Projected Coupled Diffusion (PCD)

• PCD is a framework characterized by the explicit projection of dynamics and coupling interactions, spanning both stacked-membrane hydrodynamics and constrained generative modeling.
• In membrane hydrodynamics, PCD reduces 3D solvent flows onto 2D membranes to analyze coupled diffusion and hydrodynamic interactions under varying regimes.
• In generative modeling, PCD augments diffusion processes with coupling guidance and projection steps to enforce hard constraints and coordinate joint state sampling.

Projected Coupled Diffusion (PCD) is a term used in two distinct technical literatures. In stacked-membrane hydrodynamics, it denotes the projected in-plane diffusion dynamics of inclusions in two parallel fluid membranes, where three-dimensional solvent hydrodynamics are projected onto two-dimensional membrane planes and the motions in the two membranes are coupled through the surrounding solvent (Ramachandran et al., 2011). In diffusion-based generative modeling, it denotes a test-time framework for constrained joint generation, in which multiple pre-trained diffusion or score models are coordinated by a differentiable coupling potential and projected onto a feasible set at every sampling step (Luan et al., 14 Aug 2025). The common terminology reflects a shared structural pattern—projection plus coupling—rather than a shared disciplinary origin.

1. Terminological scope and disambiguation

The two established uses of PCD differ in state space, governing equations, and objectives. In the membrane setting, the evolving quantities are membrane velocity fields and pair-diffusion tensors. In the generative setting, the evolving quantities are noisy latent or state variables sampled by reverse-time diffusion dynamics under explicit constraints.

Use of “PCD”	Meaning of “projected”	Meaning of “coupled”
Stacked fluid membranes	3D solvent hydrodynamics projected onto 2D membrane planes	Solvent-mediated hydrodynamic interactions between membranes
Test-time joint generation	Euclidean projection onto a feasible set $C$ at each sampling step	Gradient guidance from a joint coupling cost $U(X,t)$

A common source of confusion is to treat PCD as a single method family. The published usage instead spans two formalisms with different mathematical objects: continuum low-Reynolds-number hydrodynamics in one case, and constrained stochastic generative sampling in the other (Ramachandran et al., 2011, Luan et al., 14 Aug 2025).

2. PCD in stacked-membrane hydrodynamics

In "Hydrodynamic coupling between two fluid membranes" (Ramachandran et al., 2011), the physical setting consists of two infinite parallel fluid membranes located at $z=\pm h/2$, separated by distance $h$, and embedded in a three-dimensional incompressible Newtonian solvent. Each membrane is treated as a two-dimensional incompressible viscous fluid sheet with in-plane velocity field $\mathbf{v}^{(i)}(\mathbf{r})$, while the solvent exerts tangential stresses that enter the membrane equations as interfacial traction forces. The particles are point inclusions embedded in the membranes, and the analysis is performed in the pair-mobility limit.

The defining equations are linear Stokes equations in both the membrane and solvent. For membrane $i=1,2$,

$$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$

In the solvent,

$$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$

with stress tensor

$$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$

The formulation assumes two-dimensional incompressible membrane fluids, low Reynolds number, linear response, thermal equilibrium, stick boundary conditions at membrane–solvent interfaces, and vanishing solvent flow far from the membranes. It neglects out-of-plane membrane fluctuations, interleaflet slip within each membrane, and finite-size effects of inclusions. Diffusion is related to mobility through the fluctuation–dissipation theorem,

$$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$

A central length scale is the Saffman–Delbrück length,

$U(X,t)$0

This sets the crossover between quasi-two-dimensional hydrodynamics for $U(X,t)$1 and solvent-controlled behavior for $U(X,t)$2. The presence of the second membrane modifies the effective screening through $U(X,t)$3-dependent coupling functions (Ramachandran et al., 2011).

