Mean-Field Path Integral Diffusion

Mean-Field Path Integral Diffusion (MF-PID) is a generative sampling algorithm that extends Harmonic Path Integral Diffusion to an ensemble of interacting particles, where each sample responds to a potential centred on the instantaneous empirical mean of the ensemble Chertkov (2025). The central result is a linear-interpolant theorem: under a time-varying quadratic Vlasov interaction, the mean of a cooperating ensemble follows a straight line in distribution space between its initial and target values, reducing the mean-field problem to a single-particle harmonic diffusion with a moving centre Chertkov (2025). This theorem absorbs two companion extensions that preserve exact solvability while relaxing the rigidity of the static harmonic case: Adaptive PID (AdaPID), which allows the stiffness \(\beta_t\) to follow a piecewise-constant schedule Chertkov & Behjoo (2025), and Guided Harmonic PID (GH-PID), which lets the harmonic centre \(\nu_t\) move in time to produce soft path-space guidance Chertkov (2025).

Motivation: samples that cooperate

Standard Path Integral Diffusion treats samples independently: each trajectory is drawn under a drift determined only by the target density \(p^{\mathrm{tar}}\) and the potential \(V_t\), with no information flowing between samples Behjoo & Chertkov (2025). In many applications, including demand-response control of ensembles of thermostatically controlled loads in energy systems, this independence is a modelling loss: the natural object is an interacting population whose behaviour is shaped by its own empirical distribution Chertkov (2025).

Mean-field coupling captures this structure in the classical Vlasov–McKean style McKean (1966). The ensemble potential takes the form

\[ V^{\mathrm{MF}}_t[x; p_t] = \int V_t(x - y)\, p_t(y)\, dy, \]

so that each sample sees a potential centred on the current population density. When \(V_t\) is quadratic, the interaction reduces to an attraction toward the ensemble mean \(m_t = \int y\, p_t(y)\, dy\) — samples that cooperate by tracking their collective centre Chertkov (2025). The control-theoretic setting is that of mean-field games Huang, Malhamé & Caines (2006), transported into the path-integral framework of Kappen (2005) and Chernyak, Chertkov, Bierkens & Kappen (2014).

AdaPID — time-varying stiffness

Adaptive PID (AdaPID) replaces the constant stiffness \(\beta\) of H-PID with a time-dependent schedule \(\beta_t\) Chertkov & Behjoo (2025). Piecewise-constant schedules \(\beta_t = \beta_k\) on segments \([t_k, t_{k+1}]\) preserve linear solvability: on each segment the Hopf–Cole-linearised operator is a harmonic oscillator with fixed frequency, and the transition between segments is handled by matching Gaussian Green functions at each breakpoint. The resulting drift is expressed in closed form through a Riccati system whose blocks are solved segment by segment Chertkov & Behjoo (2025).

The motivation is speciation control. In generative diffusion, samples commit to modes of a multi-modal target \(p^{\mathrm{tar}}\) at a characteristic time that depends on \(\beta\) Behjoo & Chertkov (2025). Low \(\beta_t\) early encourages mode exploration; high \(\beta_t\) late sharpens commitment. AdaPID introduces a Quality-of-Sampling diagnostic — the distribution of sample margins across modes — to score candidate schedules analytically Chertkov & Behjoo (2025). Because the closed-form drift is differentiable in the breakpoints and stiffness values, schedule optimisation becomes a low-dimensional non-convex problem rather than an inner loop of neural training.

GH-PID — moving harmonic centre

Guided Harmonic PID (GH-PID) extends H-PID in an orthogonal direction: the centre of the harmonic trap is itself a time-dependent control \(\nu_t\) Chertkov (2025). The potential becomes

\[ V_t(x) = \tfrac{\beta_t}{2}\,\|x - \nu_t\|^2, \]

so a prescribed reference trajectory \(\nu_t\) guides samples along a desired path in state space while still enforcing the terminal law \(p^{\mathrm{tar}}\). The Hopf–Cole transform again yields a linear Schrödinger-like equation whose Green functions are Gaussian; the drift picks up an additional term proportional to the difference between the current state and the current centre, and remains a closed-form affine map in \(x\) Chertkov (2025). For Gaussian-mixture targets, the drift admits a further explicit decomposition into per-mode contributions weighted by responsibilities.

