Generalized Linear Chain Trick (GLCT)

• Generalized Linear Chain Trick (GLCT) is a mathematical framework that converts distributed delay dynamics into finite-dimensional ODE systems using phase-type distributions.
• Built on CTMC foundations, GLCT extends the classical Linear Chain Trick to improve model flexibility, analytical tractability, and applicability across disciplines.
• GLCT enables precise modeling of arbitrary dwell-time distributions through state augmentation and efficient numerical integration in complex systems.

The Generalized Linear Chain Trick (GLCT) is a systematic mathematical framework for incorporating arbitrary phase-type distributed delays and general dwell-time laws into systems of ordinary differential equations (ODEs). GLCT extends the classical Linear Chain Trick (LCT), which is based on Erlang (gamma) distributions, to encompass the full family of phase-type (PH) distributions by exploiting their characterization as absorption time distributions of continuous-time Markov chains (CTMCs). This formalism enables modelers to replace distributed-delay or delay differential equation (DDE) terms with a finite-dimensional ODE system, greatly enhancing the flexibility, analytical tractability, and computational efficiency of compartmental models in biology, epidemiology, ecology, and engineering (Hurtado et al., 2020).

1. Phase-Type Distributions and CTMC Foundations

A phase-type (PH) distribution is defined as the distribution of the absorption time in a finite CTMC with one absorbing state and $k$ transient states, characterized by:

• An initial probability row vector $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$ with $\sum_i \alpha_i=1$, $\alpha_i\geq0$.
• A $k\times k$ subgenerator matrix $\mathbf{T}$, where $T_{ij}\ge0$ for $i\neq j$ and $T_{ii}=-\sum_{j\neq i}T_{ij}$.

Letting $\mathbf{1}$ denote the $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$0-vector of ones, for random variable $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$1, relevant quantities are: $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$2 This general family encompasses exponential, Erlang/gamma, Coxian, hypoexponential, hyperexponential distributions, and any finite PH distribution (Hurtado et al., 2020, Hurtado et al., 2020, Hurtado et al., 2018).

2. Formal Statement and ODE Formulation of the GLCT

GLCT provides a mean-field ODE system equivalent to the introduction of distributed or delayed dynamics through a PH law. If individuals enter state $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$3 at a (possibly time-varying) rate $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$4 and experience a dwell-time $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$5, one introduces $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$6 “phases” $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$7 whose means $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$8 satisfy (Hurtado et al., 2020): $\boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_k)$9 or in component form,

$\sum_i \alpha_i=1$0

The total exit rate from $\sum_i \alpha_i=1$1 at time $\sum_i \alpha_i=1$2 is $\sum_i \alpha_i=1$3, and this flux can be routed to downstream compartments, accordant with the original system logic (Hurtado et al., 2020, Hurtado et al., 2020, Hurtado et al., 2020).

3. Classical Linear Chain Trick as a Special Case

The LCT is a specific instance of the GLCT for Erlang-distributed dwell times. For $\sum_i \alpha_i=1$4 and

$\sum_i \alpha_i=1$5

the ODEs are: $\sum_i \alpha_i=1$6 yielding the classical “linear chain” trick and, for $\sum_i \alpha_i=1$7, standard Erlang($\sum_i \alpha_i=1$8) distribution of delay/dwell time (Hurtado et al., 2020, Hurtado et al., 2020).

4. GLCT for General Distributed Delays and Mixed Kernels

GLCT enables the embedding of distributed-delay terms, such as memory kernel convolutions, into ODE systems. For a DDE

$\sum_i \alpha_i=1$9

if $\alpha_i\geq0$0 is approximated by a mixture of $\alpha_i\geq0$1 Erlang densities: $\alpha_i\geq0$2 auxiliary variables $\alpha_i\geq0$3 satisfy: $\alpha_i\geq0$4 and the total delayed input is

$\alpha_i\geq0$5

Thus, the full ODE system is (Ritschel et al., 2024, Hurtado et al., 2020): $\alpha_i\geq0$6

