An implicit–explicit method, usually abbreviated IMEX method, is a time-integration method for differential equations that treats some terms implicitly and others explicitly. The implicit part requires solving equations involving the unknown solution at a new time or stage, while the explicit part uses already available values. This division is useful when different processes impose different computational or stability requirements: a stiff term can be handled implicitly without also placing every expensive or nonlinear term inside the implicit solve.
IMEX methods form a broad class within numerical analysis, rather than a single algorithm. Their effectiveness depends on both the time-stepping formula and the choice of which terms to treat implicitly.
Mathematical setting
Consider an ordinary differential equation written in additive form:
Here may be a scalar or a vector. An IMEX method evaluates explicitly and implicitly. Typically:
- contains nonstiff terms, transport terms, or nonlinear terms that would be expensive to include in an implicit solve.
- contains stiff terms whose explicit treatment would impose a severe time-step restriction.
Stiffness commonly arises when a system contains rapidly decaying components alongside more slowly evolving components. An explicit method may then require very small steps for stability even when the solution of interest changes slowly.
The splitting is not unique. For example, if
where is a stiff linear operator and is a nonstiff nonlinear function, a natural choice is
This leads to linear implicit solves even though the full differential equation is nonlinear. In other applications, is itself nonlinear and requires a nonlinear solver.
First-order IMEX Euler method
The simplest example combines explicit Euler treatment of with implicit Euler treatment of . With a constant step size , it is
Because depends on , this is generally an equation to solve rather than a direct evaluation.
If , the update becomes
where is the identity matrix. Each step therefore requires solving a linear system. When and are constant, a matrix factorization can often be reused.
If is nonlinear, define
The new value satisfies . A solver such as Newton's method may be used, involving the Jacobian matrix
Only the implicitly treated term contributes to this Jacobian. This can make the solve substantially simpler than a fully implicit treatment of .
Under suitable smoothness and stability assumptions, IMEX Euler has first-order global accuracy.
Principal families
IMEX Runge–Kutta methods
An IMEX Runge–Kutta method uses several intermediate stage values within each time step. For an autonomous system
a common form is
followed by
The coefficients form a strictly lower-triangular explicit array. The coefficients usually form a lower-triangular implicit array, so the implicit stages can be solved sequentially. More general coupled implicit arrays are also possible.
These schemes are examples of additive Runge–Kutta methods. Their accuracy depends not only on the orders of the explicit and implicit formulas separately, but also on coupling conditions between them. Pairing two independently chosen methods of order does not automatically produce an IMEX method of order .
For nonautonomous equations, stage times must also be specified consistently.
IMEX linear multistep methods
An IMEX linear multistep method uses values from several preceding time levels. In one common form,
The newest explicit evaluation is omitted, while the implicit sum may include the unknown value .
A familiar second-order example combines Crank–Nicolson treatment of with Adams–Bashforth treatment of :
Multistep methods can reduce the number of evaluations and solves per step, but require a starting procedure. Variable time steps require coefficients appropriate to the changing step ratios.
Stability
The main motivation for IMEX integration is improved numerical stability at lower cost than full implicit integration. This does not make an IMEX method unconditionally stable.
A standard scalar test equation is
where represents the explicitly treated component and the implicitly treated component. IMEX Euler gives
with
For this scalar test, nonamplification requires
If , the method reduces to implicit Euler, which is stable throughout the left half of the complex plane. If , it reduces to explicit Euler and retains that method's bounded stability region. Thus, implicit treatment of one component does not remove restrictions associated with another component.
For matrix-valued problems, eigenvalues help characterize stability, but scalar tests are not a complete description. Noncommuting split operators and transient amplification in nonnormal systems can complicate the analysis.
Damping of fast modes
Implicit formulas differ in how strongly they damp rapidly decaying modes. Implicit Euler strongly damps such modes. Crank–Nicolson, although stable for scalar decay throughout the left half-plane, does not strongly damp modes with very large negative scaled eigenvalues; its amplification factor approaches .
