Follow-The-Leader Shallow Model

• Follow-The-Leader Shallow Model is a reduced framework that connects microscopic car-following dynamics with macroscopic first-order PDEs, exemplified by the LWR conservation law.
• It rigorously derives continuum limits using discrete density reconstructions, BV estimates, and one-sided Oleĭnik conditions to ensure convergence from finite look-ahead and nonlocal interactions.
• The model extends to collective navigation by incorporating nonlocal alignment and leader-follower couplings, elucidating threshold behavior and traveling-front dynamics in complex systems.

Follow-the-Leader (FtL) shallow models are reduced descriptions of interacting agents in which motion or decision dynamics are driven by the state of agents ahead or by a distinguished leader population. In the traffic literature, the canonical FtL model is a microscopic Lagrangian ODE system for vehicle positions whose many-particle limit is the scalar Lighthill–Whitham–Richards (LWR) conservation law; in generalized non-local variants, finite look-ahead, weighted anticipation, and rear-mirror effects still collapse to the same first-order macroscopic flux when the look-ahead horizon is fixed (Holden et al., 2017, Francesco et al., 2014, Holden et al., 2023). In collective-motion and leader–follower systems, the same FtL principle appears directly at the macroscopic level as first-order hyperbolic or drift–diffusion PDEs for follower and leader densities, and in analytically tractable stochastic decision models for follower–leader coupling in patch-leaving dynamics (Bernardi et al., 2019, Bernardi et al., 2024, Moyse et al., 4 Mar 2025).

1. Conceptual scope and state variables

Across the cited literature, FtL is a modeling principle rather than a single equation. In traffic, the interaction variable is typically the headway $x_{i+1}(t)-x_i(t)$, the spacing $s_i(t)=Z_{i+1}(t)-Z_i(t)$, or the discrete density

$\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$

with corresponding Lagrangian spacing variable

$$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$

In swarm and leader–follower PDEs, the state variables are densities such as $\rho_f,\rho_\ell$ or right-/left-moving components $u^\pm,v^\pm$, coupled through nonlocal attraction, alignment, or turning terms (Holden et al., 2023, Bernardi et al., 2019, Bernardi et al., 2024).

The traffic papers use two equivalent monotonicity conventions. One writes the speed either as a strictly decreasing function of density, $v(\rho)$, or as a strictly increasing function of spacing, $V(s)$. In the first convention the macroscopic flux is $\rho\,v(\rho)$; in the second it is $F(\rho)=\rho\,V(1/\rho)$ (Francesco et al., 2014, Holden et al., 2017). This duality is standard in Lagrangian–Eulerian reformulations of FtL systems.

The term “shallow” is explicit in the macroscopic leader–follower PDE literature, where it denotes a first-order closure for population densities and alignment fields (Bernardi et al., 2019). A plausible implication is that, in the broader FtL corpus considered here, “shallow” identifies reduced models whose dominant structure is first-order transport, conservation, or drift, even when they are derived from microscopic particle dynamics.

2. Classical traffic FtL and the LWR continuum limit

The classical one-lane FtL model places particles or vehicles in increasing order,

$s_i(t)=Z_{i+1}(t)-Z_i(t)$0

and updates each follower’s speed from the discrete density immediately to its right. In the Di Francesco–Rosini formulation, with particle mass $s_i(t)=Z_{i+1}(t)-Z_i(t)$1 and total mass $s_i(t)=Z_{i+1}(t)-Z_i(t)$2,

$s_i(t)=Z_{i+1}(t)-Z_i(t)$3

where $s_i(t)=Z_{i+1}(t)-Z_i(t)$4 is Lipschitz–$s_i(t)=Z_{i+1}(t)-Z_i(t)$5, strictly decreasing, with $s_i(t)=Z_{i+1}(t)-Z_i(t)$6 and $s_i(t)=Z_{i+1}(t)-Z_i(t)$7 (Francesco et al., 2014). Holden–Risebro use the equivalent spacing formulation

$s_i(t)=Z_{i+1}(t)-Z_i(t)$8

with $s_i(t)=Z_{i+1}(t)-Z_i(t)$9 smooth, strictly increasing, and $\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$0 (Holden et al., 2017).

Two reconstructions are central in the many-particle analysis. The first is the empirical or atomic measure,

$\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$1

and the second is the piecewise-constant density

$\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$2

The target macroscopic equation is the scalar conservation law

$\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$3

equivalently

$\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$4

with entropy solution understood in the Kružkov sense (Francesco et al., 2014, Holden et al., 2017).

