Quantum Latent Distributions

• Quantum latent distributions are constructions that represent latent variables and states using quantum sampling processes and operator-valued embeddings.
• They integrate quantum priors—such as boson sampling and quantum Boltzmann machines—into deep generative models to achieve advantages over classical methods.
• They also embed classical probability measures into quantum state spaces, providing a framework for improved expressivity and measurable performance gains.

Quantum latent distributions are a family of constructions in which latent variables, latent states, or entire probability measures are represented, sampled, regularized, or compared by quantum-mechanical objects. In current arXiv usage, the phrase does not denote a single standardized model. It refers, at minimum, to quantum priors for classical generative models such as boson samplers and quantum Boltzmann machines, latent representations that are themselves quantum states or density operators, and operator-valued embeddings that map classical probability measures into the state space of quantum mechanics (Bacarreza et al., 27 Aug 2025, Khoshaman et al., 2018, Raj et al., 19 Sep 2025, McCarty, 26 Aug 2025).

1. Terminological scope and conceptual distinctions

The literature uses the term in several technically distinct ways. In deep generative modeling, the latent distribution is the source law $P_z$ in a pushforward model $P_{g(z)} = g_\# P_z$, and it is called quantum when $P_z$ is sampled from a quantum process such as boson sampling or a quantum Boltzmann machine rather than from a Gaussian or Bernoulli prior (Bacarreza et al., 27 Aug 2025, Khoshaman et al., 2018). In quantum autoencoding and quantum generative modeling, the latent object is often not a classical random vector at all, but a reduced quantum state $\eta \in \mathcal D(\mathcal H_L)$, a mixed-state code $\zeta$, or an ensemble of density operators generated by a circuit conditioned on a classical latent variable (Raj et al., 19 Sep 2025, Wang et al., 2024, Tran et al., 27 May 2026).

A second distinction concerns whether the latent object is a prior over samples or a representation of a distribution itself. In the quantum probability metric framework, the point is not to introduce a separate latent-variable model class, but to embed a classical probability measure $\mu$ into the convex state space of quantum mechanics, producing an operator $\hat\mu$ and measuring discrepancies there (McCarty, 26 Aug 2025). The same shift appears in quantum mean embedding, where a probability distribution is represented by a normalized quantum superposition rather than by a tractable finite-dimensional feature vector (Kübler et al., 2019).

A third distinction concerns single states versus distributions over states. LPQCs emphasize that the target object can be a probability measure $Q$ on $\mathcal D(\mathcal H_S)$, not merely its average density matrix

$\overline{\rho}=\sum_j p_j \rho_j.$

This matters because two physically distinct ensembles can share the same $P_{g(z)} = g_\# P_z$0 (Tran et al., 27 May 2026). QGAA makes an analogous point operationally: its “latent distribution” is the set of encoder-produced latent quantum states $P_{g(z)} = g_\# P_z$1 and their label-conditioned structure, rather than a closed-form classical density over latent coordinates (Raj et al., 19 Sep 2025).

2. Quantum priors and samplers in deep generative models

One major lineage treats the latent distribution as the main source of representational gain. In "Quantum latent distributions in deep generative models" (Bacarreza et al., 27 Aug 2025), the latent prior is quantum when it belongs to a class $P_{g(z)} = g_\# P_z$2 of distributions that can be approximated in polynomial time on a quantum computer but not by any classical algorithm in the reference class $P_{g(z)} = g_\# P_z$3. The concrete experimental instantiation is boson sampling. The core theorem states that if a generator $P_{g(z)} = g_\# P_z$4 has an inverse $P_{g(z)} = g_\# P_z$5 that exists, is efficiently classically implementable, and is Lipschitz continuous, and $P_{g(z)} = g_\# P_z$6, then the pushforward $P_{g(z)} = g_\# P_z$7. The corresponding corollary gives an architecture-dependent separation in the sense that there exist target distributions reachable from a quantum latent that cannot be matched arbitrarily well by any classical latent within the same bounded-complexity generator class. The same paper identifies two practical mechanisms behind empirical gains: inductive bias for physically quantum or quantum-like data, and reduced factorization of latent structure, especially for multimodal targets (Bacarreza et al., 27 Aug 2025).

The quantum variational autoencoder takes a different route. In "Quantum Variational Autoencoder" (Khoshaman et al., 2018), the latent variables at inference time are classical bitstrings $P_{g(z)} = g_\# P_z$8, but their prior is the diagonal of a quantum Gibbs state,

$P_{g(z)} = g_\# P_z$9

The latent distribution is therefore quantum because it is induced by a non-commuting Hamiltonian rather than by a classical energy model. The paper derives a quantum lower bound to the ELBO using Golden-Thompson, evaluates the negative phase with continuous-time quantum Monte Carlo and population annealing, and uses the RBM limit $P_z$0 as the classical baseline (Khoshaman et al., 2018).

