Legendre Time Reduction Overview

• Legendre Time Reduction is a method that uses weighted Legendre expansions to reduce high-dimensional, time-dependent inverse problems to coupled spatial systems.
• It isolates structured temporal information to mitigate noise and computational instability, thereby improving the efficiency of inverse PDE reconstructions.
• The approach also extends to sparse spectral recovery, quantum state preparation, and memory compression, highlighting its versatility across various applications.

Searching arXiv for papers on Legendre time reduction and closely related formulations. Legendre Time Reduction denotes a family of Legendre-based reduction strategies that replace a large time-dependent, high-dimensional, or noise-sensitive problem by a smaller coefficient system in Legendre coordinates. The expression is used most explicitly for inverse PDEs, where an exponentially weighted Legendre basis in time converts a time-dependent inverse problem into coupled time-independent elliptic systems for coefficient fields (Van et al., 22 Jul 2025, Dang et al., 15 Jun 2025). In adjacent literatures, closely related language is used for sparse Legendre spectral recovery, quantum Legendre-Fenchel state preparation, and learned memory compression, all of which reduce computational burden by retaining only structured Legendre information rather than the full original representation (Hu et al., 2015, Sutter et al., 2020, Zhou et al., 2022).

1. Terminological scope

The cited literature does not use the phrase in a single uniform sense. In inverse problems, it refers to an explicit time-dimensional reduction. In sparse approximation and quantum convex conjugation, it refers to a complexity reduction obtained through Legendre-related transforms. In machine learning, it often denotes temporal summarization through fixed-size Legendre memory states rather than elimination of the time variable itself (Wu et al., 16 Dec 2025, Fong et al., 22 Dec 2025).

Usage in literature	Mechanism	Representative papers
Time-dimensional reduction	Expand the time dependence in a weighted Legendre basis and solve for spatial coefficient fields	(Van et al., 22 Jul 2025, Dang et al., 15 Jun 2025)
Sparse spectral reduction	Recover only significant Legendre coefficients instead of the full spectrum	(Hu et al., 2015)
Quantum transform speedup	Encode the discrete Legendre-Fenchel transform in a quantum state	(Sutter et al., 2020)
Temporal memory compression	Store history in fixed-dimensional Legendre states or Legendre graphs	(Zhou et al., 2022, Wu et al., 16 Dec 2025, Fong et al., 22 Dec 2025)

A separate, conceptually distinct strand uses the Legendre-Fenchel transform to construct a single-valued Hamiltonian for non-convex Lagrangians and then analyze time translation symmetry breaking; this concerns the physical meaning of Legendre duality rather than numerical reduction (Chi et al., 2013).

2. Time-dimensional reduction for inverse PDEs

The most direct formulation appears in the inverse initial-data problem for the compressible anisotropic Navier-Stokes equations. The unknown is the initial velocity field $\mathbf U_0(x)$, the data are lateral boundary observations $\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$, and the key step is to expand the velocity in time using an exponentially weighted Legendre basis. The usual Legendre polynomials on $(-1,1)$ are rescaled to $(0,T)$ as $Q_n$, and the weighted basis is defined by

$\Psi_n(t):=e^t Q_n(t),$

which is orthonormal in

$$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$

For a function $u(x,t)$, the time coefficients are

$u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$

and the truncated expansion is

$\mathbb P^N[u](x,t)=\sum_{n=0}^N u_n(x)\Psi_n(t).$

Projecting the PDE onto this basis produces a coupled elliptic system in the spatial variables,

$\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$0

with projected boundary data

$\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$1

After solving for the coefficients, the full field is reconstructed by

$\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$2

and the initial velocity is recovered by evaluating at $\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$3: $\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$4 This is the paper’s explicit “Legendre time reduction method” (Van et al., 22 Jul 2025).

A closely related construction is used for inverse elastic source recovery. There, the displacement field $\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$5 is expanded componentwise in the same weighted time basis,

$\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$6

and the wave equation

$\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$7

is reduced to the coupled spatial system

$\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$8

with projected boundary traces $\mathbf f(x,t)=\partial_\nu \mathbf u(x,t)$9 and $(-1,1)$0. The initial displacement and initial velocity are then approximated by

$(-1,1)$1

In both cases, the time variable is not discretized by marching; it is replaced by finitely many Legendre coefficients that satisfy spatial boundary-value problems (Dang et al., 15 Jun 2025).

