Global-in-Time Iterative Decoupled Methods

• Global-in-time iterative decoupled algorithms are numerical techniques that solve over full time intervals, ensuring uniform stability and error control.
• They decouple complex multiphysics problems into independent subproblems (e.g., flow, mechanics, interface) to enable efficient, parallel computation.
• The methods employ reformulations like total-pressure substitution and interface reduction, backed by rigorous convergence analysis and parameter studies.

Global-in-time iterative decoupled algorithms are numerical methods for evolutionary and multiphysics problems in which the object of iteration is the solution over an entire time interval, or over large time blocks, while the constituent subphysics, subdomains, or operator blocks are solved separately inside each iteration. In the strict sense, the iteration acts on time trajectories, space–time interface variables, or all-at-once space–time algebraic systems; in a weaker but common sense, the algorithm is local in time yet admits stability or error estimates that are uniform in time. This distinction is explicit in recent work on Biot and MPET systems, nonlinear Stokes–Darcy coupling, and time-parallel parabolic solvers (Gu et al., 2024, Cai et al., 2023, Hoang et al., 2020, Neumuller et al., 2018).

1. Conceptual scope and defining distinctions

In strict usage, “global-in-time” means that the coupled problem is posed and iterated on the full interval $(0,T)$ at once, rather than advanced sequentially with localized interface or coupling iterations. This is the formulation adopted for the nonlinear Stokes–Darcy interface problem, where the interface unknown $\lambda$ is defined on $\Gamma\times(0,T)$ and the interface equation is solved over the whole time interval by a nested Newton–Krylov method (Hoang et al., 2020). The same terminology is used for Crank–Nicolson- and BDF-based Biot algorithms in which one outer fixed-stress iteration updates the full temporal history $\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$, with a forward flow sweep and a mechanics step that is independent across time levels (Gu et al., 2024, Gu et al., 6 Aug 2025).

A weaker usage appears in time-stepping iterative decouplings. In the total-pressure MPET algorithm, the splitting is performed at each time step $t^n$, and the inner iteration converges to the monolithic solution at that time level; nevertheless, the stability and error bounds are uniform in time, and the contraction factor is independent of $\Delta t$ (Cai et al., 2023). The three-field Biot algorithm based on total pressure has the same character: the iteration is local in time, but its convergence requires no restriction on $\Delta t$ or $h$, and the coupled backward Euler scheme satisfies discrete energy estimates that are uniform over the entire interval (Gu et al., 2022). This suggests that the phrase “global-in-time” in the literature refers both to genuine space–time iteration and to time-marching schemes with uniform-in-time control, and those two meanings should not be conflated.

A related but distinct class consists of all-at-once time-parallel solvers. For implicit Euler discretizations of linear parabolic problems, the standard nonsymmetric time-global system $\mathbf B\mathbf u=\mathbf f$ is reformulated as a symmetric saddle-point system, and the resulting inexact Uzawa iteration acts on all time levels simultaneously (Neumuller et al., 2018). For block $\alpha$-circulant preconditioned all-at-once schemes, the global error iteration is analyzed directly in terms of the stability function of the underlying one-step or multistep integrator (Wu et al., 2021).

2. Reformulation patterns and operator structure

A recurring structural device is the introduction of a total-pressure variable. In Biot’s three-field formulation,

$\lambda$0

while in the quasi-static MPET system,

$\lambda$1

These substitutions transform the original displacement–pressure formulation into a generalized Stokes problem for $\lambda$2 coupled to a parabolic problem for the pressure variables (Gu et al., 2022, Cai et al., 2023). The immediate algorithmic consequence is a canonical two-block split: a Stokes-type saddle-point solve for mechanics and a diffusion–reaction solve for transport.

The same separation appears in the Crank–Nicolson and BDF-based global-in-time Biot schemes. In the global-in-time Crank–Nicolson algorithm, the flow step computes the entire sequence $\lambda$3 using the previous iterate of the total pressure differences, and the mechanics step then solves $\lambda$4 independently for each $\lambda$5 (Gu et al., 2024). In the BDF-2 full-order global-in-time algorithm, the diffusion subproblem is

$\lambda$6

followed by a generalized Stokes problem for $\lambda$7 at all time levels (Gu et al., 6 Aug 2025).

A second major reformulation pattern is interface reduction. In nonlinear Stokes–Darcy coupling, a Lagrange multiplier

$\lambda$8

represents the interface normal stress. This yields space–time Steklov–Poincaré operators $\lambda$9 and $\Gamma\times(0,T)$0, and the full coupled evolution is reduced to the interface equation $\Gamma\times(0,T)$1 on $\Gamma\times(0,T)$2 (Hoang et al., 2020). Closely related interface formulations appear in recent fluid–structure domain decomposition methods with local time stepping, where the coupled system is again posed as a space–time interface problem (Kunwar et al., 2024).

