Spectral Born Machines: Quantum Fourier Models

• Spectral Born Machines are quantum generative models defined by a Fourier sandwich circuit that uses the Born rule to generate complex, discrete data distributions.
• They apply group Fourier analysis and maximum mean discrepancy objectives to classically estimate quantum correlators and manage spectral training.
• Architectural choices and spectral truncation techniques in SBMs balance trainability with classical optimization, addressing issues like barren plateaus in quantum circuits.

Spectral Born Machines (SBMs) are a class of quantum generative models derived by viewing IQP Born Machines through the lens of group Fourier analysis and then generalizing from bitstrings to integer-structured data over $\mathbb{Z}_d$. They define a model distribution by the Born rule, use a “Fourier sandwich” circuit $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$, and train through Fourier- or graph-spectral maximum mean discrepancy objectives that are classically estimable. In the binary setting, the same perspective identifies Quantum Circuit Born Machines (QCBMs) as quantum Fourier models whose output distributions are exactly Walsh–Hadamard transforms of correlators, making the IQP Born Machine the $d=2$ special case of the broader SBM framework (Huang et al., 7 Jul 2026, Herrero-Gonzalez et al., 3 Nov 2025).

1. Formal definition and circuit structure

Given a parameterized quantum circuit $U(\theta)$ acting on an initial computational basis state $|0\rangle$, the output probability of integer-valued datum $x$ is

$p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$

i.e., the Born rule. SBMs define a model distribution $q_\theta(x)$ as the computational-basis measurement distribution of the evolved state,

$$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$

The canonical SBM circuit is a “Fourier sandwich” over $\mathbb{Z}_d^n$,

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$0

where $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$1 is the quantum Fourier transform over $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$2 and $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$3 is a diagonal phase unitary in the computational basis. For $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$4,

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$5

and for $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$6,

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$7

The Fourier basis is spanned by characters $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$8 (Huang et al., 7 Jul 2026).

In the computational basis, the diagonal layer satisfies

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$9

with phase function $d=2$0 computable in $d=2$1. A particularly useful class is generated by Heisenberg–Weyl observables,

$d=2$2

where each generator contributes a separable phase eigenvalue

$d=2$3

so that

$d=2$4

Restricting $d=2$5 by weight and degree yields strong spectral bias and reduces parameter count. The binary limit $d=2$6 recovers the IQP Born Machine, because SBMs reduce to IQP BM when $d=2$7 and $d=2$8 (Huang et al., 7 Jul 2026).

The same structure admits an equivalent Walsh–Hadamard description for qubit QCBMs. If $d=2$9 is the computational-basis measurement distribution, then

$U(\theta)$0

and

$U(\theta)$1

Thus QCBMs are spectral models whose parameters control a trigonometric–Pauli polynomial whose $U(\theta)$2-string coefficients are directly the Fourier/Walsh–Hadamard components of the output distribution (Herrero-Gonzalez et al., 3 Nov 2025).

2. Spectral representation and MMD training

The defining training perspective of SBMs is spectral. For distributions $U(\theta)$3 on $U(\theta)$4 and a translation-invariant kernel $U(\theta)$5 with Walsh–Hadamard decomposition

$U(\theta)$6

the maximum mean discrepancy diagonalizes as

$U(\theta)$7

In the IQP-QCBM notation,

$U(\theta)$8

where $U(\theta)$9 is the characteristic function and $|0\rangle$0 (Shen et al., 11 Feb 2026).

For general SBMs on $|0\rangle$1, the kernel is chosen through graph spectral analysis. Each site $|0\rangle$2 is assigned a weighted graph $|0\rangle$3 encoding the appropriate notion of label closeness, with Laplacian $|0\rangle$4. The product graph $|0\rangle$5 has Laplacian

$|0\rangle$6

eigenvectors $|0\rangle$7, and eigenvalues $|0\rangle$8. With spectral kernel $|0\rangle$9, typically the heat kernel $x$0, the MMD again diagonalizes:

$x$1

For shift-invariant graphs such as the cycle $x$2 for ordinal variables and the complete graph $x$3 for categorical variables, the $x$4 are $x$5 characters and

$x$6

The MMD then reduces to a mixture of squared Heisenberg–Weyl moment differences,

$x$7

with

$x$8

For $x$9, $p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$0, whereas for $p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$1, $p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$2 and $p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$3 (Huang et al., 7 Jul 2026).