A hybrid latent-space GAN formulation appears in "Latent Style-based Quantum GAN for high-quality Image Generation" (Chang et al., 2024). There a classical convolutional autoencoder first maps data to a bounded latent vector $P_z$1, and the quantum generator learns the empirical latent distribution rather than the pixel distribution. The generator maps classical noise $P_z$2 into circuit parameters $P_z$3, and outputs continuous latent features through expectation values,

$P_z$4

The adversarial game is a Wasserstein GAN with gradient penalty, and the paper’s theoretical contribution includes a barren-plateau analysis showing polynomially decaying gradient variance for shallow or carefully initialized circuits and exponentially vanishing gradients for polynomial-depth circuits under broad random initialization (Chang et al., 2024).

3. Latent quantum states and distributions over density operators

In models for quantum data, the latent object is frequently a state in a smaller Hilbert space. "Quantum Generative Adversarial Autoencoders" (Raj et al., 19 Sep 2025) defines an encoder

$P_z$5

and latent states

$P_z$6

A decoder reconstructs $P_z$7, and the QAE is trained with a fidelity-based reconstruction loss. The model becomes generative only after attaching a QGAN to the trained QAE: the encoder’s outputs $P_z$8 are treated as the real data source, the generator produces latent states $P_z$9, and training solves

$\eta \in \mathcal D(\mathcal H_L)$0

In this formulation, the latent distribution is explicitly not a classical VAE-style density $\eta \in \mathcal D(\mathcal H_L)$1 regularized toward an analytic prior; it is the empirical family of latent quantum states learned implicitly through adversarial matching (Raj et al., 19 Sep 2025).

A fully quantum analogue of VAE regularization is given by "$\eta \in \mathcal D(\mathcal H_L)$2-QVAE" (Wang et al., 2024). The encoder and decoder are CPTP maps,

$\eta \in \mathcal D(\mathcal H_L)$3

with latent mixed states

$\eta \in \mathcal D(\mathcal H_L)$4

The latent prior is the maximally mixed state,

$\eta \in \mathcal D(\mathcal H_L)$5

and the objective combines reconstruction and latent regularization,

$\eta \in \mathcal D(\mathcal H_L)$6

Because the latent code is a density matrix rather than a Euclidean random variable, the framework supports fidelity, quantum relative entropy, symmetric quantum relative entropy, and Wasserstein-type losses, as well as instance-level and global density-matrix training (Wang et al., 2024).

"Latent-Conditioned Parameterized Quantum Circuits as Universal Approximators for Distributions over Quantum States" (Tran et al., 27 May 2026) pushes the concept from latent states to latent-conditioned distributions over states. A classical latent variable $\eta \in \mathcal D(\mathcal H_L)$7 is sampled from a prior $\eta \in \mathcal D(\mathcal H_L)$8, mapped by neural networks to PQC parameters $\eta \in \mathcal D(\mathcal H_L)$9, and used to generate

$\zeta$0

The paper proves that LPQCs are universal approximators of probability measures on $\zeta$1 in the $\zeta$2-Wasserstein distance: for every probability measure $\zeta$3 on $\zeta$4 and every $\zeta$5, there exists an LPQC class such that $\zeta$6. It also introduces a multimodal latent prior

$\zeta$7

and a mixture-of-experts parameter map, both presented as practical devices for multimodal ensembles and for alleviating barren plateaus (Tran et al., 27 May 2026).

4. Quantum-state embeddings of classical probability distributions

A separate but increasingly influential usage treats classical distributions themselves as quantum states. "Quantum-inspired probability metrics define a complete, universal space for statistical learning" (McCarty, 26 Aug 2025) embeds a probability measure $\zeta$8 by a barycenter map

$\zeta$9

with $\mu$0 and $\mu$1. The induced quantum probability metric is the trace distance between embedded measures,

$\mu$2

For coherent states, this construction is tied directly to the Gaussian kernel by

$\mu$3

so the Gaussian RKHS picture and the quantum pure-state picture become two geometries on the same underlying embedding. The difference from MMD is precisely geometric: MMD corresponds to the Hilbert-Schmidt norm $\mu$4, whereas QPM uses the trace norm $\mu$5. The paper proves the incompleteness of reflexive embeddings on noncompact spaces, then states that every QPM completely metrizes $\mu$6 and that every Polish space has a QPM. It further proves that dual functions for the Fock QPM are dense in $\mu$7, which is the basis for its claim of a larger witness class than standard Gaussian or Laplacian RKHS functions on noncompact domains (McCarty, 26 Aug 2025).