3. Basis properties, projection identities, and regularization

The analytical viability of the method depends on spectral properties of the weighted Legendre basis. For the Navier-Stokes inverse problem, the coefficients of a smooth temporal profile satisfy

$(-1,1)$2

and the derivatives obey

$(-1,1)$3

These estimates are used to prove compatibility between projection and differentiation, including

$(-1,1)$4

and

$(-1,1)$5

The elastic paper establishes an analogous theorem for second derivatives: if the differentiated Legendre series converges in $(-1,1)$6, then the generalized derivative $(-1,1)$7 exists in that space and equals the series sum (Van et al., 22 Jul 2025, Dang et al., 15 Jun 2025).

Regularization enters because the reduced systems remain inverse and Cauchy-type problems. For anisotropic Navier-Stokes, each Picard step is computed as the minimizer of a strictly convex stabilized least-squares functional with an $(-1,1)$8 penalty and projected Neumann-data misfit on $(-1,1)$9. The paper states that the functional is strictly convex, so the minimizer exists and is unique, and the implementation uses Matlab’s lsqlin (Van et al., 22 Jul 2025).

For elasticity, the quasi-reversibility functional is

$(0,T)$0

If $(0,T)$1 and $(0,T)$2 as $(0,T)$3, then the reconstructed field and its time derivative converge strongly to the exact minimal solution in the weighted space $(0,T)$4 (Dang et al., 15 Jun 2025).

4. Numerical behavior in inverse reconstruction

The anisotropic Navier-Stokes study uses the two-dimensional domain $(0,T)$5, final time $(0,T)$6, truncation $(0,T)$7, $(0,T)$8, and $(0,T)$9, with multiplicative boundary-flux noise

$Q_n$0

The Picard iteration starts from the zero initial guess and converges monotonically, with the logarithmic relative error between successive iterates decreasing steadily. Three reconstruction tests are reported. The first yields relative errors around $Q_n$1 and $Q_n$2 for the two components’ peak values. The second yields relative errors around $Q_n$3 and $Q_n$4 while preserving the correct diagonal elliptical structure. The third reconstructs a ring and a hollow square with relative errors around $Q_n$5 and $Q_n$6. The reconstructed initial fields are reported to closely match the true ones despite the $Q_n$7 noise (Van et al., 22 Jul 2025).

The elastic reconstruction experiments are also carried out in two dimensions on $Q_n$8 with $Q_n$9, $\Psi_n(t):=e^t Q_n(t),$0, $\Psi_n(t):=e^t Q_n(t),$1, and $\Psi_n(t):=e^t Q_n(t),$2. The forward problem is solved on the larger domain $\Psi_n(t):=e^t Q_n(t),$3, and the boundary data are contaminated by

$\Psi_n(t):=e^t Q_n(t),$4

with $\Psi_n(t):=e^t Q_n(t),$5. Three tests are reported: homogeneous isotropic, heterogeneous isotropic, and anisotropic media. Across them, the method recovers the main inclusion shapes and locations well, remains robust under $\Psi_n(t):=e^t Q_n(t),$6 noise, and reconstructs complex anisotropic features such as ellipses, rings, rectangles, and diagonal inclusions, although fine edges are somewhat blurred and velocity is more sensitive than displacement (Dang et al., 15 Jun 2025).

These results place Legendre time reduction within the broader class of reduction-plus-regularization methods: the Legendre truncation removes much of the unstable temporal burden, while quasi-reversibility or stabilized least squares controls the remaining ill-posedness.

5. Related Legendre reductions outside the strict time-variable setting

A prominent neighboring example is sparse Legendre spectral recovery. Instead of expanding in time, the problem is to recover a near-optimal $\Psi_n(t):=e^t Q_n(t),$7-term Legendre approximation

$\Psi_n(t):=e^t Q_n(t),$8

without computing all coefficients. The method uses Iserles’ Legendre-to-Fourier map

$\Psi_n(t):=e^t Q_n(t),$9

applies a sparse Fourier transform to identify the large Fourier frequencies of $$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$0, and then estimates the Legendre coefficients on the detected support by least squares using reweighted Legendre polynomials

$$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$1

The headline complexity is

$$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$2

contrasting with classical $$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$3 output cost and typical $$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$4 full-transform cost. The numerical section reports that the method is faster when $$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$5, gives errors as low as $$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$6 on true coefficients in many trials, and is robust for noisy approximately sparse signals provided the significant support is detected (Hu et al., 2015).