GH-PID is the analytical backbone of the mean-field extension: in the quadratic Vlasov case, the interaction \(V^{\mathrm{MF}}_t[x; p_t]\) produces precisely a moving centre \(\nu_t = m_t\), reducing a many-body interacting problem to a single-particle GH-PID whose centre depends on the statistics of the ensemble itself Chertkov (2025).

The linear-interpolant theorem

The central theorem of the mean-field extension can be stated as follows Chertkov (2025):

Let an ensemble of \(N\) samples evolve under the MF-PID dynamics with quadratic Vlasov coupling \(V^{\mathrm{MF}}_t[x; p_t]\) of stiffness \(\beta_t\). In the mean-field limit \(N \to \infty\), the ensemble mean \(m_t = \mathbb{E}_{p_t}[x]\) follows a straight line from the initial mean \(\bar m^{(0)}\) to the target mean \(\bar m^{(1)}\): \[ m_t = (1 - t)\,\bar m^{(0)} + t\,\bar m^{(1)},\qquad t \in [0,1]. \] The ensemble dynamics then decouple into independent GH-PID trajectories with moving centre \(\nu_t = m_t\).

Two properties make the result tractable. First, the linear interpolant is an exact consequence of the quadratic structure of the Vlasov interaction; it is not a small-noise expansion or a heuristic ansatz Chertkov (2025). Second, once the mean trajectory is known, the samples evolve independently under GH-PID with centre \(\nu_t = m_t\), so the full ensemble drift is computable in closed form without resolving a fixed-point iteration over the empirical density. The many-body problem collapses to a mean-field decoupling with an analytic linear mean.

The theorem also supplies a diagnostic: any deviation from a straight-line mean trajectory in an empirical ensemble is evidence of non-mean-field structure — finite-\(N\) corrections, non-quadratic interactions, or heterogeneous samples that break the exchangeability assumption Chertkov (2025).

Relation to H-PID and two-endpoint formulations

MF-PID can be viewed as a mean-field closure of a two-endpoint H-PID problem, in which both the initial distribution \(p^{(0)}\) and the terminal distribution \(p^{(1)}\) are prescribed non-Dirac densities on \(\mathbb{R}^d\) Behjoo & Chertkov (2025); Chertkov (2025). Whereas standard H-PID transports a Dirac at the origin to \(p^{\mathrm{tar}}\), the mean-field setting naturally describes the transport of an ensemble with non-degenerate initial statistics — for instance bridging two operating regimes of a population of buildings in thermostatic control. The linear-interpolant theorem supplies the mean trajectory in this two-endpoint setting analytically, leaving only the residual fluctuations around the mean to be generated by GH-PID sampling.

All three constructions — H-PID, AdaPID, and GH-PID — sit inside the integrability hierarchy of Behjoo & Chertkov (2025), and the mean-field extension inherits their closed-form solvability by reduction to GH-PID at the level of each sample.

Connection to agent dynamics with fixed endpoints

The linear-interpolant theorem admits a simple reading: for an ensemble whose endpoints are pinned, the mean traces the straight-line interpolant between the two endpoints, and individual samples fluctuate around that mean under an exactly solvable drift. This structure is suggestive, though not claimed here as a formal correspondence, as a reference model for collections of trajectories in large language model agent systems where task instances have prescribed start and end states — a nucleation event defining the initial state and a completion event defining the terminal artefact. Under a mean-field quadratic coupling, a typed batch of such trajectories would admit the same linear-mean structure, and empirical deviations from linearity would read as finite-size or heterogeneity effects rather than as primary dynamical information. The analogy is recorded here only to indicate where an exactly solvable bridge from physics meets the data structures of multi-agent systems; the formal correspondence is left to future work.