This spectral diagonalization is the mechanism that makes SBMs classically trainable at scale. In the IQP setting, the same principle appears as a Walsh-diagonal positive-definite kernel

$p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$4

for which

$p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$5

with residual vector $p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$6 (Slim et al., 26 May 2026).

3. IQP realizations and trainability

An $p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$7-qubit IQP-QCBM prepares $p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$8 and measures all qubits in the $p(x;\theta) = |\langle x|U(\theta)|0\rangle|^2,$9 basis, with

$q_\theta(x)$0

where the generators are Pauli-$q_\theta(x)$1 strings indexed by $q_\theta(x)$2. For $q_\theta(x)$3, the characteristic function is

$q_\theta(x)$4

and depends only on the generators that anticommute with $q_\theta(x)$5. Defining

$q_\theta(x)$6

the paper derives a subset-sum expansion over

$q_\theta(x)$7

which yields exact formulas for the variances of $q_\theta(x)$8 and its partial derivatives under symmetric i.i.d. initialization (Shen et al., 11 Feb 2026).

The central trainability result is the uniform-initialization theorem. With i.i.d. uniform $q_\theta(x)$9, define the critical rank

$$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$0

Then

$$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$1

and, for $$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$2,

$$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$3

while the derivative is zero for $$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$4. These variances monotonically decrease as more generators are added. Hence, the gradient magnitude is explicitly controlled by the linear-algebraic property $$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$5 (Shen et al., 11 Feb 2026).

This result makes the kernel spectrum and the generator topology jointly decisive. If the kernel spectrum is sufficiently flat or the architecture forces $$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$6 for all $$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$7, then $$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$8. Conversely, if $$q_\theta(x) = \operatorname{Tr}\!\left[(|x\rangle\langle x| \otimes I)\rho_\theta\right], \qquad \rho_\theta = U(\theta)|0\rangle\langle 0|U(\theta)^\dagger.$$9 concentrates an inverse-polynomial fraction on frequencies with $\mathbb{Z}_d^n$0, then

$\mathbb{Z}_d^n$1

avoiding exponential suppression. Hamming-weight-dependent kernels are “low-weight-biased” if $\mathbb{Z}_d^n$2 concentrates mass on small $\mathbb{Z}_d^n$3, and in structured IQP topologies such kernels can avoid exponential gradient suppression at lower-weight frequencies (Shen et al., 11 Feb 2026).

The architecture dependence is explicit in the examples analyzed. For product-state generators, $\mathbb{Z}_d^n$4. For a 2D lattice with single-qubit and nearest-neighbor $\mathbb{Z}_d^n$5, $\mathbb{Z}_d^n$6 with $\mathbb{Z}_d^n$7. For sparse Erdős–Rényi graphs $\mathbb{Z}_d^n$8 with $\mathbb{Z}_d^n$9, one has $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$00 and low-weight modes remain subexponential or polylogarithmic in rank. For the complete graph with all single- and two-qubit generators, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$01 for any $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$02, implying $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$03 and exponential suppression everywhere; in this case no kernel can prevent a barren plateau (Shen et al., 11 Feb 2026).

The same analysis identifies an alternative trainable regime. If $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$04 are i.i.d. Gaussian with $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$05 and $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$06 for constant $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$07 and $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$08, then for any frequency $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$09 and any generator $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$10,

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$11

Under mild conditions, the lower bound for the MMD gradient is polynomial, avoiding barren plateaus. The paper notes that this concentrates $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$12 near identity, which can ease classical simulability; use carefully when preserving potential advantage matters (Shen et al., 11 Feb 2026).

The complexity-theoretic picture does not coincide with trainability. Product and 2D lattice families are not $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$13 anti-concentrated. Sparse ER with $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$14, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$15, satisfies $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$16 and is anti-concentrated. The complete graph is also anti-concentrated but untrainable at initialization. The existence theorem in the sparse ER case therefore connects a trainable low-weight regime with hardness arguments based on anti-concentration and the Bremner-style polynomial-hierarchy consequences for additive-error classical sampling (Shen et al., 11 Feb 2026).