A second distinct meaning appears in the Quantum Legendre-Fenchel Transform. There the discrete convex conjugate

$$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$7

is not output as a classical list; instead, the algorithm prepares a quantum state encoding the transformed values. On regular dual grids the expected runtime is

$$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$8

while in the adaptive case with $$L^2_{e^{-2t}}(0,T)=\left\{u\in L^2(0,T):\int_0^T e^{-2t}|u(t)|^2\,dt<\infty\right\}.$$9 it becomes

$u(x,t)$0

For multivariate functions with $u(x,t)$1, the paper states an exponential quantum speedup in dimension $u(x,t)$2, and proves classical and quantum lower bounds by reduction from unstructured search (Sutter et al., 2020).

Other reduction axes are explicitly spatial, angular, or price-based rather than temporal. In an inverse memory convection-diffusion problem, a tensor-product Legendre basis $u(x,t)$3 is used for spatial dimensional reduction, and the truncation acts as a spectral filter on unstable high spatial frequencies (Van et al., 18 Jun 2026). In forward-time Black-Scholes reconstruction, shifted Legendre polynomials in the asset-price variable reduce the PDE to the ODE system

$u(x,t)$4

and the truncation serves as a spectral cutoff that also relaxes the $u(x,t)$5-degeneracy at the zero-price boundary (Nguyen et al., 30 May 2026). In the time-fractional radiative transport equation, Fourier transform in space plus Legendre expansion in angle yields the tridiagonal fractional system

$u(x,t)$6

whose solution is expressed through the matrix Mittag-Leffler operator $u(x,t)$7 (Machida, 2016). In slab-geometry transport, finite volume in space and Legendre polynomials in angle are combined with dynamical low-rank approximation, reducing memory from $u(x,t)$8 to $u(x,t)$9 when rank $u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$0 (Peng et al., 2019).

6. Temporal compression and representation-level uses

In machine learning, Legendre time reduction usually means temporal compression rather than time-variable elimination. FiLM formulates long-term forecasting as sequence compression via the Legendre Projection Unit,

$u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$1

where $u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$2 is a fixed-size memory state containing Legendre coefficients of a recent window. For an $u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$3-Lipschitz signal, the approximation error satisfies

$u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$4

and for bounded $u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$5-th derivatives,

$u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$6

FiLM then applies Fourier filtering to suppress high-frequency noise and a low-rank factorization in the frequency layer. The paper reports about $u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$7 relative MSE reduction in multivariate forecasting and about $u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$8 in univariate forecasting, and also reports plug-in improvements for several backbones (Zhou et al., 2022).

FLAME uses Legendre Memory as a continuous online function approximation that compresses a continuous-time history $u_n(x):=\int_0^T e^{-2t}u(x,t)\Psi_n(t)\,dt,$9 into a fixed-size state $\mathbb P^N[u](x,t)=\sum_{n=0}^N u_n(x)\Psi_n(t).$0. It distinguishes translated Legendre (LegT), used in encoding and local environmental compression, from scaled Legendre (LegS), used in decoding and SSD initialization. The paper explicitly states that it does not propose a distinct mechanism called “Legendre time reduction”; instead, it uses Legendre Memory as an efficient temporal encoding and compression mechanism for long-range inference (Wu et al., 16 Dec 2025).

A still more geometric usage appears in the Symplectic Reservoir representation, where the “reduction” is representational. The state is constrained to remain on Legendre graphs

$\mathbb P^N[u](x,t)=\sum_{n=0}^N u_n(x)\Psi_n(t).$1

and the discrete-time update

$\mathbb P^N[u](x,t)=\sum_{n=0}^N u_n(x)\Psi_n(t).$2

is derived from a Hamiltonian system. The paper’s structural theorem states that the relevant Legendre-preserving maps are exactly of the form

$\mathbb P^N[u](x,t)=\sum_{n=0}^N u_n(x)\Psi_n(t).$3

so the reservoir preserves Legendre duality step by step. This is not a reduction of time in the numerical-analysis sense, but a compression of dynamics into a constrained primal-dual representation (Fong et al., 22 Dec 2025).

Legendre Time Reduction therefore names not one algorithm but a recurrent methodological pattern: Legendre coordinates are used to isolate the structured part of temporal or transform behavior, after which inversion, approximation, or learning is carried out on a reduced coefficient system rather than on the full original object.