Pseudo-Time Continuation

• Pseudo-time continuation is a family of techniques that uses an auxiliary time-like parameter to navigate solution branches and bypass singularities in nonlinear systems.
• It is applied across various fields, including numerical continuation, steady-state analysis of PDEs, time-series forecasting, and even quantum and holographic formulations.
• Key implementations involve pseudo-arclength continuation for traversing folds, pseudo-transient methods for reaching equilibrium, and analytic continuation in complex time setups.

Pseudo-time continuation denotes a family of constructions in which progression is indexed by an auxiliary time-like parameter rather than by physical time or by a directly imposed control variable. In one canonical usage, a fictitious parameter $s$ is introduced to follow a solution branch of an underdetermined nonlinear system through folds and turning points. In another, a steady or differential-algebraic problem is embedded in an artificial evolution whose long-time limit is the desired equilibrium. Other literatures use the same language for conditional continuation of a time series, reparametrisation of time in quantum mechanics, analytic continuation in complex time, or evolution of inverse-problem optima along a parameter path (Aruliah et al., 2013, Wang et al., 2014, Sak et al., 2012, Kapoor, 2016, Hart et al., 28 Aug 2025). This suggests a common structure: direct evolution is replaced by an auxiliary parameterization designed to restore tractability, robustness, or access to otherwise hidden solution structure.

1. Conceptual scope and recurring structure

Across the literature, pseudo-time continuation is not a single algorithm but a family of closely related strategies. The shared mechanism is the replacement of an inconvenient parameterization by an auxiliary one. In numerical continuation for nonlinear equations, the original parameter $\lambda$ is replaced by an arclength-like variable $s$. In pseudo-transient schemes, the steady problem $F(x)=0$ is replaced by an artificial evolution. In stochastic time-series analysis, a finite observed record is extended by conditional simulation of future values. In quantum and holographic settings, physical time itself is reparametrized or analytically continued (Aruliah et al., 2013, Wang et al., 2014, Sak et al., 2012, Kapoor, 2016).

Setting	Auxiliary variable	Purpose
Underdetermined nonlinear systems	$s$	Traverse folds and turning points
Steady DAEs/PDEs and nonlinear solves	pseudo-time step $\delta_n$, $\Delta \tau$, $t$	Regularized march to a steady state
Time-series continuation	future index $t+1,\dots,t+m$ under a fitted model	Conditional simulation of continuations
Quantum and holographic formulations	$\sigma$, complexified time or radial contour	Reparametrisation or analytic continuation

In the continuation literature, the basic object is often a one-dimensional solution set of

$\lambda$0

or, equivalently, $\lambda$1 with $\lambda$2. In pseudo-transient formulations for DAEs, one instead starts from

$\lambda$3

and uses artificial time stepping to approach an equilibrium (Aruliah et al., 2013, Wang et al., 2014). In the SARIMA setting, “simulation of the continuation of a given time series” means generating

$\lambda$4

conditional on $\lambda$5, rather than returning only the conditional mean path (Sak et al., 2012).

2. Pseudo-arclength continuation of solution branches

Natural parameter continuation advances $\lambda$6 directly and solves $\lambda$7. It fails at folds, where the curve is no longer monotone in $\lambda$8 and the Jacobian $\lambda$9 becomes singular. Pseudo-arclength continuation replaces $s$0 by a pseudo-time or arclength-like variable $s$1, seeks

$s$2

and advances by tangent prediction plus correction in an augmented system (Aruliah et al., 2013).

At a known point $s$3, with unit tangent $s$4, the predictor is

$s$5

The corrector solves

$s$6

or, in compact form,

$s$7

This formulation preserves well-posedness at folds because the hyperplane constraint replaces direct stepping in $s$8 (Aruliah et al., 2013).

The parallel adaptive method PAMPAC generalizes this by exploring several predictor-corrector sequences concurrently. It launches multiple trial step lengths, reuses unconverged corrector iterates as seeds for new predictions, and organizes the resulting computation as a rooted tree with node colors $s$9. The method prunes paths according to pseudo-arclength-per-iteration efficiency and reports speed-up by a factor of two on a multi-core CPU and a factor of three on a small cluster for the high-dimensional examples studied (Aruliah et al., 2013).

A particularly explicit realization appears in amorphous carbon under shear. There the unknown pair is $F(x)=0$0, where $F(x)=0$1 are atomic positions and $F(x)=0$2 is applied shear strain, constrained by

$F(x)=0$3

The continuation variable is the pseudo-arclength $F(x)=0$4, so the solution path is $F(x)=0$5. In a 4096-atom amorphous carbon sample modeled with a machine-learned ACE interatomic potential, Hessian-free pseudo-arclength continuation resolves avalanche structure by following both minima and index-1 saddles through bifurcations. This yields a chain

$F(x)=0$6

and an event-driven description of stress drops that eliminates time-step effects of athermal quasi-static strain stepping (Birks et al., 30 Jan 2026).

