Bayesian Quantum Circuit Overview

• Bayesian Quantum Circuit (BQC) is a quantum framework that integrates Bayesian inference into parameterized circuits using explicit registers for latent and data variables.
• It factorizes the circuit into modules for prior, likelihood, and posterior, enabling efficient sampling and robust representation of complex distributions.
• BQC employs Bayesian parameter learning and advanced optimization techniques like SGLD to reduce circuit depth and improve noise robustness in quantum machine learning tasks.

A Bayesian Quantum Circuit (BQC) is a quantum computational framework that realizes the modular structure of a Bayesian probabilistic model within a parameterized quantum circuit (PQC) architecture. BQC leverages explicit quantum registers for latent variables and data, constructing circuits such that the measurement statistics yield joint, marginal, and posterior distributions emulating Bayesian inference. The BQC formalism encompasses not only the generation and estimation of data distributions but also Bayesian learning over the circuit parameters themselves, supporting advanced quantum machine learning protocols and efficient quantum representations for classical Bayesian networks (Du et al., 2018, Duffield et al., 2022, Borujeni et al., 2020, Tucci, 2014).

1. Circuit Architecture and Fundamental Structure

A BQC factorizes a quantum circuit into explicit modules associated with Bayesian quantities:

• An ancillary register of $n_a$ qubits encodes the discrete latent variable $z \in \{0,1\}^{n_a}$ (the prior).
• A data register of $n_x$ qubits encodes observable variables $x \in \{0,1\}^{n_x}$.
• The overall unitary $U_{\text{total}}(\lambda, \theta)$ decomposes as:

$$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$

where $U_P(\lambda)$ prepares the prior on the ancilla register, and $U_L(\theta)$ implements the likelihood via entangling operations between ancilla and data qubits.

Measurement in the computational basis yields samples $(z, x)$ with joint probability:

$$p_{\text{model}}(z, x; \lambda, \theta) = |\langle z, x | U_{\text{total}}(\lambda, \theta) | 0 \rangle^{\otimes (n_a+n_x)}|^2$$

Marginal and conditional distributions (posterior over $z \in \{0,1\}^{n_a}$0, marginal of $z \in \{0,1\}^{n_a}$1) result from suitable summation over measurement outcomes (Du et al., 2018).

2. Parameterization: Prior, Likelihood, Posterior

The BQC parameterization follows the Bayesian inference hierarchy:

• Prior Preparation ($z \in \{0,1\}^{n_a}$2): Composed of $z \in \{0,1\}^{n_a}$3 layers, each with single-qubit $z \in \{0,1\}^{n_a}$4 rotations per ancilla and a fixed entangling gate pattern (e.g., ring of CNOTs). Parameter set $z \in \{0,1\}^{n_a}$5 thus encodes the variational prior $z \in \{0,1\}^{n_a}$6.
• Likelihood Transformation ($z \in \{0,1\}^{n_a}$7): Incorporates $z \in \{0,1\}^{n_a}$8 layers per data qubit. Each layer contains:
  • Controlled $z \in \{0,1\}^{n_a}$9 rotations from each ancilla to each data qubit (modeling $n_x$0).
  • Local mixing (additional $n_x$1 rotations and entanglers) on the data register.
• Posterior: The conditional $n_x$2 is directly accessible by measurement statistics.

This explicit factorization mitigates mode collapse and enables direct sample generation from the learned model (Du et al., 2018). Ancilla-based priors supply a mechanism for diversity and efficient representation of multimodal distributions.

3. Bayesian Learning over Circuit Parameters

BQC supports full Bayesian treatment of circuit parameters as random variables:

• Given a cost function $n_x$3 (e.g., loss for a generative or supervised task), one defines a prior $n_x$4 over parameters and constructs a (generalized) likelihood $n_x$5.
• The posterior over $n_x$6 is:

$n_x$7

• A Laplace ($n_x$8) prior enables automatic sparsification of parameters, leading to zeroing of small angles and reduced quantum circuit depth without sacrificing performance.
• Dimension reduction is achieved via MAP estimation:

$n_x$9

and optional Laplace approximation to the posterior for further parameter pruning (Duffield et al., 2022).

• Posterior sampling can be performed with stochastic gradient Langevin dynamics (SGLD):

$x \in \{0,1\}^{n_x}$0

where $x \in \{0,1\}^{n_x}$1. SGLD escapes local minima and produces samples from the full Bayesian posterior (Duffield et al., 2022).

4. Bayesian Quantum Circuits for Classical Graphical Models

BQC provides explicit quantum embeddings of classical Bayesian networks and probabilistic graphical models:

• Root Nodes: Prepared with single-qubit $x \in \{0,1\}^{n_x}$2 rotations such that $x \in \{0,1\}^{n_x}$3 encodes marginal probabilities.
• Non-root Nodes: Realized by multi-controlled $x \in \{0,1\}^{n_x}$4 gates conditioned on parental configurations, encoding conditional probability tables (CPTs) into amplitudes.
• Multi-state Variables: Encoded using binary or unary quantum representations over multiple qubits, with cascaded controlled rotations to achieve the desired amplitudes.
• Ancilla Decomposition: $x \in \{0,1\}^{n_x}$5-controlled $x \in \{0,1\}^{n_x}$6 gates decomposed into $x \in \{0,1\}^{n_x}$7 Toffoli gates and up to $x \in \{0,1\}^{n_x}$8 ancillas, yielding polynomial-time compilation with classical resources matching Bayesian network structure (Borujeni et al., 2020).
• Gate-level circuits for Bayesian structure learning (e.g., Tucci’s ordered/structural modular models) allow amplitude estimation of network posteriors $x \in \{0,1\}^{n_x}$9 via Grover-like iterations (Tucci, 2014).

