Quantum Fokker-Planck solvers as trainable generative models

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: We're looking at the paper: "Quantum Fokker-Planck solvers as trainable generative models".

Mira: Here's the gist of the paper. The gist The authors develop and analyze generative models based on quantum solvers of stochastic differential equations, which enables sampling through computational-basis measurements and is efficiently controlled by coefficients of the block-encoded diagonal potential Quantum Generative Model Construction The model is constructed by parameterizing the potential of an overdamped Langevin process rather than using common approaches based on layered variational quantum circuits with parameterized rotations The associated Fokker-Planck (FP) equations are Hermitianized and solved using a linear combination of cosine operators implemented through quantum singular value transformation By encoding the square root of the stationary density, the model enables sampling through computational-basis measurements and is efficiently controlled by coefficients of the block-encoded diagonal potential The model is trained to match empirical distributions within a family described by FP equations, demonstrating parameter recovery for multivariate probability distributions Fokker-Planck Generative Model The generative model starts by considering a d-dimensional overdamped Langevin process with additive, isotropic noise, defined as dXt = −∇V (Xt) dt + σ dWt The probability density p(x, t) of an ensemble of such trajectories obeys the Fokker–Planck equation ∂tp = ∇ · p ∇V + σ 2/2 ∆p = −∇ · J, where J = −p∇V − σ 2/2 ∇p is the probability current Imposing the equilibrium condition J = 0 gives the stationary density pss(x) = Z−1 exp−2V (x)/σ squared, where Z = R e − 2V (x)/σ squared dx The model distribution is then defined as pθ(x) = 1/Z(θ) exp − 2/σ squared V (x; θ), where V (x; θ) is parameterized by the coefficients θ ∈ R K of a fixed dictionary of basis functions This model is physical because the fitted θ are the coefficients of the drift of a stochastic differential equation consistent with the data Quantum Implementation The quantum model takes θ as input and prepares a state with the discretized Born-rule distribution pθ of Eq.

Kai: (2) embedded into amplitudes The workflow is split into four stages: (A) transformation of the Fokker–Planck generator into a positive semidefinite operator H = A†A whose ground state is √pθ; (B) a block encoding of the discretized A in which θ enters only through the diagonal potential term (C) preparation of the ground state by an imaginary-time filter e−tH implemented as a linear combination of cosine operators (LCCO); and (D) measurement and a loss that drives the update of θ Hermitianization involves casting Eq.

Mira: (1) into imaginary-time Schr¨odinger form, ∂tψ = −Hψ, with H = −σ 2/2 ∇2 + Veff(x) The stationary state solution is retrieved by ground state preparation of the imaginary-time Schr¨odinger equation Training and Trainability The optimization problem differs from training a generic parameterized quantum circuit because the optimizer sees K variables regardless of the coarseness of the grid set by N or the data dimensionality d The only modeling choice is the dictionary, which encodes prior knowledge about the physical system in a way that is independent of grid resolution The optimization problem is closer to model discovery than to distribution fitting because the optimum is a drift field, and likelihood-based training is convex in natural parameters because the stationary state is an exponential family Resource Estimation The total training cost is given by Ctrain = Neval × ns / Psucc × Cprep The circuit width is logarithmic in the grid and in the quadrature order, and the state preparation cost Cprep is polynomial in N, d, t, D and log(1/ϵ) rather than in the Nd grid size The required imaginary time t is set by the gap ∆ of H Discussion and Conclusions The architecture avoids common pitfalls of scaling variational quantum machine learning models because the circuit is fixed and algorithmically structured The success of the proposed architecture over classical methods is most promising for many coupled stochastic degrees of freedom, as a d-dimensional grid with N points per axis occupies d log2 N qubits The framework opens a route to identifying regimes in which quantum state preparation offers an advantage over classical sampling How it works The training objective is to minimize the distance between generated and empirical distributions, using losses such as the Kullback–Leibler divergence KL(p∗∥pθ) or the squared maximum mean discrepancy (MMD2) Minimization involves matching K moments of the data, which is a convex problem with a unique optimum characterized by K moments of the data The model density has the exponential family form pθ(x) = expP j ηjvj (x) − log Z(θ), where natural parameters ηj = −2θj/σ squared and sufficient statistics vj

90: The convergence of the learning algorithm to the target is supported by numerical results showing that the landscape converges to the exact one as they are refined Approximation Errors The distance between the distribution sampled from the circuit and the exact stationary density pss is bounded by several contributions, including grid discretization error O(∆x 2) and finite imaginary time error e−t∆ The dominant approximation error in present experiments arises from the finite-order Hubbard–Stratonovich representation, since the oscillatory terms cos2√tχσmax(A) become increasingly difficult to resolve as the evolution time, domain range, or spectral range increases The success probability satisfies Psucc ≥ ζ squared e−2tE0 + O(ϵfilt), where ϵfilt is the combined quadrature and polynomial error described above Resource Summary The circuit width is logarithmic in the grid and in the quadrature order, and the state preparation cost Cprep is polynomial in N, d, t, D and log(1/ϵ) rather than in the Nd grid size The total training cost is NiterCtrain for Niter optimization steps The number of trainable parameters K ≤ d+r/r− 1 The number of queries to UA per state preparation is q = O√t χmaxα¯A + log(1/ϵ) The imaginary time itself is t = O∆−1 log(1/(ϵζ)) The required circuit width is nqubits = dn + ⌈log2 (n + 1)⌉ + ⌈log2 d⌉ + ⌈log2 D+⌉ + 7 The cost per query to UA is CA = O(d(n squared + rn)) gates per query The total resource requirements are summarized in Table I Coefficient Vector Monomial Dictionary The coefficient vectors and corresponding monomial dictionaries for Figs.

