Dispersion Formation Control

• Dispersion Formation Control is a framework that regulates the spread of multi-agent distributions via global measures like covariance spectra and statistical laws.
• It employs both centralized and distributed methods, incorporating dynamic consensus and stochastic formulations to guarantee convergence under defined conditions.
• The approach spans diverse applications—from robotics and fluid dynamics to photonics—using spectral-phase engineering and geometric constraints to manage dispersion.

Dispersion formation control denotes a family of control objectives in which the regulated quantity is a measure of spread, dispersion, or dispersive structure rather than only a set of local geometric constraints. In the literature, this includes direct regulation of the covariance spectrum of a multi-agent distribution, relaxation of covariance steering to minimum-dispersion control for stochastic diffusion processes, geometric or pressure-mediated tuning of effective dispersivity in channels and networks, and spectral-phase engineering for pulse formation, compression, and dispersive wavebreaking (Chen et al., 24 Sep 2025, Chertovskih et al., 2024, Lee et al., 2021, Geng et al., 14 Apr 2026, Anderson et al., 2018). The common feature is that the controlled object is global or distributed—such as a covariance matrix, a law $\mu_t$, an effective diffusivity $D_e$, a group-delay profile $\tau_g(\lambda)$, or a wavebreaking profile—while the actuation is typically local, boundary-based, or geometrically encoded.

1. Conceptual scope and mathematical definitions

A recent formulation makes the notion explicit for multi-agent systems by defining the spatial mean and covariance of agent positions as

$$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$

with the control objective specified through covariance similarity: $$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$ The corresponding dispersion error is the eigenvalue mismatch $e_\lambda = [e_1,\ldots,e_d]^\top$, $e_j=\lambda_j-\lambda_j^*$. Because $C$ is computed in barycentric coordinates, the objective is invariant to global translations, and its eigenvalues are invariant to rotations (Chen et al., 24 Sep 2025).

A stochastic-control formulation uses a different, law-level notion of dispersion. On a fixed horizon $I=[0,T]$, the controlled diffusion

$$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$

is embedded into a Mayer problem $D_e$0. Dispersion is encoded either through terminal moments around a target $D_e$1 or through the trace of the covariance,

$D_e$2

which is made linear in the law by lifting to a product-space process $D_e$3 with an independent copy $D_e$4 (Chertovskih et al., 2024).

These definitions distinguish dispersion formation control from classical formation control based on distances, bearings, or relative positions alone. A common misconception is that dispersion objectives are merely another parameterization of shape constraints. The covariance-based formulation instead targets a global variable associated with the agent distribution, and the stochastic formulation targets a law-dependent terminal quantity rather than a prescribed geometric embedding (Chen et al., 24 Sep 2025, Chertovskih et al., 2024). This suggests that “formation” in this context refers as much to distributional structure as to rigid geometry.

2. Law-level and covariance-spectrum control

For covariance-spectrum regulation, the centralized control law is expressed in barycentric coordinates $D_e$5. If $D_e$6 are orthonormal eigenvectors of $D_e$7 with eigenvalues $D_e$8, and $D_e$9, then the controller is

$\tau_g(\lambda)$0

This law preserves the centroid,

$\tau_g(\lambda)$1

and preserves the eigenvectors of $\tau_g(\lambda)$2. The induced eigenvalue dynamics are

$\tau_g(\lambda)$3

with Lyapunov function

$\tau_g(\lambda)$4

For $\tau_g(\lambda)$5, the paper proves almost global exponential convergence, excluding the measure-zero set corresponding to $\tau_g(\lambda)$6 (Chen et al., 24 Sep 2025).

A distributed realization reconstructs barycentric coordinates and covariance through dynamic average consensus. The local estimators are

$\tau_g(\lambda)$7

and

$\tau_g(\lambda)$8

With a slow-fast cascade parameterized by $\tau_g(\lambda)$9 and $$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$0, singular perturbation analysis yields convergence to $$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$1 for sufficiently small $$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$2, assuming an undirected connected graph, aligned frames, $$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$3, and $$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$4 (Chen et al., 24 Sep 2025).

