Linear Chain Trick (LCT) in Delay Systems

Updated 27 January 2026

Linear Chain Trick (LCT) is a method that converts delay-differential equations with memory kernels, such as Erlang kernels, into finite chains of ordinary differential equations.
It systematically replaces distributed delays with cascades of first-order ODEs, enabling rigorous theoretical analysis and practical numerical simulation.
Generalizations like the GLCT extend LCT to phase-type and oscillatory kernels, broadening its application in modeling complex systems such as epidemic and biological dynamics.

The Linear Chain Trick (LCT) is a widely utilized analytical and computational device for converting delay- and memory-dependent dynamical systems—particularly those with distributed delays—into finite-dimensional systems of ordinary differential equations (ODEs). This conversion facilitates both theoretical analysis and algorithmic approaches to model identification, simulation, and control. The LCT is exact under Kirchoff-type integral formulations with Erlang (gamma, integer-shape) kernels and can be generalized to encompass more flexible phase-type kernels or perturbed/oscillatory memory structures.

1. Foundations: Formulation and Derivation

The LCT applies to systems in which the evolution of the state $x(t)\in\mathbb{R}^d$ depends not only on its instantaneous value but also on a convolution integral of its own history with a memory kernel $K(s)$ : $\dot{x}(t) = f\left(x(t), \int_0^\infty K(s)\,x(t-s)\,ds \right)$ with $K(s)$ typically integrable over $[0,\infty)$ . The classic LCT targets the case where $K(s)$ is the Erlang (or gamma with integer shape) distribution: $K_{p,a}(s) = \frac{a^{p} s^{p-1} e^{-a s}}{(p-1)!}$ with mean delay $\mu = p/a$ and variance $\sigma^2 = p/a^2$ .

The main LCT result states that the memory integral

$z_p(t) := \int_0^\infty K_{p,a}(s) u(t-s) ds$

can be represented exactly as the last state in a $p$ -stage ODE cascade: $\begin{aligned} \dot{z}_1(t) &= a (u(t) - z_1(t)),\ \dot{z}_j(t) &= a (z_{j-1}(t) - z_j(t)), \quad j=2,\dots, p. \end{aligned}$ Thus, any delay convolution with an Erlang kernel is equivalent to a finite set of first-order ODEs, reducing the infinite-dimensional DDE to a tractable ODE system (Alanazi et al., 20 Jan 2026, Mielke et al., 18 Oct 2025, Hurtado et al., 2018, Ritschel et al., 2024, Plötzke et al., 2024).

When applied, the original DDE becomes an extended ODE system with both the “main” state variables and the auxiliary chain variables ( $z_j$ ): $\begin{aligned} \dot{x}(t) &= f(x(t), z_p(t)) \ \dot{z}_1(t) &= a(x(t) - z_1(t)) \ \dot{z}_j(t) &= a(z_{j-1}(t) - z_j(t)),\quad j=2,\dots,p. \end{aligned}$ This approach is exact for any convolution against an Erlang kernel.

2. Theoretical Extensions: Generalizations Beyond Erlang

While the classical LCT is strictly valid for Erlang/gamma kernels with integer shape, several generalized frameworks have been proposed:

Generalized Linear Chain Trick (GLCT): The GLCT extends LCT to arbitrary phase-type (PH) distributions, precisely those distributions arising as absorption times of continuous-time Markov chains (CTMCs) with finitely many transient states. A PH distribution is parameterized by an initial substate distribution $\boldsymbol{\alpha}\in\mathbb{R}^k$ and a subgenerator matrix $A\in\mathbb{R}^{k\times k}$ (off-diagonals nonnegative, rows sum to nonpositive values).

The mean-field equations are

$\frac{d\mathbf{x}}{dt} = \boldsymbol{\alpha}^{T}I(t) + A^{T} \mathbf{x}(t)$

where $\mathbf{x} \in \mathbb{R}^k$ represents the occupancy of each transient phase. Any Erlang, hypoexponential, Coxian, or mixture thereof can be handled within this formalism (Hurtado et al., 2018, Hurtado et al., 2020, Hurtado et al., 2020).

Oscillatory and Non-Erlang Kernels: For kernels with explicit oscillatory modulation (e.g., seasonal delay kernels), the convolution term can be decomposed into sums involving damped exponentials and trigonometric functions, leading to real ODE chains augmented by auxiliary variables for each frequency component. The resulting ODE system increases in dimension proportional to the number of harmonics and chain orders represented (Mielke et al., 18 Oct 2025).

Limits of the LCT: The LCT is not applicable for kernels which are non-integer-shape gammas (general $\alpha\notin\mathbb{N}$ ) or for heavy-tailed and multi-modal delay distributions that cannot be captured by finite chains. In such cases, phase-type or spectral (e.g., Laguerre collocation) approaches provide principal alternatives (andò et al., 26 Nov 2025).