3. Pseudo-transient continuation for steady states and nonlinear PDE systems

Pseudo-transient continuation treats a steady problem as the asymptotic limit of an artificial evolution. In the DAE formulation used for power-system long-term stability,

$F(x)=0$7

with pseudo-time step $F(x)=0$8 updated by Switched Evolution Relaxation: $F(x)=0$9 This iteration is exactly one Newton correction of a backward Euler step started from the previous time point, so it is a middle ground between integration and steady state calculation (Wang et al., 2014). In long-term power-system stability analysis, the method is modified to switch back to implicit integration when discrete variables jump and algebraic consistency is lost. On the 145-bus example, implicit trapezoidal simulation of the full long-term model took 122.39 s, whereas $s$0 starting at $s$1 took 16.12 s; in the QSS model, $s$2 overcame numerical difficulty at 40 s and took 21.75 s (Wang et al., 2014).

For semilinear elliptic PDEs, the same principle is formulated in a Hilbert space. With $s$3, the pseudo-evolution is

$s$4

and backward Euler leads to the PTC step

$s$5

For the singularly perturbed semilinear elliptic problem

$s$6

this is coupled with adaptive finite element Galerkin discretization and robust a posteriori residual estimators, producing a fully adaptive PTC-Galerkin scheme that decides between another PTC step and mesh refinement based on the relative size of linearization and discretization residuals (Amrein et al., 2016).

For the fully nonlinear elliptic Monge–Ampère equation

$s$7

two related schemes are analyzed. The pseudo-transient continuation method uses

$s$8

with $s$9 or $\delta_n$0. The pure pseudo time marching method uses

$\delta_n$1

which requires only Poisson solves per step. Both are proven to converge to a strictly convex solution of the natural discrete variational formulation with $\delta_n$2 conforming approximations, assuming existence of a strictly convex discrete solution (Awanou, 2013).

A related stabilization idea appears in CFD Newton–Krylov continuation. Residual smoothing modifies the right-hand side of the pseudo-transient Newton system while leaving the Newton Jacobian unchanged: $\delta_n$3 For small pseudo-time steps, this reduces to a local nonlinear smoother; for large pseudo-time steps, it recovers exact Newton and its quadratic convergence properties (Mavriplis, 2018).

In steady incompressible Navier–Stokes CFD, pseudo-transient continuation is also combined with learned local time-step prediction. The element-wise iteration is

$\delta_n$4

where the local pseudo-time step on each element is predicted by a neural network from patchwise geometric, flow, and residual information. On back-step and Couette benchmarks, the learned local pseudo-time stepping improved convergence relative to the CFL-based controllers used in COMSOL (Zandbergen et al., 2023).

4. Conditional continuation, optimization flows, inverse problems, and learning

In time-series analysis, pseudo-time continuation takes a probabilistic form. Given observed data $\delta_n$5, a fitted SARIMA model defines the conditional distribution of

$\delta_n$6

Forecasting via predict() returns the conditional mean $\delta_n$7 and conditional variance, whereas conditional simulation generates full sample paths from the joint conditional distribution. The paper’s arima.condsim function implements this recursively on the differenced series, simulates innovations $\delta_n$8, and then undifferences to the original scale. For the monthly international airline passengers example, Monte Carlo means from 10,000 continuations were extremely close to the predict() forecast path, validating the continuation distribution (Sak et al., 2012).

In unconstrained optimization, pseudo-transient continuation is built from the generalized gradient flow

$\delta_n$9

Linearizing an implicit Euler step leads to the explicit pseudo-transient update

$\Delta \tau$0

The pseudo-time step $\Delta \tau$1 is updated with a trust-region-style ratio

$\Delta \tau$2

and switching preconditioning chooses an L-BFGS inverse in well-conditioned phases and a Hessian inverse in ill-conditioned phases. On 47 standard test problems, the method solved all 47 problems, whereas the trust-region fminunc failed on 6 and the quasi-Newton fminunc failed on 13; for large-scale problems ($\Delta \tau$3), the computational time was about 1% of trust-region fminunc and about $\Delta \tau$4 of quasi-Newton fminunc (Luo et al., 2020).

For parameterized inverse problems, pseudo-time continuation follows the optimizer itself as auxiliary PDE parameters vary. With

$\Delta \tau$5

the optimality condition $\Delta \tau$6 yields, by the Implicit Function Theorem,

$\Delta \tau$7

for a path $\Delta \tau$8. Predictor–corrector integration of this ODE is accelerated by an adaptive quasi-Newton Hessian preconditioner combining parametric BFGS and block secant updates from PCG history (Hart et al., 28 Aug 2025).

A different notion appears in Monte Carlo sampling. For a single-minimum bound state sampled optimally with Gaussian proposals of scale $\Delta \tau$9, the authors argue that one can associate a pseudo-time with the configuration sequence. In the small-step diffusive regime,

$t$0

Optimal sampling occurs at

$t$1

But this pseudo-time is lost when transport among distinct bound states takes place, because symmetry is then restored and no unique time-like ordering survives (He et al., 2023).