Title and authors: Kai: Let's talk about what this paper actually proposes with the title "Quantum Fokker-Planck solvers as trainable generative models." It suggests using quantum solvers to build generative models that are trained to fit physical stochastic processes.

Mira: Basically, they take a known physics model—the Fokker-Planck equation for an overdamped Langevin process—and solve it using quantum techniques, specifically these cosine operators and singular value transformation. Then they turn the solution into a trainable structure where the parameters theta are coefficients of a block-encoded diagonal potential.

Lev: What I find interesting is that by encoding the square root of the stationary density, they enable sampling through computational-basis measurements, which bypasses some of those complex classical mixing times we usually have to deal with when sampling from these kinds of distributions.

Kai: And it's trainable at scale; they show that this approach remains trainable even when you increase the size and complexity of the problem. They are focusing on the physics information itself during optimization, which is key.

Mira: The paper also highlights that because they're optimizing for a drift field consistent with the data, it leans more toward model discovery than just fitting a generic distribution.

Lev: That physical interpretation of learned parameters as coefficients of a drift field makes sense; if you can identify the underlying physics, you get more robust results on real hardware when we try to run this on something like superconducting qubits.

The paper's summary: Kai: So, looking at the summary of "Quantum Fokker-Planck solvers as trainable generative models," it boils down to them defining a d-dimensional overdamped Langevin process with additive, isotropic noise, defined by d X t = -grad V(X t) dt + sigma d W t.

Mira: Their probability density obeys the Fokker–Planck equation where the probability current J is related to the gradient of V and sigma two/two and when that current is zero, you get an equilibrium stationary density p ss(x) = Z-one (-2V(x)/sigma two).

Lev: The generative model they construct then defines a distribution p theta(x) = one/Z(theta) (-two/sigma squared V(x; theta)), where V is parameterized by coefficients theta from a fixed dictionary of basis functions.

Kai: The quantum implementation workflow is four stages: first, transforming the Fokker–Planck generator into a positive semidefinite operator H = A A, whose ground state is the square root of p theta; then block encoding the discretized A so theta enters only through the diagonal potential term; then preparing that ground state using an imaginary-time filter implemented as a linear combination of cosine operators; and finally, measurement and a loss that drives the update of theta.

Mira: They show this is physical because the fitted theta end up being the coefficients of the drift of a stochastic differential equation that actually matches the data they're fitting.

Lev: The training objective is to minimize distance between generated and empirical distributions using losses like Kullback–Leibler divergence or squared maximum mean discrepancy, which they state is a convex problem with a unique optimum characterized by K moments of the data.

The paper's improvements: Kai: One improvement they discuss is that the optimization problem differs from training a generic parameterized quantum circuit because the optimizer sees K variables regardless of the coarseness of the grid set by N or dimensionality d.

Mira: They say that the only modeling choice they have is the dictionary, which encodes prior knowledge about physical systems in a way that is independent of grid resolution. This makes it closer to model discovery than just fitting a specific distribution.

Lev: And they note that likelihood-based training is convex in natural parameters because the stationary state has an exponential family form, which simplifies the optimization landscape compared to non-convex problems.

Kai: They also give a resource estimation for training cost C train = N eval times n s / P succ times C prep. The circuit width is logarithmic in the grid and in the quadrature order, and the state preparation cost C prep is polynomial in N, d, t, D and (one/epsilon) rather than being dependent on the full grid size <ref:2610.31330#pg2>.

Mira: They also detail approximation errors, saying they are bounded by grid discretization error of order x squared and finite imaginary time error of order e-t <ref:2610.31330#pg2>. The dominant error they see in experiments comes from the finite-order Hubbard–Stratonovich representation.

Lev: That means the success probability is bounded by P succ at least zeta squared e-2tE zero + O(epsilon filt), where epsilon filt includes both quadrature and polynomial errors. It’s important to know these error terms because they dictate how accurate we can expect the sampling to be on real hardware.

Conclusion: Kai: So, wrapping up the paper "Quantum Fokker-Planck solvers as trainable generative models," the main implication is that we can use quantum solvers to build generative models for physical stochastic processes, offering a way to sample from unknown distributions in one shot using computational-basis measurements.

Mira: They've shown that this approach is structured and trainable at scale, with the parameter count K being independent of N, the number of grid points per dimension. This makes it very robust for high-dimensional problems, even when the stationary density isn't a simple Gaussian.

Lev: From a hardware standpoint, they estimate the required circuit width is logarithmic in both the grid and quadrature order, which keeps things manageable compared to methods that scale exponentially with problem dimensions.

Kai: They also point out that this framework opens a route to identifying regimes where quantum state preparation can actually offer an advantage over classical sampling methods for coupled stochastic degrees of freedom.

Mira: The paper emphasizes that the success probability depends on the gap of the operator H, and they found that convergence to the exact distribution happens as they refine these parameters.

Lev: My main concern, as always, is those approximation errors; specifically, the finite-order Hubbard–Stratonovich representation causes issues when evolution time or domain range increases because those oscillatory terms become hard to resolve.

Kai: That's a fair point about the error analysis; knowing where the fidelity drops is crucial for building usable quantum hardware for these tasks.

Mira: So, in summary, this work moves us toward using quantum solvers not just for simulation, but as the core engine of a trainable generative model that learns physical drift fields directly from data.