The law-level stochastic formulation reaches a similar objective through Fokker–Planck duality. If $$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$5, then

$$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$6

and the backward dual variable $$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$7 solves

$$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$8

The central analytical result is an exact “$$p_c = \mu = \frac{1}{N} \sum_{i=1}^N p_i, \qquad C = \Sigma = \frac{1}{N} \sum_{i=1}^N (p_i - p_c)(p_i - p_c)^\top,$$9-order variation” of the cost: $$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$0 where $$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$1. The induced law-feedback control is

$$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$2

A Krasovskii–Subbotin constructive motions algorithm with Monte Carlo estimation of $$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$3 and empirical laws implements this controller numerically (Chertovskih et al., 2024).

In the stochastic Theta-model experiment, the method learns a pseudo-Markovian control structure $$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$4. With $$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$5, $$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$6, $$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$7, $$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$8, and $$C \sim C^* \Longleftrightarrow C \text{ and } C^* \text{ have identical spectra}.$$9 steps per unit time, the optimization converges in 3 iterations for $e_\lambda = [e_1,\ldots,e_d]^\top$0, reducing the average cost from $e_\lambda = [e_1,\ldots,e_d]^\top$1 to $e_\lambda = [e_1,\ldots,e_d]^\top$2, while $e_\lambda = [e_1,\ldots,e_d]^\top$3 produces an even stronger denoising effect (Chertovskih et al., 2024).

3. Distributed geometric realizations in multi-agent formations

A symmetry-based distributed strategy enforces dispersed formations through rotational constraints alone. Agents obey

$e_\lambda = [e_1,\ldots,e_d]^\top$4

on a cycle graph $e_\lambda = [e_1,\ldots,e_d]^\top$5, but the interaction graph $e_\lambda = [e_1,\ldots,e_d]^\top$6 can be a spanning-tree subgraph with only $e_\lambda = [e_1,\ldots,e_d]^\top$7 edges. For rotation angle $e_\lambda = [e_1,\ldots,e_d]^\top$8,

$e_\lambda = [e_1,\ldots,e_d]^\top$9

the canonical regular-$e_j=\lambda_j-\lambda_j^*$0-gon constraint is $e_j=\lambda_j-\lambda_j^*$1. The potential

$e_j=\lambda_j-\lambda_j^*$2

induces the gradient flow $e_j=\lambda_j-\lambda_j^*$3, equivalently $e_j=\lambda_j-\lambda_j^*$4, where $e_j=\lambda_j-\lambda_j^*$5. The resulting matrix-weighted Laplacian is positive semidefinite with $e_j=\lambda_j-\lambda_j^*$6 and $e_j=\lambda_j-\lambda_j^*$7, yielding exponential convergence to the rotationally symmetric manifold $e_j=\lambda_j-\lambda_j^*$8. A maneuvering extension adds feedforward terms for translations, rotations, and scalings through $e_j=\lambda_j-\lambda_j^*$9, $C$0, and $C$1 (Martinez et al., 1 Oct 2025).

A 3D distributed controller for quadrotors realizes dispersion-like formations from local relative positions alone. Each agent uses

$C$2

with

$C$3

and optional scale augmentation

$C$4

where $C$5 or $C$6. Gains are computed offline from the SDP

$C$7

with $C$8. The design requires no global position information or inter-vehicle communication and remains convergent when each actual velocity satisfies the positive-projection condition $C$9 (Fathian et al., 2018). In this framework, uniformly spaced planar or 3D dispersion formations are realized by choosing $I=[0,T]$0 and $I=[0,T]$1 as a regular $I=[0,T]$2-gon, lattice patch, regular polyhedron, or spherical-shell arrangement (Fathian et al., 2018).