The computational recipe is also explicit. For finite atomic measures, one forms $\mu$8, computes a Gram factorization $\mu$9, sets $\hat\mu$0, and obtains the relevant eigenvalues from $\hat\mu$1. The resulting discrepancy is

$\hat\mu$2

The paper presents this as a drop-in replacement for MMD with analytic gradients, while also noting the cost increase from $\hat\mu$3 kernel evaluations for MMD to typically $\hat\mu$4 eigenvalue computation for QPM (McCarty, 26 Aug 2025).

"Quantum Mean Embedding of Probability Distributions" (Kübler et al., 2019) offers a closely related but distinct representation. Given a quantum feature map $\hat\mu$5, the quantum mean embedding is the normalized superposition

$\hat\mu$6

with normalization

$\hat\mu$7

The key relation is

$\hat\mu$8

which makes QME a quantum-mechanical re-expression of kernel mean embedding rather than an alternative latent prior. Under a universal kernel, the representation remains injective over probability measures (Kübler et al., 2019).

5. Empirical performance across application domains

Empirical work with quantum latent priors has been most extensive in GAN-like settings. On a synthetic quantum dataset generated from 8 indistinguishable photons in a 16-channel random optical circuit, the boson-sampler latent achieved $\hat\mu$9, compared with $Q$0 for the distinguishable-photon sampler, $Q$1 for the Gaussian latent, and $Q$2 for the Bernoulli latent. On QM9 at latent size $Q$3, the boson sampler reported FCD $Q$4, valid unique $Q$5, and novel $Q$6, outperforming the distinguishable-photon, Bernoulli, and Gaussian baselines. The same paper reports that results from the ORCA Computing PT-2 real photonic boson sampler closely matched the simulated boson sampler and still outperformed the classical baselines, while DDGAN experiments on CIFAR-10 gave comparable FID across Gaussian and photonic latents rather than a clear quantum gain (Bacarreza et al., 27 Aug 2025).

For image generation in latent space, LaSt-QGAN reports MNIST FID $Q$7, IS $Q$8, JSD(features) $Q$9, and JSD(images) $\mathcal D(\mathcal H_S)$0 for Circuit 3 at depth 6, compared with a classical GAN with hidden layers $\mathcal D(\mathcal H_S)$1 at FID $\mathcal D(\mathcal H_S)$2, IS $\mathcal D(\mathcal H_S)$3, JSD(features) $\mathcal D(\mathcal H_S)$4, and JSD(images) $\mathcal D(\mathcal H_S)$5. On SAT4, the reported values were FID $\mathcal D(\mathcal H_S)$6, IS $\mathcal D(\mathcal H_S)$7, JSD(features) $\mathcal D(\mathcal H_S)$8, and JSD(images) $\mathcal D(\mathcal H_S)$9 for LaSt-QGAN, versus FID $\overline{\rho}=\sum_j p_j \rho_j.$0, IS $\overline{\rho}=\sum_j p_j \rho_j.$1, JSD(features) $\overline{\rho}=\sum_j p_j \rho_j.$2, and JSD(images) $\overline{\rho}=\sum_j p_j \rho_j.$3 for the classical baseline. The same study reports that FID below 20 on MNIST is reached in fewer than 20 epochs and that finite-shot latent generation becomes essentially indistinguishable from the infinite-shot result by about 512 shots (Chang et al., 2024).

For quantum data generation, QGAA reports average generated-state fidelities $\overline{\rho}=\sum_j p_j \rho_j.$4 for $\overline{\rho}=\sum_j p_j \rho_j.$5 and $\overline{\rho}=\sum_j p_j \rho_j.$6 for $\overline{\rho}=\sum_j p_j \rho_j.$7, with average absolute energy errors $\overline{\rho}=\sum_j p_j \rho_j.$8 for $\overline{\rho}=\sum_j p_j \rho_j.$9 and $P_{g(z)} = g_\# P_z$00 for $P_{g(z)} = g_\# P_z$01. The same paper reports average QAE reconstruction fidelities of about $P_{g(z)} = g_\# P_z$02 for $P_{g(z)} = g_\# P_z$03 and $P_{g(z)} = g_\# P_z$04 for $P_{g(z)} = g_\# P_z$05, and interprets the larger LiH error as evidence that the latent distribution was learned only approximately because the underlying compression was itself more difficult (Raj et al., 19 Sep 2025).

