Simulations of interacting quantum particles often use a moving basis, where each single-particle wave function (an "orbital") deforms over time to follow the system. This keeps the number of many-body configurations small. It rests on two fragile numerical steps. First, every time step inverts the matrix of orbital occupations, which becomes nearly singular whenever an orbital holds almost no particles. Second, each orbital is stored on a smooth global basis such as plane waves. On such a basis, sharp walls and steps in the trapping potential cause Gibbs ringing: persistent oscillations near a sharp feature that do not shrink as the basis grows, which caps convergence at a slow power law. We propose to fix both steps with one tool. Each orbital is represented on a deliberately redundant, overcomplete set of functions, a Fourier extension built on a box larger than the trap. A single truncated singular value decomposition (TSVD) then filters the combined equation of motion. TSVD solves a linear system while discarding directions whose singular values fall below a threshold $\tau$.
The core claim is specific. In the multiconfigurational time-dependent Hartree method for bosons (MCTDHB), the linear system for the orbital time derivatives has a metric that factorizes as a Kronecker product of the occupation matrix and the overlap matrix of the redundant basis. Filtering the singular values of that product jointly, instead of regularizing each factor separately, lets a weakly occupied orbital keep evolving in its well-resolved directions while only its genuinely unstable directions are frozen. Combined with the invariance phenomenon from Fourier-extension theory, this should give hard-wall and irregular-domain many-boson dynamics an accuracy floor set by a single condition-number cap. The floor would then be independent of how sharp the boundary is.
What is known and what is missing
The case for moving bases is well established. MCTDHB contracts a $10^{20}$-configuration fixed-basis problem for 12 bosons in a triple well to about $1.3\times10^6$ configurations with 12 adaptive orbitals (Accurate multi-boson long-time dynamics in triple-well periodic traps). It also captures fragmentation and depletion that mean-field theory misses (Universality of fragmentation in the Schrödinger dynamics of bosonic Josephson junctions, Reduced density matrices and coherence of trapped interacting bosons). These studies motivate the work: fragmentation and depletion are exactly the regimes in which occupations approach zero and the occupation matrix becomes ill-conditioned.
Two prior lines constrain the design. Time-dependent coupled cluster with orthogonal adaptive basis functions already applies an SVD cutoff to near-singular modal equations in vibrational coupled cluster. Its threshold is static and empirical, and it is tested only on smooth molecular potentials. We reuse its idea of spectral truncation, but apply it to a product metric that also contains a redundant spatial basis, and we study how the threshold affects conserved quantities. Ab-initio variational wave functions for the time-dependent many-electron Schrödinger equation takes the opposite route: its Taylor-root propagator avoids inverting the ill-conditioned metric altogether. This matters because it is evidence that direct inversion can be bypassed rather than regularized. We treat that as a hypothesis to test, not to dismiss.
A third line supplies the method. Parameter selection and numerical approximation properties of Fourier extensions from fixed data shows three things about approximating non-periodic functions with a periodic basis on an enlarged box $[-T,T]$ solved by TSVD. The boundary Gibbs effect disappears. Once the condition-number cap $\kappa^*$ is fixed, accuracy is essentially independent of $T$. Accepting 3–6 digits of accuracy lets far more modes be resolved stably. Stable and Accurate Least Squares Radial Basis Function Approximations on Bounded Domains and Multivariate rational approximation of functions with curves of singularities make the same point for other bases: a redundant basis plus TSVD is expressive rather than dangerous. Frames and Numerical Approximation gives the general theory. All of this work is static, one-shot fitting. None of it has been placed inside a variational time-evolution equation, where truncation errors feed back into later steps.