3. Practical Methodologies and Algorithmic Implementation

Application of the LCT consists of several systematic steps:

Identify the Memory Kernel: Determine whether the target delay distribution can be represented as an Erlang (or, more generally, phase-type) density. For empirical distributions, fit or approximate a phase-type representation via established tools (e.g., moment-matching, EM algorithms, BuTools suite).
Transform the DDE to ODE: Replace the delay convolution with an ODE chain of auxiliary variables as described above. For mixed or oscillatory kernels, construct a block diagonal system encompassing all chains and their harmonics.
Parameter Identification and Model Fitting: Given observations, fit system parameters (rates, chain lengths, kernel weights) by simulation-based least squares, single-shooting, or sparse regression (as in SINDy with LCT augmentation) (Alanazi et al., 20 Jan 2026, Ritschel et al., 2024).
Numerical Solution: Use standard ODE solvers, leveraging any sparsity or block structure for computational efficiency. Stiff solvers or semi-implicit schemes are recommended for high-dimensional chains or with large/fast chain rates (Plötzke et al., 2024).

A table summarizing typical LCT structures:

Kernel type	ODE Chain Structure	Applicable LCT Variant
Erlang (integer shape)	Linear cascade of $k$ compartments	Classical LCT
Phase-type	Matrix ODE for $n$ transient phases	GLCT
Mixed/oscillatory	Real+complex chains for each kernel component	LCT with harmonics

4. Applications: Identification, Simulation, and Analysis

Model Identification: LCT enables data-driven discovery of nonlinear distributed-delay systems, as in the SINDy-LCT framework (“LCT–SIND³y”), which jointly infers governing ODEs, mean delay, and delay dispersion from time series by augmenting the candidate function library with chain-auxiliary variables. Grid search or more sophisticated optimization selects the best kernel parameters based on out-of-sample mean squared error and information criteria (Alanazi et al., 20 Jan 2026).

Delay Kernel Estimation: For systems where the memory kernel is unknown, it can be approximated as a weighted sum of Erlang kernels (mixed-Erlang) and identified via ODE-based least squares fitting, with the LCT enabling simulation and gradient-based parameter estimation. Fitted kernels converge to the original memory profile as the number of chain components increases (Ritschel et al., 2024).

Simulation of Epidemic and Population Dynamics: In epidemiology, replacement of the exponential waiting-time assumption in compartment models by Erlang or mixed-Erlang (via the LCT) introduces realistic delay lags, sharper change-point dynamics, and more regular epidemic peaks without affecting the final size relations. The number of internal chain stages modulates variance and peak timing (Plötzke et al., 2024).

Oscillatory and Multimodal Memory: For systems with oscillatory memory kernels, the LCT can be modified by incorporating chains for each frequency, enabling analysis of resonance, phase shift, and Hopf bifurcation structure in, e.g., population dynamics under oscillatory external forcing (Mielke et al., 18 Oct 2025).

5. Rigorous Limitations and Extensions

The LCT has critical limitations:

Restriction to Kernel Structure: Exact closure holds only for Erlang and, via GLCT, phase-type models. Arbitrary distributions (Weibull, lognormal, heavy-tailed) require PH approximation, which may introduce additional dimensions and computational cost (Hurtado et al., 2018, Hurtado et al., 2020).
Breakdown for Non-integer Gamma Kernels: For non-integer shape gamma kernels, there is no finite ODE closure; the Laplace transform acquires a non-rational (fractional) power, and only infinite chains or spectral approximations (e.g., pseudospectral discretization) are viable (andò et al., 26 Nov 2025).
Numerical Challenges: Increased number of chain stages $k$ or phases $n$ leads to numerical stiffness, demanding adaptive or implicit solvers, and elevated computational costs scaling linearly or quadratically with system size (Plötzke et al., 2024).
Empirical Delay Heterogeneity: In systems with non-unimodal or time-varying kernels, multiple chains or more elaborate PH representations must be employed for faithful representation. The chain configuration (length, branching, transition rates) directly affects achievable match to empirical dwell-time variance and higher moments (Hurtado et al., 2018).

6. Impact and Domain-Specific Best Practices

The LCT, its phase-type generalizations, and associated model-reduction techniques have had widespread impact in the life sciences, engineering, and physical systems. In infectious disease modeling, LCT-driven compartmental ODEs more accurately reproduce observed lag structures, epidemic peaks, and changepoint responses—critical for public health forecasting. In neural, gene-regulatory, and reactor kinetics contexts, convolutional memory effects are rendered tractable within the ODE paradigm.

Best practices emerging from the literature:

Select chain length $k$ based on empirical mean $T$ and variance $\mathrm{Var}:\; k = \mathrm{round}(T^2/\mathrm{Var})$ (Plötzke et al., 2024).
Fit PH parameters directly to negative-density kernel estimates via EM or moment-matching (Hurtado et al., 2018, Hurtado et al., 2020).
Use GLCT for non-Erlang or multimodal dwell times; prefer LCT when integer-shape gamma is adequate (Hurtado et al., 2018, Hurtado et al., 2020).
Exploit block-tridiagonal structure in Jacobian formation for efficient implicit integration and parallel computation (Plötzke et al., 2024).
In SINDy-type model discovery, augment the candidate library with chain-state features and select delay parameters via BIC or cross-validated MSE (Alanazi et al., 20 Jan 2026).

The LCT remains a foundational technique for bridging biologically or physically realistic delay formulations and mathematically or computationally tractable ODE systems. Its extensions, especially to phase-type and oscillatory memory, expand its application domain while highlighting core mathematical constraints inherent to finite-dimensional ODE representation of infinite-dimensional memory effects.