An earlier artificial-swarm formulation uses a virtual spring–damper model with Newtonian dynamics,

$I=[0,T]$3

where $I=[0,T]$4. For leaderless dispersion, each agent selects its three closest neighbors $I=[0,T]$5 and imposes the trigonal-planar condition

$I=[0,T]$6

Simulation examples with 50 and 100 agents show that sensing range $I=[0,T]$7 controls whether the swarm remains fragmented or becomes a single dispersed connected component (Jeong et al., 2014).

These geometric realizations clarify a second misconception: dispersion formation control does not necessarily eliminate explicit shape information. In some formulations, such as covariance-spectrum regulation, the target is purely distributional; in others, such as rotational symmetry or equal-distance 3D formations, dispersion is encoded by a sparse set of symmetry or spacing relations rather than by full rigidity constraints (Chen et al., 24 Sep 2025, Martinez et al., 1 Oct 2025, Fathian et al., 2018).

4. Geometry-, flow-, and transport-based control of dispersion

In microfluidic Taylor–Aris dispersion, the controlled quantity is the effective axial dispersivity

$I=[0,T]$8

where $I=[0,T]$9 is determined entirely by cross-sectional geometry. For bowed rectangular channels with aspect ratio $$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$0 and deflection $$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$1, inward bowing flattens the velocity profile and reduces dispersion. The asymptotic optimum is

$$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$2

Experiments on a five-layer microchip with $$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$3 found inward-bowing minima consistent with this prediction; for $$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$4, the minimum occurred near $$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$5, close to $$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$6, and varying $$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$7 changed $$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$8 by nearly an order of magnitude (Lee et al., 2021).

A more general geometric formulation studies periodically corrugated axisymmetric channels with local radius $$X_t = x_0 + \int_0^t f_s(X_s,u_s)\,ds + \int_0^t \sigma_s(X_s,u_s)\,dW_s$$9. The long-time effective diffusivity

$D_e$00

admits an exact auxiliary-cell representation. In the Fick–Jacobs regime $D_e$01,

$D_e$02

whereas in the wide-channel limit $D_e$03,

$D_e$04

The analysis distinguishes smooth-neck and compartmentalized channels, and in several regimes connects $D_e$05 to mean first-passage times through $D_e$06 (Mangeat et al., 2017).

In drinking water networks, dispersion becomes a controlled transport mechanism within a multi-species advection–dispersion–reaction model,

$D_e$07

with dynamic regime switching based on

$D_e$08

The model couples chlorine, a fictitious reactant, and THMs, and embeds the discretized dynamics into an MPC problem with state and actuation constraints. A time-dependent controllability analysis uses the finite-horizon controllability matrix and Gramian to weight booster stations according to current hydraulic reachability (Elsherif et al., 2024). The results show that including dispersion materially changes predicted residuals and byproducts: in the BLA-M network, an advection-only model underestimates chlorine at Junction J1 by approximately $D_e$09 relative to ADR and by approximately $D_e$10 at J2; dynamic ADR/AR switching adds only approximately $D_e$11 runtime overhead (Elsherif et al., 2024).

At the nanoscale, a structured optical field provides another transport-based realization. In an optical vortex lattice generated by two perpendicular standing waves with phase difference $D_e$12, flexible dumbbells formed by two $D_e$13 nm gold spheres joined by a polymer can disperse, rotate, or remain trapped depending on chain length and optical intensity. For $D_e$14, the enhanced diffusivity satisfies the empirical scaling

$D_e$15

and reaches a pronounced maximum near $D_e$16. For some parameters, the displacement distribution develops exponential tails while the MSD remains linear in time, producing a Brownian yet non-Gaussian regime. Hydrodynamic coupling is essential to this effect, while secondary optical scattering tends to delay the return to Gaussian statistics (Meléndez et al., 2018).