The convergence theory is rigorous in several regimes. Di Francesco–Rosini prove that, for nonnegative initial data in $\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$5 with compact support, the empirical measures converge for each $\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$6 in the $\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$7-Wasserstein topology and the discretised densities converge a.e. and in $\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$8 to the unique entropy solution; the proof uses a discrete maximum principle, discrete BV estimates, and a discrete one-sided Oleĭnik-type condition (Francesco et al., 2014). Holden–Risebro give a short proof under $\rho_i(t)=\frac{\ell}{x_{i+1}(t)-x_i(t)},$9 assumptions on the initial discrete density and a structural bound $$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$0, obtaining

$$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$1

as $$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$2, where $$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$3 is the piecewise-constant density reconstruction (Holden et al., 2017).

These derivations establish a central fact of the FtL shallow paradigm: the first-order LWR PDE is not merely a phenomenological continuum closure, but a rigorous dense-traffic limit of microscopic car-following dynamics.

3. Non-local FtL traffic models and robustness of the macroscopic law

Holden–Risebro generalize the traffic FtL law by allowing each driver to consider an arbitrary but finite number of vehicles ahead, as well as the vehicle directly behind. For $$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$4 vehicles of length $$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$5 on a periodic road of length $$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$6, with rear-bumper positions $$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$7, one fixes a Lipschitz–continuous, non-increasing fundamental diagram

$$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$8

nonnegative weights

$$y_i(t)=\frac{1}{\rho_i(t)}=\frac{x_{i+1}(t)-x_i(t)}{\ell}.$$9

and a rear-mirror parameter $\rho_f,\rho_\ell$0. The local and weighted forward speeds are

$\rho_f,\rho_\ell$1

and the ODE system is

$\rho_f,\rho_\ell$2

with periodicity $\rho_f,\rho_\ell$3 (Holden et al., 2023).

The continuum limit is taken in the regime

$\rho_f,\rho_\ell$4

with $\rho_f,\rho_\ell$5 fixed. The discrete density and spacing variables are

$\rho_f,\rho_\ell$6

and the initial density satisfies $\rho_f,\rho_\ell$7, is $\rho_f,\rho_\ell$8-periodic, and is bounded away from $\rho_f,\rho_\ell$9 and $u^\pm,v^\pm$0 (Holden et al., 2023).

The analytical mechanism is Lagrangian-to-Eulerian. In Lagrangian coordinates one obtains

$u^\pm,v^\pm$1

Discrete entropy inequalities yield uniform $u^\pm,v^\pm$2-stability, BV decay, and time-Lipschitz bounds; a piecewise-constant reconstruction $u^\pm,v^\pm$3 is compact in $u^\pm,v^\pm$4. In Eulerian coordinates, the density

$u^\pm,v^\pm$5

is uniformly bounded in $u^\pm,v^\pm$6, and a subsequence converges to $u^\pm,v^\pm$7. Testing against Kruzhkov entropies identifies the limit and shows that the non-local terms collapse to the local flux $u^\pm,v^\pm$8, giving

$u^\pm,v^\pm$9

as the macroscopic PDE (Holden et al., 2023).

The main theorem states that $v(\rho)$0 converges, up to a subsequence, in

$v(\rho)$1

to the unique entropy solution of the LWR equation. The paper further states that the limit is independent of the finite look-ahead horizon $v(\rho)$2, the precise weights $v(\rho)$3, and the parameter $v(\rho)$4, provided $v(\rho)$5 remains fixed as $v(\rho)$6. This is the precise sense in which the result underscores the robustness of the classical LWR model. By contrast, if $v(\rho)$7 as $v(\rho)$8, or if genuinely long-range kernels scale with $v(\rho)$9, different non-local macroscopic models may arise (Holden et al., 2023).

4. First-order leader–follower PDE formulations

In the swarm and collective-navigation literature, the FtL shallow model is written directly at the PDE level. For fixed populations of leaders and followers, the basic first-order system has leader density $V(s)$0 and follower density $V(s)$1,

$V(s)$2

$V(s)$3

with the follower velocity given by the nonlocal alignment closure

$V(s)$4

Here $V(s)$5 is the interaction kernel, $V(s)$6 is the follower diffusivity, $V(s)$7 is the leader speed, $V(s)$8 is the leader direction, and $V(s)$9 are relative conspicuousness weights (Bernardi et al., 2019).

