Projector-Splitting Time Evolution

• Projector-splitting time evolution is a geometric integrator that decomposes the tangent-space projection into sequential subflows, avoiding matrix inversions.
• The method demonstrates exactness for matrices with true fixed rank and remains robust under over-approximation, facilitating adaptive rank updates.
• It is widely applied in dynamical low-rank approximations, tensor networks, kinetic PDEs, and machine learning to improve stability and computational efficiency.

Searching arXiv for recent and foundational papers on projector-splitting time evolution and closely related dynamical low-rank / TDVP methods. Projector-splitting time evolution denotes a class of geometric time-integration procedures in which an evolution constrained by a projector is decomposed into simpler subflows that can be solved successively. In the dynamical low-rank setting, the projector is the orthogonal projector onto the tangent space of the manifold of fixed-rank matrices, and the resulting scheme propagates a low-rank approximation directly, rather than evolving a full matrix and recomputing a truncated factorization at every step. The formulation introduced for matrices by Lubich and Oseledets is explicit, computationally inexpensive, and notable for robustness when the core factor becomes ill-conditioned or nearly singular (Lubich et al., 2013). Across later literature, closely related constructions appear in tensor-network dynamics, multiconfiguration quantum propagation, kinetic equations, spatial projection methods for quasiperiodic PDEs, and low-rank optimization or training, although the phrase is not used in exactly the same sense in every subfield.

1. Variational and geometric foundation

The basic object in dynamical low-rank approximation is a time-dependent matrix $A(t)\in\mathbb{R}^{m\times n}$, or the solution of a matrix ODE $\dot A = F(A)$. One seeks an approximation $Y(t)$ of fixed rank $r\ll \min(m,n)$, constrained to the manifold

$$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$

The defining variational principle requires $\dot Y(t)$ to be the best tangent approximation to $\dot A(t)$, or to $F(Y(t))$ in the ODE setting. This yields the projected evolution

$\dot Y(t)=P(Y(t))\,F(Y(t)),$

where $P(Y)$ is the orthogonal projector onto the tangent space $\dot A = F(A)$0 (Lubich et al., 2013).

Writing

$\dot A = F(A)$1

with $\dot A = F(A)$2 and $\dot A = F(A)$3 column-orthonormal and $\dot A = F(A)$4 invertible, together with gauge conditions

$\dot A = F(A)$5

gives the standard factor equations. Direct integration of those equations becomes problematic when $\dot A = F(A)$6 is ill-conditioned or nearly singular, which corresponds to over-approximating the actual rank. The projector formulation avoids placing $\dot A = F(A)$7 at the center of the numerical method (Lubich et al., 2013).

A key identity is the explicit tangent-space projector

$\dot A = F(A)$8

or equivalently

$\dot A = F(A)$9

This splittable structure is the starting point of projector-splitting time evolution: the projected vector field is decomposed into three simpler pieces associated with row-space projection, overlap correction, and column-space projection (Lubich et al., 2013).

2. Construction of the matrix projector-splitting integrator

The first-order integrator is a Lie–Trotter splitting of the projected evolution into three subproblems corresponding to the three terms of the tangent-space projector. In factored form, these become linear updates for $Y(t)$0, then for $Y(t)$1, then for $Y(t)$2. For known matrix increments $Y(t)$3, the practical KSL ordering is

$Y(t)$4

factorized as $Y(t)$5, followed by the core correction

$Y(t)$6

and then

$Y(t)$7

factorized as $Y(t)$8, after which

$Y(t)$9

This algorithm never uses $r\ll \min(m,n)$0 or $r\ll \min(m,n)$1; it uses matrix additions and QR or SVD on thin matrices only (Lubich et al., 2013).

The same splitting admits a symmetric second-order version by composing the first-order step with its adjoint at half step size. For known $r\ll \min(m,n)$2, Lubich and Oseledets give an explicit formula involving $r\ll \min(m,n)$3, $r\ll \min(m,n)$4, and $r\ll \min(m,n)$5; the method is second-order accurate for the projected evolution and still requires no matrix inversions (Lubich et al., 2013).

The construction extends directly to matrix ODEs $r\ll \min(m,n)$6. In that case, $r\ll \min(m,n)$7 is replaced by a numerical approximation based on $r\ll \min(m,n)$8: first order by $r\ll \min(m,n)$9, and second order by a predictor–corrector using a linear approximation to $$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$0 inside the symmetric splitting. This keeps the scheme fully explicit while preserving the geometric structure of the fixed-rank manifold (Lubich et al., 2013).