5. Spectral-phase engineering, pulse formation, and dispersive wave control

In integrated photonics, dispersion formation control is realized by engineering spectral phase directly. Meter-long chirped spiral Bragg gratings on ultra-low-loss SiN use a linearly chirped Bragg period

$D_e$17

with local Bragg mapping

$D_e$18

to synthesize a target reflection phase $D_e$19, group delay

$D_e$20

and dispersion

$D_e$21

Apodized index modulation suppresses side-lobes and group-delay ripple. On a SiN platform with propagation loss approximately $D_e$22 dB/m, a meter-scale grating folded into an Archimedean spiral occupies approximately $D_e$23 and yields nanosecond-scale delays with low insertion loss (Geng et al., 14 Apr 2026).

Two experimentally emphasized regimes illustrate the design trade-off. A narrowband, high-dispersion CSBG with $D_e$24 m achieved a measured bandwidth of approximately $D_e$25 nm centered at $D_e$26 nm and $D_e$27 ns/nm, with no observable group-delay ripple. A broadband CSBG with $D_e$28 m achieved a dispersion bandwidth of approximately $D_e$29 nm and $D_e$30 ps/nm, with ripple $D_e$31 across the central band. The associated dispersion–bandwidth products were approximately $D_e$32 ns·nm for the high-dispersion device and at least $D_e$33 ns·nm for the broadband device, while on-chip insertion loss remained approximately $D_e$34 dB (Geng et al., 14 Apr 2026). A common misconception in this setting is that the relevant dispersion is simply intrinsic waveguide GVD; the analysis states explicitly that dispersion here is dominated by the distributed Bragg reflection phase rather than intrinsic material or waveguide GVD (Geng et al., 14 Apr 2026).

These phase profiles support pulse formation and compression of a $D_e$35-GHz electro-optic comb. An initial pulse with FWHM approximately $D_e$36 ps was compressed after CSBG reflection to autocorrelation-derived FWHM values of $D_e$37 ps at $D_e$38 nm, $D_e$39 ps at $D_e$40 nm, and $D_e$41 ps at $D_e$42 nm. The on-chip average and peak powers after compression were approximately $D_e$43 mW and $D_e$44 W, respectively, and the same on-chip compressed comb enabled wavelength-swept CARS microscopy with improved temporal stability relative to fiber-based compression (Geng et al., 14 Apr 2026).

An earlier microresonator implementation achieves broadband dispersion engineering by using multiple concentric silica wedges as an on-chip analog of multi-cladding fibers. In wedge-disk resonators with diameter approximately $D_e$45 mm, wedge angles and radial positions tune $D_e$46 and $D_e$47 over an octave-spanning band from approximately $D_e$48 nm to $D_e$49 nm while maintaining $D_e$50. Double-wedge devices tune $D_e$51 into the $D_e$52–$D_e$53 kHz range at $D_e$54 nm, and quadruple-wedge devices tune $D_e$55 to approximately zero or positive values around the pump (Yang et al., 2015).

In attosecond pulse formation, the control variable is the driver chirp or GDD rather than a geometric phase structure. For a chirped Gaussian driver with spectral phase $D_e$56, TDSE calculations and experiment show that isolated attosecond pulse CEP jitter is minimized when the driving pulse is near its Fourier limit but with slightly negative chirp. In the wedge-scan coordinate, the optimum occurs near $D_e$57 rad, where the transfer sensitivity

$D_e$58

is smallest (Kothe et al., 2017). The same work emphasizes that a wedge scan is not a pure CEP offset: wedge insertion changes $D_e$59, $D_e$60, and higher-order dispersion together (Kothe et al., 2017).

Dispersive hydrodynamic wavebreaking provides a boundary-controlled analog. In a viscous fluid conduit, the dispersionless limit of the conduit equation is

$D_e$61

with characteristics $D_e$62. By tracing characteristics backward from a desired breaking profile $D_e$63, the boundary input obeys the implicit map

$D_e$64

which yields explicit boundary waveforms for step, box, triangle, and N-wave targets. Experiments and simulations achieved better than $D_e$65 relative error in breaking height and $D_e$66 in breaking time, and numerical predictions agreed with the full conduit equation within $D_e$67 in breaking height and $D_e$68 in breaking time (Anderson et al., 2018).