Model Element	Circuit Mapping	Quantum Resource
Marginal $U_{\text{total}}(\lambda, \theta)$0	$U_{\text{total}}(\lambda, \theta)$1	1 qubit
CPT $U_{\text{total}}(\lambda, \theta)$2	Multi-controlled $U_{\text{total}}(\lambda, \theta)$3	$U_{\text{total}}(\lambda, \theta)$4 controls, $U_{\text{total}}(\lambda, \theta)$5 ancilla
Multi-state node	Cascaded binary/unary encoding	$U_{\text{total}}(\lambda, \theta)$6 or $U_{\text{total}}(\lambda, \theta)$7 qubits

Preparation accurately matches classical marginals in simulation, with polynomial resources for bounded in-degree networks (Borujeni et al., 2020).

5. Training Objectives and Optimization

BQC training targets both generative and semi-supervised learning setups:

• Unsupervised/GAN-style: Fit the empirical distribution $U_{\text{total}}(\lambda, \theta)$8 by minimizing KL divergence:

$U_{\text{total}}(\lambda, \theta)$9

or via the MMD objective:

$$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$0

where $$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$1 is a kernel-induced feature embedding.

• Semi-supervised/Classification: For labeled pairs $$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$2, minimize the negative log-posterior:

$$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$3

Possibly include MMD regularization as $$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$4.

Gradients for $$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$5-rotation parameters are estimated by the parameter-shift rule, with classical optimizers (Adam, SGD) applied for update (Du et al., 2018).

6. Empirical Results and Resource Analysis

Simulation and hardware validation confirm BQC performance:

• In generative learning, BQC achieves $$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$6 mode coverage on discrete benchmarks (e.g., 2×2 Bars-and-Stripes), with $$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$7 fewer quantum shots needed than baseline quantum circuit Born machines.
• Semi-supervised binary MNIST task (using $$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$8, $$U_{\text{total}}(\lambda, \theta) = U_L(\theta)\ [U_P(\lambda) \otimes I_{\text{data}}]$$9 PCA features) yields $U_P(\lambda)$0 accuracy with BQC versus $U_P(\lambda)$1 for classical variational classifiers with only $U_P(\lambda)$2 labels per class.
• Laplace-regularized MAP optimization prunes $U_P(\lambda)$3–$U_P(\lambda)$4 of parameterized gates, cuts $U_P(\lambda)$5 of two-qubit gates post-compilation, halves necessary hardware runtime, and improves noise robustness (total-variation distance reduction by $U_P(\lambda)$6–$U_P(\lambda)$7) (Duffield et al., 2022).
• For quantum learning of Bayesian networks, polynomial resources are required for sampling and marginalized posteriors as long as maximum in-degree is bounded (Borujeni et al., 2020), and quantum amplitude estimation can recover network probabilities via Grover/adaptive fixed-point techniques (Tucci, 2014).
• Direct data sampling is realized by applying $U_P(\lambda)$8 to $U_P(\lambda)$9 and measuring, with every shot yielding a valid $U_L(\theta)$0. Ancilla-driven latent sampling enables generation of new, previously unseen data combinations.

7. Implications, Extensions, and Limitations

BQC enables modular quantum inference with advantages in expressivity, uncertainty quantification, and computational efficiency:

• Explicit Bayesian structure (prior-likelihood-posterior) reduces effective quantum circuit depth compared to monolithic PQCs.
• Bayesian treatment of parameters provides uncertainty quantification and robustness against optimization landscapes with plateaus or multiple minima; SGLD sampling in parameter space further improves exploration and avoids local traps (Duffield et al., 2022).
• Circuit sparsification by $U_L(\theta)$1-induced MAP estimation directly combats hardware noise and coherence limitations in near-term quantum hardware.
• The approach generalizes to quantum representations of complex graphical models and may be extended to Bayesian quantum learning of model structure (such as discovery of Bayesian network graphs), for which circuits to estimate posteriors $U_L(\theta)$2 are explicitly constructed (Tucci, 2014).
• Extensions include higher-order Bayesian priors (von Mises, Gaussian-process), variational Bayesian quantum inference, and approximate Bayesian computation or pseudo-marginal MCMC.
• Limitations stem from the scalability of controlled rotations with large parent sets, the need for efficient ancilla allocation, and the computational cost of advanced prior structures and posterior sampling; parameter tuning (sparsity $U_L(\theta)$3, noise $U_L(\theta)$4) is nontrivial.
• In the quantum Bayesian network setting, resource usage is polynomial for bounded in-degree; without such bounds, the scalable quantum advantage is less clear (Borujeni et al., 2020, Tucci, 2014).

BQC thus forms a substantial module for quantum probabilistic machine learning, facilitating both efficient synthesis of classical Bayesian models and fully quantum-native Bayesian learning (Du et al., 2018, Duffield et al., 2022, Borujeni et al., 2020, Tucci, 2014).