A more structured hyperbolic formulation splits both followers and leaders into right- and left-moving components $\rho\,v(\rho)$0 and $\rho\,v(\rho)$1, with transport speeds $\rho\,v(\rho)$2 and $\rho\,v(\rho)$3, and turning rates

$\rho\,v(\rho)$4

where the perceived signal is

$\rho\,v(\rho)$5

The attraction term $\rho\,v(\rho)$6 depends on the total density $\rho\,v(\rho)$7, and the alignment term $\rho\,v(\rho)$8 is determined by which neighbors contribute to orientation selection (Bernardi et al., 2024).

The 2024 model distinguishes three degrees of leadership:

Regime	Leader rule	Reported behavior
Indifferent leaders (M1)	Leaders are a fixed translating Gaussian at speed $\rho\,v(\rho)$9; $F(\rho)=\rho\,V(1/\rho)$0	Cohesive and target-directed migration only for short times
Observant leaders (M2)	Leaders reverse direction according to an external cue $F(\rho)=\rho\,V(1/\rho)$1, but do not align to followers	Robust rightward flocking for moderate $F(\rho)=\rho\,V(1/\rho)$2
Persuadable leaders (M3)	Leaders combine social alignment with the external bias $F(\rho)=\rho\,V(1/\rho)$3	Widest parameter band of successful guidance

For followers, M2 and M3 include leader orientations in the alignment term through conspicuousness coefficients $F(\rho)=\rho\,V(1/\rho)$4, with $F(\rho)=\rho\,V(1/\rho)$5 encoding conspicuousness bias. In M3, leaders themselves align socially while also responding to the external cue $F(\rho)=\rho\,V(1/\rho)$6 (Bernardi et al., 2024).

These models make the FtL idea explicit at the population scale: leader influence enters not through a discrete headway law, but through nonlocal alignment operators, velocity-jump closures, and directional switching rates.

5. Thresholds, traveling fronts, and long-time asymptotics

The macroscopic leader–follower PDEs exhibit threshold behavior in alignment, bias, and leader fraction. In the collective-navigation model, linear stability and nonlinear bifurcation analysis give conditions on the alignment strength $F(\rho)=\rho\,V(1/\rho)$7, cue strength $F(\rho)=\rho\,V(1/\rho)$8, conspicuousness ratio $F(\rho)=\rho\,V(1/\rho)$9, and leader-to-follower ratio $s_i(t)=Z_{i+1}(t)-Z_i(t)$00 to secure a unique stable steady state with a $s_i(t)=Z_{i+1}(t)-Z_i(t)$01 rightward orientation and spatial coherence: $s_i(t)=Z_{i+1}(t)-Z_i(t)$02 with

$s_i(t)=Z_{i+1}(t)-Z_i(t)$03

For M3 with $s_i(t)=Z_{i+1}(t)-Z_i(t)$04, numerical values reported are

$s_i(t)=Z_{i+1}(t)-Z_i(t)$05

and

$s_i(t)=Z_{i+1}(t)-Z_i(t)$06

Excessive alignment is not uniformly beneficial: the paper states that $s_i(t)=Z_{i+1}(t)-Z_i(t)$07 may lead to fragmentation (Bernardi et al., 2024).

The same model admits a traveling-front theorem. Under sufficiently large attraction coupling $s_i(t)=Z_{i+1}(t)-Z_i(t)$08, turning rates satisfying $s_i(t)=Z_{i+1}(t)-Z_i(t)$09, and bias parameters satisfying the threshold conditions above, there exists a traveling-wave solution

$s_i(t)=Z_{i+1}(t)-Z_i(t)$10

monotonically connecting left and right states. The wave speed is

$s_i(t)=Z_{i+1}(t)-Z_i(t)$11

and the wave is spectrally stable in the weighted norm if $s_i(t)=Z_{i+1}(t)-Z_i(t)$12 lies strictly between the bifurcation thresholds (Bernardi et al., 2024).

In the fixed-population PDE model, a Peclét-type estimate quantifies the minimal leader number needed for sustained drift. With leaders uniformly distributed over a region of volume $s_i(t)=Z_{i+1}(t)-Z_i(t)$13, the critical number is

$s_i(t)=Z_{i+1}(t)-Z_i(t)$14

and the corresponding critical leader fraction satisfies

$s_i(t)=Z_{i+1}(t)-Z_i(t)$15

The transient/long-time picture is kernel-sensitive: short-range top-hat kernels generate sharper fronts and stronger transient cohesion but may require more leaders to achieve global drift, while long-range Gaussian kernels lower the critical leader number at the expense of weaker local cohesion. If alignment is replaced by an inhomogeneous law $s_i(t)=Z_{i+1}(t)-Z_i(t)$16, then there is a critical leader density $s_i(t)=Z_{i+1}(t)-Z_i(t)$17 separating isotropic diffusion from polarization and collective drift (Bernardi et al., 2019).