6. Metrics, misconceptions, and emerging directions

Across these domains, evaluation is organized around different but structurally analogous metrics. Covariance-based formation control uses eigenvalue errors $D_e$69, Lyapunov decay, and convergence of $D_e$70 (Chen et al., 24 Sep 2025). Law-based stochastic control tracks terminal cost decrement through the exact increment formula and its Monte Carlo implementation (Chertovskih et al., 2024). Geometric transport problems use $D_e$71, $D_e$72, $D_e$73, $D_e$74, and MFPT-based asymptotics (Lee et al., 2021, Mangeat et al., 2017). Water-network control uses residual and THM box constraints together with controllability ranks and Gramian traces (Elsherif et al., 2024). Photonic systems use group-delay linearity, ripple, insertion loss, bandwidth, and DBP; attosecond stabilization uses $D_e$75 and $D_e$76; wavebreaking control uses breaking height, breaking time, and the slope-inflection criterion $D_e$77 (Geng et al., 14 Apr 2026, Kothe et al., 2017, Anderson et al., 2018).

Several recurring misconceptions are explicitly contradicted by the literature. First, dispersion formation control is not identical to conventional rigid-shape control; regulating a covariance spectrum or a law can leave many admissible geometric realizations (Chen et al., 24 Sep 2025, Chertovskih et al., 2024). Second, some “dispersed” geometric formations still require explicit structural encoding: equal distances, rotation symmetries, or graph rigidity do not emerge automatically from the word “dispersion” alone (Fathian et al., 2018, Martinez et al., 1 Oct 2025). Third, in optical pulse control, wedges should not be treated as pure CEP shifters, and in CSBGs the dominant dispersion mechanism is the distributed Bragg reflection phase, not intrinsic GVD (Kothe et al., 2017, Geng et al., 14 Apr 2026). Fourth, in reactive transport networks, advection-only models can materially bias both residual and DBP predictions, especially in low-velocity regions (Elsherif et al., 2024).

The limitations identified in the cited work are likewise domain-specific but conceptually aligned. Law-feedback Fokker–Planck control can suffer discontinuity and possible ill-posedness of the nonlocal feedback PDE, and high-dimensional Monte Carlo estimation remains variance-sensitive (Chertovskih et al., 2024). Covariance-spectrum control permits many admissible terminal configurations and becomes only asymptotically convergent for rank-deficient $D_e$78 (Chen et al., 24 Sep 2025). Rotational-symmetry formulations admit collapsed equilibria when the projection onto the symmetric manifold is zero, and collision avoidance is not built into the basic potential (Martinez et al., 1 Oct 2025). Quadrotor dispersion formations require appropriate sensing graphs and do not autonomously “spread out” without encoded pattern or distance data (Fathian et al., 2018). Microchannel and corrugated-channel asymptotics depend on shape idealizations, while water-network MPC inherits model-calibration and linearization errors (Lee et al., 2021, Mangeat et al., 2017, Elsherif et al., 2024).

The stated future directions indicate a widening scope for the field. On the stochastic side, higher-order adjoint constructions for $D_e$79-polynomial costs, variance reduction, density-constrained formations, and mean-field games are natural extensions (Chertovskih et al., 2024). On the multi-agent side, open problems include delays, noise, asynchronous updates, higher-order dynamics, obstacle constraints, and tracking time-varying dispersion or higher moments beyond covariance (Chen et al., 24 Sep 2025). In integrated photonics, extending SiN spiral Bragg structures beyond $D_e$80 m is presented as a route to delays $D_e$81 ns and larger DBP, while co-integration with sources, modulators, and detectors is identified as a scalable architecture for compact “dispersion engines” (Geng et al., 14 Apr 2026). Taken together, these directions suggest that dispersion formation control is evolving from a collection of domain-specific techniques into a broader systems framework for shaping spread, spectral phase, and macroscopic transport through distributed actuation and structured dynamics.