A third pattern is all-at-once operator symmetrization. For implicit Euler discretizations of linear parabolic equations, the nonsymmetric time-global matrix $\Gamma\times(0,T)$3 is left-preconditioned by $\Gamma\times(0,T)$4, producing the symmetric positive definite operator

$\Gamma\times(0,T)$5

which is then embedded in a symmetric saddle-point system (Neumuller et al., 2018). This algebraic recasting is the basis for robust global-in-time preconditioning.

3. Principal algorithmic families

The contemporary literature contains several distinct but structurally related families.

Family	Representative structure	Decoupling mechanism
Total-pressure poromechanics	generalized Stokes block + parabolic block	alternate transport and mechanics solves
Space–time interface methods	interface equation on $\Gamma\times(0,T)$6	independent subdomain solves with interface updates
All-at-once time-parallel solvers	global space–time linear system	Fourier/DST or circulant diagonalization across time
DG correction iterations	DG weak formulation on each time slab	preconditioned Picard or correction equations

In total-pressure poromechanics, the split may be time-stepping or genuinely global-in-time. The MPET method alternates a symmetric positive definite pressure solve and a generalized Stokes solve at each time level (Cai et al., 2023). The Biot Crank–Nicolson and BDF-2 methods extend this idea to strict global-in-time fixed-stress iterations, with the mechanics stage partially parallel-in-time (Gu et al., 2024, Gu et al., 6 Aug 2025). The unified thermo-/multiple-network poroelasticity model further generalizes the pattern to displacement, two generalized pressures, and total pressure, with a transport block solved over the full interval and a mechanics block decoupled from it at each iteration (Gu et al., 31 Mar 2026).

In interface methods, the nonlinear Stokes–Darcy problem is solved by a nested Newton–Krylov procedure in which each Krylov matrix–vector product requires one linearized Stokes and one linearized Darcy solve over the full time interval (Hoang et al., 2020). In global fluid–structure domain decomposition, Steklov–Poincaré and Robin formulations again shift the iteration to interface traces on space–time slabs (Kunwar et al., 2024).

In all-at-once time-parallel methods, the emphasis is on linear algebra. The inf-sup-based saddle-point reformulation for parabolic equations uses an inexact Uzawa iteration with a time-parallel preconditioner built from a Discrete Sine Transform (Neumuller et al., 2018). The block $\Gamma\times(0,T)$7-circulant class instead diagonalizes an $\Gamma\times(0,T)$8-circulant temporal preconditioner by FFT, so each iteration reduces to independent solves for the time modes (Wu et al., 2021).

Finally, DG-based iterative time integration starts from the weak DG formulation on each interval and applies a preconditioned Picard iteration. This yields explicit, implicit, and semi-implicit SDG schemes with arbitrary order, and a multilevel extension intended for coupling with space–time multigrid and parallel-in-time methods (Li et al., 2016).

4. Convergence theory, contraction mechanisms, and uniform-in-time bounds

The sharpest results concern contraction in norms built on time derivatives of the coupling variable. For the Crank–Nicolson global-in-time Biot iteration,

$\Gamma\times(0,T)$9

with

$\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$0

and $\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$1 is independent of $\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$2 (Gu et al., 2024). The BDF-2 global-in-time full-order model satisfies the analogous contraction

$\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$3

with

$\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$4

(Gu et al., 6 Aug 2025).

The thermo-/multiple-network poroelasticity scheme obtains a continuous contraction

$\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$5

and a fully discrete contraction

$\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$6

with

$\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$7

(Gu et al., 31 Mar 2026). Here the transport stabilizations enter the contraction factor explicitly.

In local-in-time total-pressure splittings, the iteration is still unconditional with respect to $\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$8. For MPET, if $\{(\mathbf u^n,\xi^n,p^n)\}_{n=0}^N$9, then

$t^n$0

while convergence for $t^n$1 is proved by contradiction (Cai et al., 2023). For the three-field Biot iteration, the theorem gives

$t^n$2

with

$t^n$3

and convergence holds even for $t^n$4 (Gu et al., 2022).

In all-at-once parallel-in-time methods, the contraction constants are tied to time-integrator stability rather than PDE parameters. For block $t^n$5-circulant iterations, stable one-step methods satisfy

$t^n$6

and stable linear multistep methods satisfy

$t^n$7

with bounds independent of $t^n$8 and $t^n$9 (Wu et al., 2021). For implicit-Euler all-at-once parabolic solvers, spectral bounds for the DST preconditioner yield Uzawa convergence rates independent of the number of time steps, final time, and spatial mesh sizes (Neumuller et al., 2018).

Not all strict global-in-time algorithms have complete proofs. For the nonlinear Stokes–Darcy space–time interface method, there is no full analytic convergence theory for the nonlinear global-in-time interface algorithm; the work instead relies on well-posedness and numerical evidence (Hoang et al., 2020).