These results clarify a recurrent misconception: stronger alignment or stronger leader forcing does not automatically improve coherent guidance. In the cited PDE analyses, successful FtL transport depends on parameter windows, kernel structure, and the way leader bias is integrated into the alignment operator.

6. Related shallow formulations, higher-order variants, and limits of validity

Outside traffic and swarm transport, follower–leader coupling also appears in shallow stochastic decision models. In the patch-leaving framework of social foraging, each agent carries a decision variable $s_i(t)=Z_{i+1}(t)-Z_i(t)$18 that evolves by

$s_i(t)=Z_{i+1}(t)-Z_i(t)$19

with threshold $s_i(t)=Z_{i+1}(t)-Z_i(t)$20, deterministic travel time $s_i(t)=Z_{i+1}(t)-Z_i(t)$21, and reset $s_i(t)=Z_{i+1}(t)-Z_i(t)$22 after moving between two non-depleting patches. Leaders ignore social input; followers incorporate leader signals either through pulsatile arrival/departure updates or through a counting rule based on the instantaneous fraction of leaders in the patch. Under a large-$s_i(t)=Z_{i+1}(t)-Z_i(t)$23 mean-field approximation and time-scale separation, the follower dynamics reduce to effective drift corrections

$s_i(t)=Z_{i+1}(t)-Z_i(t)$24

for pulsatile coupling, and

$s_i(t)=Z_{i+1}(t)-Z_i(t)$25

for counting coupling (Moyse et al., 4 Mar 2025).

The analytical output is renewal-theoretic: one obtains patch-leaving distributions, equilibrium patch residence times, equilibrium accuracy, equilibrium leaving densities, and a cohesion metric $s_i(t)=Z_{i+1}(t)-Z_i(t)$26 based on damped oscillations of patch occupancy. The paper states that counting coupling $s_i(t)=Z_{i+1}(t)-Z_i(t)$27 and pulsatile arrival coupling $s_i(t)=Z_{i+1}(t)-Z_i(t)$28 increase follower accuracy in inferring the higher-quality patch, prolong patch residence, and typically damp oscillations, whereas misinformation through false rewards or biased beliefs reduces follower accuracy and patch residence but may increase cohesion (Moyse et al., 4 Mar 2025). A plausible implication is that FtL shallow modeling functions here as a reduced decision-theoretic analogue of leader-guided transport.

A higher-order boundary case is the Bando–follow-the-leader car-following model, where followers have both positions and velocities: $s_i(t)=Z_{i+1}(t)-Z_i(t)$29 Under assumptions on the leader trajectory and initial data, Cao–Gong–Keimer prove a uniform strict lower and upper bound on headways, hence collision-freeness on $s_i(t)=Z_{i+1}(t)-Z_i(t)$30; when the leader speed is constant $s_i(t)=Z_{i+1}(t)-Z_i(t)$31, the unique equilibrium is $s_i(t)=Z_{i+1}(t)-Z_i(t)$32, $s_i(t)=Z_{i+1}(t)-Z_i(t)$33, and if

$s_i(t)=Z_{i+1}(t)-Z_i(t)$34

all followers converge exponentially to that equilibrium (Cao et al., 27 Feb 2026). This model is not first-order in the PDE sense, but it delineates one route by which inertia and relative-velocity effects extend the FtL framework beyond shallow closures.

The limits of validity are explicit in the continuum-limit literature. For non-local traffic FtL, fixed finite look-ahead collapses to LWR, but if $s_i(t)=Z_{i+1}(t)-Z_i(t)$35 with $s_i(t)=Z_{i+1}(t)-Z_i(t)$36, or if non-compactly supported kernels scale with $s_i(t)=Z_{i+1}(t)-Z_i(t)$37, different non-local macroscopic models may emerge (Holden et al., 2023). The same source lists multi-lane traffic, second-order follow-the-leader laws, networks, non-periodic boundary conditions, and stochastic perturbations as possible extensions. Within the present corpus, these regimes mark the frontier where the shallow FtL description is either generalized or supplanted by richer mesoscopic or higher-order dynamics.