Quantum generative modeling of compressed scientific latents has also been tested outside image and molecule domains. In CFD, a VQ-VAE compresses $P_{g(z)} = g_\# P_z$06 vorticity snapshots at $P_{g(z)} = g_\# P_z$07 into a 7-dimensional latent vector, discretizes each latent dimension into 256 bins, and compares seven independent 8-qubit QCBMs and seven independent 10-qubit QGAN circuits against a single-layer LSTM baseline. The reported average minimum distances are approximately $P_{g(z)} = g_\# P_z$08 for QCBM, $P_{g(z)} = g_\# P_z$09 for QGAN, and $P_{g(z)} = g_\# P_z$10 for LSTM, and QCBM is nearest neighbor to over 1600 out of 1999 original codebook vectors (Hsain et al., 27 Dec 2025). In topology optimization, a variational quantum circuit generates a bounded latent code

$P_{g(z)} = g_\# P_z$11

which is projected and decoded into material fields. On the tip-loaded cantilever benchmark at iteration 200, the 5-qubit quantum encoding reported compliance $P_{g(z)} = g_\# P_z$12 and diversity $P_{g(z)} = g_\# P_z$13, compared with compliance $P_{g(z)} = g_\# P_z$14 and diversity $P_{g(z)} = g_\# P_z$15 for the matched classical baseline; other benchmarks show that the advantage is problem-dependent rather than universal (Tabarraei, 20 Jun 2025).

Distribution-embedding approaches also report downstream gains. In generative moment matching networks, QPM improves image quality on MNIST relative to MMD, and on CelebA-64 the paper reports that MMD could not reject the null in a two-sample test with batch size 1000 ($P_{g(z)} = g_\# P_z$16), while QPM gave $P_{g(z)} = g_\# P_z$17. On MNIST, both metrics detect differences, but QPM yields lower $P_{g(z)} = g_\# P_z$18-values and better samples (McCarty, 26 Aug 2025).

6. Limitations, adjacent usages, and open questions

The strongest claims in this literature are conditional. The boson-sampler separation theorem assumes an invertible, efficiently classically implementable, Lipschitz generator inverse, and the same paper explicitly states that quantum advantage is not universal: negative results on StyleGAN/CIFAR-10 and some QM9 hyperparameter settings show that changing the latent distribution alone did not reliably help (Bacarreza et al., 27 Aug 2025). QPMs gain completeness and expressivity at higher computational cost and under explicit assumptions on a closed embedding $P_{g(z)} = g_\# P_z$19 and a characteristic kernel; the matrix square-root and eigenvalue step can be numerically delicate when points are nearly coincident, and if a kernel does not admit a clean square-root kernel the method requires an approximation choice (McCarty, 26 Aug 2025). QVAE training remains limited by the expense of CT-QMC and by the looseness of the quantum bound as the transverse field $P_{g(z)} = g_\# P_z$20 increases (Khoshaman et al., 2018). LPQC universality is proved for compact latent spaces with positive, continuous priors, while the implemented training objective is a practical optimal-transport surrogate rather than exact Wasserstein minimization (Tran et al., 27 May 2026).

A recurrent misconception is that the phrase always names a latent prior for a generative model. The term is also used in broader or adjacent senses. "Quantum Latent Semantic Analysis" represents documents as normalized wave functions, latent topics as subspaces, and topic probabilities as squared projection amplitudes, with an interference term absent from classical mixture models (González et al., 2019). "Quantum Latent Gauge and Coherence Selective Forces" uses “latent distribution” for a conserved coherence current

$P_{g(z)} = g_\# P_z$21

which couples to a hidden $P_{g(z)} = g_\# P_z$22 gauge field only when quantum coherence is present (Horchani, 26 Nov 2025). In sequence modeling, generalized hidden Markov models with quantum or post-quantum latent belief states and a complex unitary wave-function model with Born-rule readout both use the language of latent distributions, but their primary object is a history-conditioned predictive state geometry rather than a generative-model prior (Riechers et al., 10 Jul 2025, Nebli et al., 24 Feb 2026).

Taken together, these works suggest a stable core meaning and a broad periphery. The stable core is that quantum latent distributions replace or augment classical latent structure by using quantum sampling laws, quantum states, or operator-valued embeddings. The broad periphery is terminological: “latent distribution” may also denote a geometric topic state, a coherence-selective current, or a predictive belief state. The field’s central unresolved question is therefore not whether a single quantum latent formalism exists, but when a given quantum latent construction yields a measurable advantage over classical priors, classical embeddings, or classical latent-state geometries.