Consequently, an IMEX method based on Crank–Nicolson may exhibit persistent numerical oscillations in stiff components even when it satisfies a stability criterion. Stability, damping, and accuracy are distinct properties.
Application to partial differential equations
IMEX methods are widely used after a partial differential equation has been discretized in space. A finite difference method or finite element method then produces a large system of ordinary differential equations.
Consider the advection–diffusion equation
After spatial discretization, it has the schematic form
where approximates advection and approximates diffusion. A common IMEX choice treats explicitly and implicitly.
For standard discretizations, explicit diffusion typically imposes a restriction of the form
whereas explicit advection typically imposes a Courant–Friedrichs–Lewy restriction of the form
The constants depend on the spatial discretization and time integrator. Implicit treatment of diffusion can remove the explicit diffusive restriction while leaving the advective restriction in place. On fine grids, this can substantially reduce the number of required time steps.
Other applications include:
- Reaction–diffusion equations, with rapid reactions or diffusion treated implicitly.
- Fluid models based on the Navier–Stokes equations, with selected viscous or fast-wave terms treated implicitly.
- Systems combining transport with stiff source terms.
- Higher-order dissipative equations, where explicit treatment of high spatial derivatives can impose especially restrictive time steps.
The appropriate split is problem-dependent: diffusion is not always the stiffest process, and reaction terms are not always best treated implicitly.
Accuracy, conservation, and limiting behavior
Convergence and order reduction
Formal order describes the error as under suitable assumptions. It does not guarantee that the same error behavior persists uniformly as stiffness increases.
IMEX methods may exhibit order reduction: their observed convergence rate can be lower than their classical order, particularly in stiff problems with time-dependent boundary conditions or insufficient stage accuracy. The treatment of boundary data and implicit-solver tolerances can also affect the measured error.
Conservation and positivity
An IMEX scheme does not automatically preserve conservation laws, nonnegativity, or a discrete energy inequality.
For example, let be a linear conserved quantity. If
separately, standard additive updates preserve that quantity in exact arithmetic, assuming the stage equations are solved consistently. Conservation that arises only through cancellation between and needs additional care.
Positivity and monotonicity generally require suitable spatial discretizations, time-integration coefficients, and possibly step-size restrictions. Implicit treatment alone does not establish these properties.
Singular limits
Some systems contain a small parameter:
As tends to zero, solutions may approach an equilibrium constraint such as . An appropriately constructed asymptotic-preserving method can remain stable without resolving the fastest scale and reduce to a consistent discretization of the limiting problem.
Not every IMEX method has this property. Uniform accuracy in is a further requirement, distinct from asymptotic preservation.
Relation to other approaches
IMEX integration is related to, but not identical with, operator splitting. An operator-splitting method typically advances separate subproblems in succession, such as an evolution under followed by an evolution under . An additive IMEX method combines contributions from both terms within a coupled time-stepping formula. Both approaches introduce interactions between the components, but their order conditions and error mechanisms differ.
The term semi-implicit method is often used for IMEX schemes, though its meaning varies. It can also describe linearization, extrapolation of nonlinear coefficients, or implicit treatment of selected variables rather than a fixed additive split.
Computational trade-offs
An IMEX method can be attractive when implicit treatment of a restricted part of the model permits much larger time steps while keeping the solve manageable. Linear stiff terms, reusable factorizations, and structured sparse matrices are particularly favorable.
Its advantages diminish when the explicitly treated terms remain strongly restrictive, when the implicit solve is expensive, or when the desired accuracy already requires small steps. A poor splitting can also leave important fast interactions explicit.
Practical performance therefore depends on several costs together: function evaluations, linear or nonlinear iterations, memory use, communication, and the number of accepted steps. A larger stable step is useful only if the resulting solution is sufficiently accurate and the work per step does not outweigh the reduction in step count.