3. Mobility kernels, diffusion coefficients, and limiting regimes in membranes

In Fourier space, with in-plane wavevector $U(X,t)$4, magnitude $U(X,t)$5, and transverse projector

$U(X,t)$6

the coupled mobility takes a $U(X,t)$7 block form. Using

$U(X,t)$8

the two-membrane kernels are

$U(X,t)$9

$z=\pm h/2$0

with

$z=\pm h/2$1

The symmetry relations are $z=\pm h/2$2 and $z=\pm h/2$3. The cross-membrane kernel carries a $z=\pm h/2$4 factor, which behaves as $z=\pm h/2$5 at large $z=\pm h/2$6, encoding exponential attenuation across the solvent gap (Ramachandran et al., 2011).

Longitudinal and transverse pair-diffusion coefficients are defined by aligning the $z=\pm h/2$7-axis with the line of centers: $z=\pm h/2$8 In scaled variables

$z=\pm h/2$9

the real-space coefficients are Bessel-transform integrals: $h$0 where

$h$1

Several asymptotic regimes are explicit. For $h$2, the membranes effectively decouple, $h$3, and the same-membrane coefficients reduce to the classical single-membrane results. In the quasi-two-dimensional regime $h$4,

$h$5

while for $h$6,

$h$7

Thus longitudinal correlations decay more slowly than transverse ones. The paper attributes this anisotropy to incompressibility and the tensorial structure of the Oseen projector.

The limit $h$8 is also distinctive. Both same- and cross-membrane diffusion approach a rescaled single-membrane form,

$h$9

with effective screening length

$\mathbf{v}^{(i)}(\mathbf{r})$0

The physical interpretation given is increased dissipation due to membrane proximity and halved amplitudes because the stacked pair shares momentum flux into the surrounding solvent (Ramachandran et al., 2011).

The reported numerical trends are monotone decay of $\mathbf{v}^{(i)}(\mathbf{r})$1 with $\mathbf{v}^{(i)}(\mathbf{r})$2, appreciable dependence on $\mathbf{v}^{(i)}(\mathbf{r})$3 for $\mathbf{v}^{(i)}(\mathbf{r})$4, and vanishing cross-membrane coefficients for $\mathbf{v}^{(i)}(\mathbf{r})$5. With water as solvent and lipid bilayer viscosities, $\mathbf{v}^{(i)}(\mathbf{r})$6 is order microns, so membranes within microns can show strong cross-membrane hydrodynamic coupling. Experimental observables proposed for inferring this PCD include correlated displacements in single-particle tracking, analysis of pair-correlation functions, and fluorescence correlation spectroscopy of labeled inclusions.

4. PCD in test-time constrained joint generation

In "Projected Coupled Diffusion for Test-Time Constrained Joint Generation" (Luan et al., 14 Aug 2025), PCD is a training-free, test-time framework for joint sampling of $\mathbf{v}^{(i)}(\mathbf{r})$7 correlated variables $\mathbf{v}^{(i)}(\mathbf{r})$8 from $\mathbf{v}^{(i)}(\mathbf{r})$9 pre-trained diffusion or score models under hard constraints. The joint state is $i=1,2$0, each model provides a score $i=1,2$1, and the feasible set is $i=1,2$2 with $i=1,2$3. Constraints may be expressed as equalities or inequalities, and the paper instantiates them as convex trajectory velocity constraints and latent convex hulls formed by exemplars for image pairs.

The method augments baseline samplers such as Langevin Monte Carlo (LMC), DDPM, and a DPS variant with two operations. The first is coupled guidance: a differentiable potential $i=1,2$4 contributes a gradient to each coordinate update,

$i=1,2$5

in LMC-style form, or

$i=1,2$6

in DDPM-style form. The second is projection: $i=1,2$7 For convex nonempty $i=1,2$8, this guarantees feasibility at every sampling step.