4. Classical surrogation, truncation, and deployment discrepancy

The spectral viewpoint is independent of the loss function. QCBMs are identified as a quantum Fourier model independently of the loss function, because the Born distribution over computational-basis bitstrings is exactly the Walsh–Hadamard transform of a set of $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$17-string correlators. This permits direct spectrum truncation. The paper studies $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$18-order truncation,

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$19

and Random Fourier Correlator truncation,

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$20

These approximations preserve normalization if $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$21 is included, but not positivity; thus, training must use losses defined on $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$22 (Herrero-Gonzalez et al., 3 Nov 2025).

The omitted spectrum quantitatively controls the approximation error. If $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$23, then the deterministic upper bound is

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$24

and for Haar-random unitaries,

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$25

The paper recasts training as regression in an RKHS induced by the parity kernel

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$26

which yields random-feature approximations over frequencies $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$27 with a sampling distribution proportional to $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$28 (Herrero-Gonzalez et al., 3 Nov 2025).

The resulting dequantization theorem states that if the sampling distribution over frequencies is efficiently samplable, polynomially concentrated, and aligned with the optimal quantum spectrum, then RFC-based classical training achieves, with high probability and $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$29, an excess risk within $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$30 of the optimal quantum risk. The paper also gives a discrepancy theorem for train-classical, deploy-quantum pipelines:

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$31

The first term is the feature-cardinality gap; the second is surrogate mismatch. Together, they quantify deployment-induced distribution shifts (Herrero-Gonzalez et al., 3 Nov 2025).

Two surrogate families are analyzed in detail. Tensor-network surrogates use Matrix Product States and RMPS averages, with closed-form variance formulas for marginals and truncated probabilities. Pauli-propagation surrogates propagate observables in the Heisenberg picture and, for IQP circuits, give a closed-form surrogate for truncated probabilities with flip-budget $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$32. The paper studies IQP circuits, matchcircuits, Heisenberg-chain circuits, and Haldane-chain circuits, derives analytic variance formulas for matchcircuits, and characterizes a new dynamical Lie algebra for the Haldane chain. The overarching conclusion is that not all correlators are trainable; high-order terms may vanish or offer little gradient signal individually, which motivates spectral truncation and surrogate training on a controllable subset (Herrero-Gonzalez et al., 3 Nov 2025).

5. Implementations, software, and empirical demonstrations

The software and hardware-facing realization of SBMs is now split between fully general qudit implementations and specialized IQP pipelines. The paper “Spectral Born machines: classically trainable quantum generative models for discrete data” makes the SBM training machinery available in a new tcdq module of the PennyLane software platform. The module provides batched HW moment estimators implementing the $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$33 factoring, graph spectral kernel utilities for cycle and complete graphs, unbiased MMD U-statistic estimators with JAX autodiff, gate-set construction helpers for weight/degree restrictions and sparse $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$34 selection by empirical Fourier magnitude, and GPU-ready pipelines for large $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$35, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$36, and $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$37 (Huang et al., 7 Jul 2026).

For tractable $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$38, expectation values of HW displacements can be estimated to inverse polynomial additive precision by uniform sampling over $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$39:

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$40

where $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$41. This is unbiased, with $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$42 samples for additive error $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$43. Batching many moments and $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$44-samples can be reduced to matrix multiplications, with overall complexity

$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$45

which is practical with low weights such as one- and two-site gates (Huang et al., 7 Jul 2026).

A distinct implementation line appears in the 64-qubit calorimeter study, which trains an IQP Born machine on real high-energy-physics calorimeter shower images and compiles the trained model into a single sampling-hard IQP circuit for quantum deployment. The pipeline has three components: a Mixture-of-IQP (\moiqp{}) architecture, a Pearson-Stabilized Correlation Kernel (\psck{}), and an exact deferred-measurement compilation of \moiqp{} into a single IQP circuit on $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$46 qubits (\ciqp{}). With $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$47, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$48, and $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$49, the compiled circuit acts on $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$50 qubits. Across five seeds at $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$51, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$52 epochs, the model reaches $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$53 against a $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$54 encoding-fidelity floor on the training split and $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$55 on a held-out test split, versus a Liu--Wang baseline at $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$56. The compiled \ciqp{} reproduces the \moiqp{} marginal to $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$57 times the Monte Carlo noise floor (Slim et al., 26 May 2026).