3. Exactness, robustness, and adaptive-rank consequences

A central exactness result states that if $$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$1 has rank at most $$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$2 for all $$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$3, then with $$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$4 the KSL splitting step is exact, $$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$5. This property depends on the KSL ordering; other orderings such as KLS generally lose it (Lubich et al., 2013).

The more distinctive result concerns over-approximation. If the true matrix is $$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$6 of rank $$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$7, perturbed by a small full-rank term $$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$8, then the rank-$$\mathcal{M}_r=\{Y\in\mathbb{R}^{m\times n}:\operatorname{rank}(Y)=r\}.$$9 and rank-$\dot Y(t)$0 projector-splitting trajectories satisfy

$\dot Y(t)$1

uniformly in time, with $\dot Y(t)$2 independent of $\dot Y(t)$3, $\dot Y(t)$4, and $\dot Y(t)$5. In other words, the excess-rank components remain small even when the core $\dot Y(t)$6 becomes very ill-conditioned (Lubich et al., 2013).

This is not the behavior of standard integrators applied directly to factor equations containing $\dot Y(t)$7 and $\dot Y(t)$8. In the numerical experiments of Lubich and Oseledets, the implicit midpoint rule becomes unstable or fails in over-approximation regimes, and a KLS variant also loses robustness, whereas KSL and its symmetrized version maintain stable, accurate results and the expected order. For example, with $\dot Y(t)$9 and over-approximation $\dot A(t)$0, midpoint fails, KLS-based methods lose nominal behavior, and KSL-based methods remain stable (Lubich et al., 2013).

Two algorithmic consequences follow directly. First, robustness under over-approximation makes adaptive rank increase feasible, because the new degrees of freedom can be introduced through very small singular directions without ever evaluating $\dot A(t)$1. Second, one projector-splitting step along the linear path $\dot A(t)$2 can serve as an efficient truncation or retraction back to the rank-$\dot A(t)$3 manifold in low-rank optimization methods (Lubich et al., 2013).

4. Tensor-network and multiconfiguration formulations

In multiconfiguration time-dependent Hartree, the tangent-space projector of the MCTDH manifold can be written as

$\dot A(t)$4

and, after orthogonalizing single-hole functions and introducing non-orthogonal single-particle functions, the SPF part becomes

$\dot A(t)$5

The resulting projector-splitting algorithm yields linear equations for the transformed SPFs and avoids direct inversion of the single-particle density matrix. In this setting, projector splitting refers not to a Hamiltonian splitting but to a splitting of the tangent-space projector of the Tucker/MCTDH manifold (Bonfanti et al., 2018).

For matrix product states, Haegeman and collaborators formulated TDVP time evolution by splitting the projector onto the MPS tangent space rather than using a Suzuki–Trotter splitting of the Hamiltonian. The resulting algorithm resembles finite-system DMRG so closely that it can be implemented by changing only a few lines of code, is compatible with arbitrary Hamiltonians including long-range interactions, and yields DMRG as a special case of imaginary time evolution with infinite time step (Haegeman et al., 2014).

The same geometric idea persists at deeper tensor-network levels. In ML-MCTDH, a projector-splitting integrator based on a gauge in which both SPFs and SHFs are orthonormal removes the singular reduced-density inverses responsible for weak-entanglement instabilities. On large spin-boson benchmarks, this implementation was reported to require roughly $\dot A(t)$6–$\dot A(t)$7 orders of magnitude fewer Hamiltonian evaluations and $\dot A(t)$8–$\dot A(t)$9 orders of magnitude fewer Hamiltonian applications than standard ML-MCTDH, with calculations including up to $F(Y(t))$0 bath modes and wavefunctions containing up to $F(Y(t))$1 parameters (Lindoy et al., 2021).

In two-dimensional tensor networks, an iPEPS time-step algorithm from first principles applies a short-time evolution that enlarges the bond dimension and then variationally projects back to the fixed-$F(Y(t))$2 iPEPS manifold by maximizing overlap using CTMRG environments. This is a projector step in a variational tensor-network manifold rather than a tangent-space KSL step, but it is explicitly cast as a projector-based time-evolution strategy and provides proof of principle for real-time and imaginary-time evolution on an infinite 2D lattice (Czarnik et al., 2018).

A TD-DMRG benchmark on the Fenna–Matthews–Olson complex found that TDVP-MU and TDVP-PS yield the same result when the time step size is converged, that both are more accurate than P&C-RK4, and that TDVP-PS tolerates a larger time step size than TDVP-MU; GPU acceleration of the heavy tensor contractions sped up TDVP-MU and TDVP-PS by up to $F(Y(t))$3 times (Li et al., 2019).

