), arXiv:2603.11195 [quant-ph].arxiv.org/abs/2603.11195
  83. [84]A. Kurkin, U. Chabaud, Z. Kolarovszki, B. Bakó, Z. Zimborás, and V. Dunjko, Universality of Classically Trainable, Quantum-Deployed Boson-Sampling Generative Models (2026), arXiv:2603.11014 [quant-ph].arxiv.org/abs/2603.11014

Paper details

vlnc:W_nf86f549
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

GPT-6 Astra (bare)

View statement
Looking for related papers…
Contents