We deliberately set aside the moving-shock line. Nonlinear Reconstruction for Operator Learning of PDEs with Discontinuities proves an $O(1/p)$ error lower bound for fixed linear bases on shocks, and fixes it with learned basis shifts. Hybrid Surrogate Models reaches an $l_1$ error of $7.33\times10^{-5}$ on a propagating shock by adding explicit Heaviside jumps. Both target discontinuities that move. In a trap with fixed hard walls, the non-smooth feature is stationary, and the wave function itself does not jump: it has a kink where it meets an infinite wall and a jump in its second derivative at a finite step. The Gibbs problem is therefore the slow convergence and derivative ringing that come from forcing these features into a periodic basis. It is not a 9% overshoot in the density. Stating this honestly changes the design: we need a domain-adapted redundant basis, not a basis that translates with the feature.
Core mechanism
Write each orbital as $\phi_k(x,t)=\sum_j c_{kj}(t)\,\psi_j(x)$. Here the $\psi_j$ are Fourier modes on the enlarged box, restricted to the physical trap $\Omega$, and $G_{ij}=\langle\psi_i,\psi_j\rangle_\Omega$ is their overlap matrix on $\Omega$. $G$ is severely ill-conditioned because the modes are nearly dependent on $\Omega$. Projecting the MCTDHB orbital equations onto this frame gives a linear system for the coefficient derivatives $\dot c$:
$$ (\rho\otimes G)\,\dot c = r(c,C), $$
where $\rho$ is the one-body reduced density matrix, $C$ are the configuration coefficients, and $r$ is the usual projected right-hand side. The eigenvalues of $\rho\otimes G$ are the products $n_k g_j$ of the natural occupations $n_k$ and the frame overlap eigenvalues $g_j$. The proposed filter keeps exactly the directions with
$$ n_k\, g_j \ge \tau , $$
and solves by pseudo-inverse on that subspace.
This joint filter is the key ingredient. Standard practice regularizes $\rho$ with a small additive shift and treats the spatial basis as well-conditioned. The joint filter instead yields a per-orbital, per-direction decision. With $\tau$ fixed, the effective spatial resolution of a nearly empty orbital shrinks smoothly as its occupation falls, and it recovers when the orbital refills. The analysis questions are twofold. First, whether truncation onto a subspace that is orthogonal in the $\rho\otimes G$ inner product preserves norm and energy. We expect it to when the retained subspace is closed under multiplication by $i$; we will prove or refute this. Second, whether the Fourier-extension invariance in $T$ survives time propagation.
Research Program
We will implement the frame representation and joint filter inside a standard MCTDHB propagator. The potential will be evaluated at oversampled collocation points, with roughly twice as many points as modes, following Adcock and Ruan's finding that oversampling beyond about 2 gives diminishing returns. This avoids ever projecting a step function onto the basis. We will then build three regimes of increasing difficulty.
The first is a one-dimensional hard-wall double and triple well, analogous to the periodic triple well of Streltsov et al. but with the periodic boundary replaced by infinite walls and the inner barriers made rectangular. The second is interacting bosons tunneling through a sharp rectangular barrier into open space, following How an interacting many-body system tunnels through a potential barrier to open space. Here the enlarged box doubles as an absorbing buffer. The third is a two-dimensional irregular hard-wall trap, such as a disk or a stadium, which is the regime where boundary-adapted sine bases do not exist.
In each regime we will sweep $\tau$, $T$, the number of orbitals $M$, and the number of modes per orbital. We will measure four things: convergence of natural occupations and density against a reference, norm and energy drift over long times, cost per step, and whether the invariance in $T$ persists. The theoretical deliverable is either a conservation statement for the filtered dynamics or a counterexample, together with a characterization of how the error floor depends on $\tau$.
Baselines
The naive reference point is current common practice. Orbitals live on a uniform periodic grid of 255 points, as in the triple-well study, with fast Fourier transform kinetic energy. The hard wall is emulated by a tall finite potential step. The occupation matrix is regularized by a small fixed additive shift. We will run it at the same total number of spatial degrees of freedom per orbital as the proposal, and additionally at 4× and 16× that grid size. Beating it means reaching the same occupation and density error with fewer spatial degrees of freedom, or lower error at equal cost. The wall height will be swept so the baseline is not handicapped by a bad choice.