5. PDEs, kinetic equations, and stability theory

For the Vlasov–Poisson equation, a low-rank projector-splitting integrator constrains the dynamics to a manifold

$F(Y(t))$4

and splits the tangent-space projector into three subprojections. In four and six phase-space dimensions, one time step is reduced to two systems of two- or three-dimensional advection equations and a small matrix ODE; with hierarchical dynamical low-rank approximation, the same step is further reduced to a set of four or six systems of one-dimensional advection equations whose size is still equal to the rank (Einkemmer et al., 2018).

A different but related use of the phrase appears in the nonlinear quasiperiodic Schrödinger equation, where the method combines a spatial projection onto a finite quasiperiodic spectral space with Strang splitting in time. The fully discrete update has the form

$F(Y(t))$5

and the main error estimate is

$F(Y(t))$6

which gives second-order accuracy in time and spectral accuracy in space under the stated regularity assumptions (Jiang et al., 2024).

For linear hyperbolic and parabolic equations, recent stability analysis of the low-rank projector-splitting integrator distinguishes discretize-then-project and project-then-discretize formulations. For hyperbolic equations, the full upwind–Euler scheme is $F(Y(t))$7-stable if $F(Y(t))$8, whereas Lie–Trotter PSI with Euler gives $F(Y(t))$9 for both DtP and PtD; Strang splitting with SSP-RK2 significantly enlarges the stability region, to approximately $\dot Y(t)=P(Y(t))\,F(Y(t)),$0 for DtP and $\dot Y(t)=P(Y(t))\,F(Y(t)),$1 for PtD. For parabolic equations, despite the negative S-step, unconditional stability is recovered with Crank–Nicolson in all substeps or with a hybrid scheme using backward Euler in K and L and forward Euler in S (Zhang et al., 21 Jul 2025).

These PDE examples show two distinct roles for projection. In kinetic low-rank dynamics, the projector is the tangent-space projector of a low-rank manifold of functions. In quasiperiodic Schrödinger discretization, the projector $\dot Y(t)=P(Y(t))\,F(Y(t)),$2 is a spatial spectral projector intertwined with time splitting. The formal similarity is real, but the underlying geometric object is different.

6. Optimization, machine learning, and broader meanings of the term

Projector-splitting ideas have moved from approximation of prescribed dynamics to optimization and training. In the matrix setting, one splitting step along $\dot Y(t)=P(Y(t))\,F(Y(t)),$3 already acts as an efficient retraction onto the fixed-rank manifold, which is why the original paper explicitly points to optimization algorithms for low-rank matrices as a second application area besides time evolution (Lubich et al., 2013).

In dynamical low-rank training of neural networks, the projected gradient flow

$\dot Y(t)=P(Y(t))\,F(Y(t)),$4

inherits the same geometric stiffness issues as matrix ODEs. A recent augmented backward-corrected projector-splitting integrator introduces

$\dot Y(t)=P(Y(t))\,F(Y(t)),$5

so that $\dot Y(t)=P(Y(t))\,F(Y(t)),$6, removes the descent obstruction of backward-corrected PSI, reduces the number of QR factorizations per step from two to one, and proves both a robust global error bound

$\dot Y(t)=P(Y(t))\,F(Y(t)),$7

and a loss decrease inequality for sufficiently small $\dot Y(t)=P(Y(t))\,F(Y(t)),$8 (Kusch et al., 5 Feb 2025).

The expression also appears in open quantum systems in a more abstract projector-operator sense. There, a time-dependent superoperator projector $\dot Y(t)=P(Y(t))\,F(Y(t)),$9 splits relevant and irrelevant sectors of Liouville space, and generalized TCL theory produces time-local master equations for $P(Y)$0. With generalized Kawasaki–Gunton projectors, this leads to nonlinear master equations on ansatz manifolds determined by relevant observables. In that setting, “projector-splitting time evolution” refers to following a moving relevant subspace in operator space rather than splitting the tangent projector of a fixed-rank matrix manifold (Meretukov et al., 2023).

A persistent misconception is therefore that projector-splitting time evolution always means splitting the Hamiltonian. In the foundational low-rank, MPS, and MCTDH literature, what is split is the projector onto a tangent space, not the Hamiltonian itself (Lubich et al., 2013). In other literatures, the phrase can denote operator splitting combined with a spatial or statistical projector. The common core is geometric restriction of dynamics by projection; the specific projector, manifold, and meaning of the split depend on the application domain.