The baseline updates are standard. For unconstrained single-variable LMC,

$i=1,2$9

Projected LMC replaces this by

$$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$0

For DDPM, the per-variable reverse step used in the paper is

$$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$1

For $$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$2, the LMC-style PCD update is

$$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$3

$$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$4

The DDPM-style PCD update first performs the base reverse step, then applies coupled guidance, then projects: $$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$5

$$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$6

$$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$7

$$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$8

The paper also defines a posterior-sampling variant using Tweedie-denoised estimates,

$$\eta_m \nabla^2 \mathbf{v}^{(i)}-\nabla p^{(i)}+\mathbf{f}_s^{(i)}+\mathbf{F}^{(i)}=0, \qquad \nabla\cdot \mathbf{v}^{(i)}=0.$$9

and then uses $$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$0 for guidance. This is presented as a way to evaluate the coupling cost on denoised states rather than directly on noisy intermediates (Luan et al., 14 Aug 2025).

5. Constraint handling, applications, and empirical behavior in generative PCD

Constraint handling is application-specific. For convex trajectory constraints, the paper projects using a batched ADMM routine. For one trajectory $$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$1 with horizon $$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$2, time step $$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$3, and initial point $$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$4, the projection solves

$$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$5

subject to

$$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$6

For latent convex hulls, projection is reduced to the simplex-constrained least-squares problem

$$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$7

which the paper solves by Mirror Descent with a negative-entropy mirror map. Nonconvex constraints are described as conceptually compatible with PCD when an efficient projector or proximal mapping exists, but uniqueness and convergence guarantees do not generally hold.

Three application families are reported: multi-robot motion planning, PushT robot manipulation, and paired face generation. The coupling costs differ by domain. In multi-robot planning, the paper studies a log barrier,

$$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$8

and a squared hinge distance,

$$\eta \,\tilde{\nabla}^2 \mathbf{v}^{(j)}-\tilde{\nabla} p^{(j)}=0, \qquad \tilde{\nabla}\cdot \mathbf{v}^{(j)}=0,$$9

For PushT, it also considers a DPP-based diversity cost

$$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$0

For paired faces, the coupling is an age-group XOR contrast implemented with a latent classifier, and projection is used to enforce membership in exemplar-based latent convex hulls (Luan et al., 14 Aug 2025).

Setting	Selected PCD results	Baseline contrast
Multi-robot, $$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$1, Highways, $$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$2	PCD-SHD: SU$$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$3, RS$$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$4, CS$$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$5, DA$$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$6; PCD-LB: SU$$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$7, RS$$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$8, CS$$\boldsymbol{\sigma}^{(j)}=-p^{(j)}\mathbf{I}+\eta\left[\tilde{\nabla}\mathbf{v}^{(j)}+(\tilde{\nabla}\mathbf{v}^{(j)})^{\mathrm T}\right].$$9, DA$$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$0	Vanilla Diffuser: SU$$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$1, RS$$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$2, CS$$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$3, DA$$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$4; Diffuser+projection: SU$$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$5, RS$$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$6, CS$$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$7, DA$$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$8
Multi-robot, $$\mathbf{D}^{(ij)}(\mathbf{r})=k_B T\,\mathbf{G}^{(ij)}(\mathbf{r}).$$9, Empty, $U(X,t)$00	PCD-SHD: SU$U(X,t)$01, RS$U(X,t)$02, CS$U(X,t)$03, DA$U(X,t)$04; PCD-LB: SU$U(X,t)$05, RS$U(X,t)$06, CS$U(X,t)$07, DA$U(X,t)$08	Illustrates the stronger adherence trade-off of LB
PushT, $U(X,t)$09, $U(X,t)$10	PCD-LB: DTW$U(X,t)$11, DFD$U(X,t)$12, CS$U(X,t)$13, TC$U(X,t)$14; PCD-DPP: DTW$U(X,t)$15, DFD$U(X,t)$16, CS$U(X,t)$17, TC$U(X,t)$18	Baseline DP: DTW$U(X,t)$19, DFD$U(X,t)$20, CS$U(X,t)$21, TC$U(X,t)$22; DP+projection: CS$U(X,t)$23 but lower DTW/DFD than PCD
Paired faces, two-exemplar setup	PCD: XOR$U(X,t)$24, M/F$U(X,t)$25, SE-CLIP$U(X,t)$26, SE-LPIPS$U(X,t)$27, IS-LPIPS$U(X,t)$28	SD+P: XOR$U(X,t)$29, M/F$U(X,t)$30; SD+C+text: XOR$U(X,t)$31, M/F$U(X,t)$32