5. Implementation: Fitting PH Distributions and Model Construction

Practical use of the GLCT involves the following steps (Hurtado et al., 2020, Hurtado et al., 2018):

1. Select or fit a dwell-time law: Empirically estimate a target cumulative distribution $\alpha_i\geq0$7. Fit a PH distribution PH($\alpha_i\geq0$8) of order $\alpha_i\geq0$9 to the empirical or desired distribution, using expectation-maximization, moment-matching, or tools such as BuTools and EMpht.
2. Augment state space: Introduce $k\times k$0 new ODE variables corresponding to the phases.
3. ODE construction: Formulate the ODE subblock $k\times k$1, with $k\times k$2.
4. Exit flux calculation: Compute the aggregate absorption flux as $k\times k$3, and input this to downstream components.
5. Integration: Incorporate the GLCT-derived ODEs with the existing model structure and use standard integrators (e.g., Runge–Kutta) for simulation (Hurtado et al., 2020, Hurtado et al., 2018).

If the delay kernel is not adequately captured by a single PH law, use a finite mixture (hyper-Erlang or hyperexponential), enabling arbitrary approximation of positive-valued distributions, effectively leveraging the density of the PH family (Ritschel et al., 2024, Hurtado et al., 2020).

6. Applications: Model Generalization and Analytical Advantages

GLCT systematically generalizes ODE compartment models with rich dwell-time structure. Examples include:

• Epidemiological models: SEIR/SEIRS models with PH-distributed latent and infectious periods. The entire infected/latent class is replaced by multidimensional phase-chains, yielding increased realism and facilitating the computation of threshold quantities (e.g., $k\times k$4) via structured next-generation matrices (Hurtado et al., 2020, Hurtado et al., 2020).
• Population and ecological models: Stage-structured predator-prey systems (e.g., Rosenzweig–MacArthur), where maturation and lifespan durations can take arbitrary PH forms, reflecting empirical measurements or biological detail (Hurtado et al., 2020, Hurtado et al., 2020).
• Data-driven discovery: Model identification strategies such as SINDy can directly incorporate GLCT-based chain variables to recover distributed-delay equations from time series, estimating delay mean and dispersion within a sparse regression framework (Alanazi et al., 20 Jan 2026).

GLCT affords several advantages:

• Flexibility: It contains the LCT, Coxian, mixture models, and allows state-dependent or time-varying PH parameters.
• Analytical tractability: Closed-form moment and Laplace-transform expressions enable direct analysis of dynamical properties (e.g., stability, bifurcation).
• Computational efficiency: The matrix–vector formulation exploits optimized linear algebra back-ends and accommodates high-dimensional distributed-delay models with computational ease.
• Structural clarity: The underlying CTMC structure grounds the modeling in first-principles probabilistic interpretations (Hurtado et al., 2020, Hurtado et al., 2020).

7. Assumptions, Limitations, and Extensions

Assumptions include sufficient regularity for ODE well-posedness, non-negative, normalized memory kernels, and tractability of the fitted PH law (Ritschel et al., 2024, Hurtado et al., 2020). Limitations may arise if the number of phases or mixture components is too large, leading to stiffness or identifiability issues. Extensions under development include:

• Generalization beyond Erlang to arbitrary PH structures and competing Poisson processes (Hurtado et al., 2018).
• Mixed or state/time-dependent phase transitions for modeling seasonality, feedback, or other dynamic effects (Hurtado et al., 2020).
• Use in data identification pipelines, where delay-related parameters and the governing ODE form are co-inferred from times series via machine learning frameworks (Alanazi et al., 20 Jan 2026, Ritschel et al., 2024).

GLCT unifies and extends the LCT and its variants, providing a comprehensive, theoretically grounded, and computationally robust methodology for embedding arbitrary delay/distributional assumptions in ODE-based models (Hurtado et al., 2020, Hurtado et al., 2020, Hurtado et al., 2020, Hurtado et al., 2018).