5. Parallelism, local time stepping, and model reduction

A principal motivation for strict global-in-time decoupling is parallelism in the time direction. In the global-in-time Crank–Nicolson Biot algorithm, the mechanics solves for $\Delta t$0 are independent and in parallel, while the flow solve remains a forward sweep; the paper therefore characterizes the method as having a partially parallel-in-time feature (Gu et al., 2024). The same work gives the speedup estimate

$\Delta t$1

when $\Delta t$2 processors are used for the mechanics stage (Gu et al., 2024).

Global-in-time interface methods also support asynchronous temporal resolution. In nonlinear Stokes–Darcy coupling, different time grids $\Delta t$3 and $\Delta t$4 are coupled by $\Delta t$5 projections $\Delta t$6 and $\Delta t$7, so nonmatching time grids are handled without introducing a separate mortar time grid (Hoang et al., 2020). The fluid–structure domain decomposition methods are explicitly designed for local time stepping in the subsystems (Kunwar et al., 2024).

Reduced-order acceleration has been incorporated directly into global-in-time decouplings. The POD-based BDF-2 Biot algorithm keeps the diffusion step at full order but replaces most generalized Stokes solves by solves in iteration-dependent POD spaces built from snapshot times $\Delta t$8 (Gu et al., 6 Aug 2025). Theoretical estimates show convergence to a neighborhood controlled by the ROM error,

$\Delta t$9

and the heterogeneous injection–production test reports about $\Delta t$0 minutes/iteration for the FOM, about $\Delta t$1 minutes/iteration for ROM with the full index set, and about $\Delta t$2 minute/iteration for a sparse Legendre-based snapshot set, achieving $\Delta t$3 speed-up while maintaining high accuracy (Gu et al., 6 Aug 2025).

At the linear-algebraic end, the DST-preconditioned all-at-once parabolic solver has theoretical parallel complexity $\Delta t$4 in the number of time steps, and the cost per iteration is

$\Delta t$5

(Neumuller et al., 2018).

6. Applications, limitations, and recurrent misunderstandings

The application range is broad. Total-pressure MPET decoupling has been used for brain flow simulations (Cai et al., 2023). Global-in-time Biot algorithms have been tested on manufactured solutions, Barry–Mercer, and Mandel’s problem (Gu et al., 2024). The unified thermo-/multiple-network formulation covers both linear thermo-poroelasticity and Barenblatt–Biot dual-network models (Gu et al., 31 Mar 2026). Space–time interface methods address nonlinear Stokes–Darcy coupling with nonconforming time grids (Hoang et al., 2020). Global iterative solvers have also been studied for the Kadanoff–Baym equations in nonequilibrium DMFT, where the unknowns are two-time Green’s functions rather than PDE fields (Gašperlin et al., 12 Dec 2025).

Several misconceptions recur in this literature. First, not every decoupled iterative algorithm is global-in-time in the strict sense: the MPET and three-field Biot total-pressure schemes iterate per time step, even though their stability and error estimates are uniform in time (Cai et al., 2023, Gu et al., 2022). Second, global-in-time does not automatically mean fully parallel-in-time: in Biot fixed-stress schemes, the flow sweep remains sequential and only the mechanics stage is independent across time levels (Gu et al., 2024). Third, global-in-time formulations do not automatically outperform time-stepping. For the Kadanoff–Baym equations, several stable global solvers exhibit an iteration count that scales roughly linearly with the number of time steps, and with dense matrix–matrix convolutions the total work scales as $\Delta t$6, which is asymptotically worse than standard time-stepping (Gašperlin et al., 12 Dec 2025). The same work identifies a propagating residual front as the mechanism behind that scaling, and shows that a standard forward fixed point iteration does not converge stably at long propagation times (Gašperlin et al., 12 Dec 2025).

A final limitation concerns parameter robustness. Total-pressure formulations control locking and enable stable Stokes pairs, but the iteration rate need not be parameter-robust in the strongest sense. In MPET, the contraction factor depends on $\Delta t$7, $\Delta t$8, and the storage coefficients; the paper explicitly does not claim a fully parameter-robust rate independent of those physical parameters (Cai et al., 2023).

Taken together, these developments show that the modern theory of global-in-time iterative decoupled algorithms is organized around a small set of structural ideas: reformulation by total pressure or interface variables, iteration in space–time norms, decoupling into canonical subproblems, and preconditioning or stabilization strong enough to preserve uniform-in-time control. The main unresolved questions concern nonlinear convergence theory in the strict global-in-time setting, the interaction between compression and iteration in genuinely large space–time systems, and the extent to which time-parallel efficiency can be obtained without sacrificing the robust stability properties that have made these methods technically attractive (Hoang et al., 2020, Gašperlin et al., 12 Dec 2025).