In PINNs, pseudo-time stepping modifies the empirical PDE residual. Instead of minimizing only $t$2, one minimizes

$t$3

at collocation points. The paper argues that the main benefit is not simply easing optimization; when combined with collocation-point resampling, pseudo-time stepping reveals and avoids spurious solutions hidden by the empirical residual loss. The adaptive step size is chosen from a finite-difference surrogate

$t$4

which approximates inverse local residual-Jacobian magnitude and aims for the largest stable step without per-problem tuning (Wang et al., 26 Apr 2026).

5. Complex-time formulations in quantum theory and holography

In quantum mechanics, pseudotime is a literal reparametrisation of time. With a local time scaling function $t$5,

$t$6

and pseudo-Hamiltonian

$t$7

classical dynamics in the homogeneous formalism becomes equivalent to the original dynamics constrained by $t$8. In the Hamiltonian path integral, reparametrisation leads to the Duru–Kleinert formula, and in the operator formalism it induces modified derivatives

$t$9

This structure generates supersymmetric partner Hamiltonians, shape invariance, and isospectral deformation from the choice of $t+1,\dots,t+m$0 (Kapoor, 2016).

In holography, pseudo-time continuation takes the form of analytic continuation into time-like or complex-time directions. In dS/CFT, pseudo entropy is defined from the reduced transition matrix

$t+1,\dots,t+m$1

with pseudo entropy

$t+1,\dots,t+m$2

Because $t+1,\dots,t+m$3 need not be Hermitian, the pseudo entropy is generally complex. The paper relates holographic pseudo entropy in de Sitter to time-like entanglement entropy in CFTs by analytic continuation, and argues that the imaginary part of pseudo entropy implies an emergence of time in dS/CFT (Doi et al., 2022).

A more geometric formulation appears in no-boundary de Sitter. Extremal surface areas are defined by contour integrals in the complex time plane, with the radial/time coordinate $t+1,\dots,t+m$4 allowed to be complex and the contour $t+1,\dots,t+m$5 deformable as long as no singularities are crossed. For $t+1,\dots,t+m$6, both timelike+Euclidean extremal surfaces and purely complex extremal surfaces can exist, and multiple extremal surfaces can have indistinguishable areas because their contours are deformable into each other in the complex $t+1,\dots,t+m$7-plane. Here pseudo-time continuation means analytic continuation of the time coordinate into the complex plane together with contour deformation (Narayan, 31 Mar 2026).

6. Limits, failure modes, and interpretive cautions

A recurring misconception is to identify pseudo-time with physical time. In pseudo-arclength continuation, $t+1,\dots,t+m$8 orders points on a branch but does not represent dynamics. In the amorphous-carbon avalanche study, the authors explicitly note that the ordering inferred from the latent structure is correlated with but not always identical to the ordering obtained from overdamped dynamics; the continuation trajectory is a geometric path in configuration–strain space (Birks et al., 30 Jan 2026). In power-system $t+1,\dots,t+m$9, pseudo-time steps are chosen for convergence efficiency rather than time accuracy, even though the method inherits stability properties of backward Euler (Wang et al., 2014). In quantum mechanics and de Sitter holography, the auxiliary time variable may be non-unitary or complex, so any literal dynamical interpretation is even less appropriate (Kapoor, 2016, Narayan, 31 Mar 2026).

A second caution concerns residuals and apparent convergence. In SARIMA continuation, the procedure treats estimated parameters as known, assumes Gaussian innovations, and is only as reliable as the fitted model; model misspecification, non-normality, and unpropagated parameter uncertainty can make simulated continuations misleading (Sak et al., 2012). In PINNs, the issue is sharper: small empirical PDE residual loss can coexist with physically incorrect solutions, because the empirical residual may admit trivial or spurious solutions. Pseudo-time stepping helps only when paired with collocation-point resampling and a stable choice of $\sigma$0 (Wang et al., 26 Apr 2026).

A third limitation is domain-specific. PAMPAC assumes that corrector steps are much more expensive than communication, and it is designed for underdetermined systems with one free parameter; its benefit diminishes when correctors are cheap or the problem structure changes (Aruliah et al., 2013). In Monte Carlo sampling, pseudo-time can be calibrated only in the optimal regime for a single-minimum bound state; once transport among distinct bound states takes place, pseudo-time can no longer be defined (He et al., 2023). In power systems, the theory is built on semi-explicit index-1 DAEs and consistent initial conditions, so discrete events require temporary return to conventional implicit integration (Wang et al., 2014).

A plausible unifying implication is that pseudo-time continuation is best understood not as a surrogate for real time but as a controlled parameterization of progression. Depending on the field, that progression may traverse a nonlinear solution manifold, relax toward a steady state, extend a stochastic process beyond observed data, follow an optimizer through parameter space, or continue amplitudes and contours into complex time. What remains invariant across these uses is the role of the auxiliary parameter as an organizing device for otherwise ill-conditioned, singular, or inaccessible structure.