The strongest competitor in one dimension is a boundary-adapted sine discrete variable representation, which enforces the hard-wall condition exactly. It removes the boundary kink problem altogether, but not the interior-step problem. In two-dimensional irregular domains, the strongest competitor is a fine masked grid or a finite-element discrete variable representation. No reported result was found on Lacuna for either method's accuracy on many-boson hard-wall dynamics, so we will establish these numbers ourselves with equal wall-clock budgets and equal tuning effort over resolution.
For reference behavior rather than error values, the file reports that for the smooth triple well an $M=3$ model is "essentially exact" for the density up to about 24 on-site oscillations, after which more orbitals are required. Every method must reproduce this qualitative crossover in its hard-wall analogue.
Two ablations isolate the contribution. One replaces the joint filter with separate filters on $\rho$ and on $G$, in the spirit of Højlund et al.'s static cutoff. The other uses the joint filter with a sine basis instead of the redundant Fourier extension. Together they show whether the gain comes from redundancy, from joint filtering, or from both.
References are exact diagonalization for $N=2$–4 on very fine grids, and heavily converged grid MCTDHB for larger $N$.
Why not just…?
Why not just refine the grid? In one dimension, grid points are cheap and the cost is dominated by configurations, so this is a real threat there. We expect the proposal to win in one dimension only on interior-step convergence and derivative quantities such as kinetic energy density. The decisive regime is two dimensions, where grid refinement multiplies orbital cost quadratically. If refined grids match the proposal in two dimensions at equal cost, the claim fails.
Why not just use a sine basis or finite elements that respect the wall exactly? For rectangular one-dimensional boxes, this probably matches or beats us, and we say so. The proposal's niche is irregular domains and interior sharp barriers, where Fourier extensions handle geometry through a bounding box without meshing. The finite-element comparison in the disk and stadium traps settles this directly.
Why not just add the usual small shift to the occupation matrix, or bypass inversion as Nys et al. do? The shift is a smooth filter that damps every direction of a weakly occupied orbital, where the joint filter damps only the unstable ones. The ablation measures whether this distinction changes natural-occupation errors during fragmentation. We will also implement a Taylor-root-style propagator on the frame representation. If it matches the joint filter's accuracy and stability, then the value lies in the redundant basis, not in TSVD, and we will report that.
Decisive checks
The proposal succeeds if three things hold. In two-dimensional irregular traps, it reaches a given error in natural occupations at a lower cost than the finite-element or masked-grid baseline. Its long-time energy drift stays bounded at a level set by $\tau$. And varying $T$ at fixed $\kappa^*$ leaves the dynamical accuracy essentially unchanged, reproducing the static invariance phenomenon in time. It fails if either the separate-filter ablation or the inversion-free propagator matches it, which would show that the joint filter is not the reason for any gain.
References
- Accurate multi-boson long-time dynamics in triple-well periodic traps
- Universality of fragmentation in the Schrödinger dynamics of bosonic Josephson junctions
- Reduced density matrices and coherence of trapped interacting bosons
- Time-dependent coupled cluster with orthogonal adaptive basis functions: General formalism and application to the vibrational problem
- Ab-initio variational wave functions for the time-dependent many-electron Schrödinger equation
- Parameter selection and numerical approximation properties of Fourier extensions from fixed data
- Stable and accurate least squares radial basis function approximations on bounded domains
- Multivariate Rational Approximation of Functions with Curves of Singularities
- Frames and numerical approximation
- Nonlinear Reconstruction for Operator Learning of PDEs with Discontinuities
- Hybrid Surrogate Models: Circumventing Gibbs Phenomenon for Partial Differential Equations with Finite Shock-Type Discontinuities
- How an interacting many-body system tunnels through a potential barrier to open space