The reported large-scale SBM demonstrations extend beyond qubit IQP models:

Study	System and model	Reported outcome
Potts/clock model	$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$58 qudits, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$59	All models reach similar train/test $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$60
Ribosomal RNA	$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$61 qudits, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$62; $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$63-qubit circuit when encoded to qubits	Model 1 final $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$64; Model 2 $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$65
Calorimeter images	IQP Born machine at $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$66 qubits, compiled to $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$67 qubits	$$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$68 train, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$69 test

In the synthetic cyclic-data experiment, four SBMs on a $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$70 periodic lattice with $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$71 and train/test size $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$72 each were trained for $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$73 iterations; the Degree 1, Degree 2, Degree 3, and All degrees models used $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$74, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$75, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$76, and $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$77 parameters, respectively, and all reached similar train/test $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$78. The largest converges fastest, consistent with overparameterization benefits without overfitting. In the ribosomal RNA experiment, positions were filtered to retain $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$79 positions out of $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$80, with train size $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$81 and test size $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$82; the two-qudit model used $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$83 parameters, the two-qudit plus selected three-qudit model used $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$84 parameters, and the reported observation was “no obvious overfitting despite $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$85M parameters” (Huang et al., 7 Jul 2026).

6. Limitations, misconceptions, and open questions

Several limitations are structural rather than implementation-specific. In the IQP trainability analysis, the average-case lower bound uses Assumption 2, namely an unstructured target distribution ensemble with mean-zero and pairwise-uncorrelated characteristic values; highly structured targets can violate it, and tighter analysis of cross-frequency covariances is then needed. Assumption 1, symmetric i.i.d. initialization, is crucial for the exact subset-sum variance formulas; asymmetric or correlated initializations require modified analysis. The small-variance Gaussian bound assumes $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$86 so that $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$87 (Shen et al., 11 Feb 2026).

A common misconception is that classical hardness of sampling and trainability necessarily coincide. The architecture comparison in the IQP setting shows the opposite. Product and 2D lattice families are trainable at low weights but are not anti-concentrated. The complete graph is anti-concentrated but not trainable at initialization. Sparse Erdős–Rényi architectures with $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$88 occupy the regime in which anti-concentration coexists with partial-spectrum trainability at low Hamming weight. This suggests that the feasible operating regime is architecture- and kernel-dependent rather than generic (Shen et al., 11 Feb 2026).

A second misconception is that spectral training guarantees accurate distribution learning in stronger distances. Minimizing MMD does not guarantee small total variation distance $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$89; MMD is controlled by TV but not vice versa. In the truncated-spectrum setting, pseudo-probabilities can be negative, so losses must be defined on $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$90. The dequantization conditions in RFC-based learning hinge on access to the optimal spectrum and its sparsity or low-degree structure; if the learned quantum distribution’s mass is spread across high-order frequencies, RFC dequantization fails (Herrero-Gonzalez et al., 3 Nov 2025).

At the model-design level, kernel choice sensitivity is substantial. A mismatch, such as using complete graphs for ordinal data, can slow training or reduce accuracy. The group structure itself can also be mismatched to the data; if the data do not exhibit cyclic or categorical structure, $$U(\theta) = (F^{\otimes n})^\dagger D(\theta) F^{\otimes n}$$91 assumptions may not help. On hardware, approximate QFT compilation errors or noise degrade sampling fidelity, even though training is classical. The papers identify several open lines: broader dequantization criteria beyond shift-invariant kernels and RFC alignment, robustness-to-noise analyses in the spectral picture, non-Abelian SBMs aligned with domain symmetries, and extensions to broader architectures including IQP universality with ancilla, boson sampling or continuous-variable Spectral Born Machines, and photonic implementations (Huang et al., 7 Jul 2026, Herrero-Gonzalez et al., 3 Nov 2025).

Within these constraints, the published SBM literature now delineates a coherent picture. SBMs are QCBMs viewed through their Fourier spectrum, with Fourier coefficients equal to directly trainable correlators or Heisenberg–Weyl moments; they admit graph-spectral MMD objectives, classically tractable estimators, and train-classical/deploy-quantum workflows; and their practical performance depends on the interaction between kernel spectrum, generator topology, initialization, and the spectral structure of the data (Huang et al., 7 Jul 2026).