The reported qualitative pattern is consistent across domains. Coupling-only improves coordination or diversity but fails to enforce hard feasibility constraints. Projection-only enforces feasibility but does not reliably induce desired joint structure. Their combination yields collision-free, velocity-compliant trajectories in robotics, increased pairwise diversity or non-intersection in PushT, and recovery of desired pairwise age contrast in faces. The paper states that projection can suppress diversity and damp guidance, so effective $U(X,t)$33 may need to be increased; on faces, $U(X,t)$34 is reported as useful because projection attenuates guidance (Luan et al., 14 Aug 2025).

The implementation details reported in the paper are correspondingly solver-dependent. For multi-robot planning, $U(X,t)$35 is used for SHD and $U(X,t)$36 for LB; PushT uses $U(X,t)$37 for DPP/DPP-PS and $U(X,t)$38 for LB/LB-PS. ADMM uses penalty $U(X,t)$39, $U(X,t)$40, and tolerance $U(X,t)$41. Mirror Descent uses learning rate $U(X,t)$42 and $U(X,t)$43. DDPM runs typically use $U(X,t)$44 reverse steps, and an optional noise amplification factor $U(X,t)$45 up to $U(X,t)$46 is reported to help recover diversity under projection. Runtime overhead is application-dependent: image pairs are reported as approximately $U(X,t)$47 slower than vanilla DDPM because of per-step projection, while memory overhead is described as negligible.

6. Relations, reductions, limitations, and conceptual significance

The two PCD formalisms share a common compositional idea: a baseline diffusion process is modified by an explicit mechanism of coupling and an explicit mechanism of projection. In membrane hydrodynamics, the projection is geometric reduction from 3D solvent flow to 2D membrane dynamics, and the coupling is long-ranged solvent-mediated hydrodynamic interaction. In constrained generative sampling, the projection is Euclidean projection onto a feasible set, and the coupling is a gradient term $U(X,t)$48 acting on a joint state. This suggests a useful editorial distinction: the acronym is structurally descriptive rather than domain-specific.

Each formulation has sharp reductions to simpler limits. In the membrane case, $U(X,t)$49 recovers the single-membrane Oseen tensor and eliminates cross-membrane diffusion, while $U(X,t)$50 yields a rescaled single-membrane form with $U(X,t)$51 (Ramachandran et al., 2011). In the generative case, PCD reduces to projected diffusion when $U(X,t)$52, and classifier guidance is recovered as a special case when one variable is fixed and $U(X,t)$53 (Luan et al., 14 Aug 2025).

The limitations are also domain-specific. In the membrane setting, the theory is restricted to point particles in infinite planar membranes under linear Stokes flow, with curvature, finite-size effects, interleaflet slip, and out-of-plane fluctuations neglected. The data further notes that finite-size inclusions modify short-distance behavior and that membrane elasticity and fluctuations reduce self-diffusion in isolated membranes but are partly suppressed in stacked or supported geometries. In the generative setting, nonconvex constraint sets may lack efficient exact projectors and lose feasibility or convergence guarantees; per-step projection can induce mode collapse or strong damping of guidance; strong coupling, especially with LB, can reduce data adherence or task completion; and scaling to large $U(X,t)$54 increases pairwise cost complexity as $U(X,t)$55.

Several misconceptions are directly ruled out by the reported results. In the membrane literature, cross-membrane correlations are not generic at arbitrary separation: they are exponentially suppressed for large $U(X,t)$56 through the $U(X,t)$57 factor. In the generative literature, hard constraints are not enforced by coupled guidance alone, and coordination is not reliably obtained by projection alone. The published experiments instead treat feasibility and coordination as separate mechanisms that must be combined when both are required (Ramachandran et al., 2011, Luan et al., 14 Aug 2025).