Carleman--Picard and time-dimensional reduction for inverse initial-data problems in nonlinear transport with memory

The paper studies an inverse initial-data problem for a quasilinear transport equation with nonlinear and memory effects. The unknown initial state is reconstructed from time-dependent measurements on the outflow boundary, with prescribed inflow data. The authors first apply a Legendre–exponential time-dimensional reduction to transform the governing equation into a finite nonlinear system in space. They then develop a Carleman-weighted and Tikhonov-regularized Picard method. At each iteration, the nonlinear terms are evaluated using the previous iterate, leading to a linear minimization problem with a unique solution. A Carleman estimate for the principal transport operator is used to prove that, for a sufficiently large Carleman parameter, the resulting Picard map is contractive on a prescribed admissible set. Consequently, the method converges from an arbitrary initial guess in that set. The authors also establish stability with respect to noisy outflow data. These analytical results concern the truncated and regularized reduced problem. Two-dimensional numerical experiments demonstrate accurate reconstruction of single and multiple inclusions, robustness with respect to noise, and rapid convergence of the Picard iteration.

The paper is organized as follows. Section 2 states the geometric propagation condition, establishes the Carleman estimate for H · ∇, and recalls the Legendre–exponential basis. Section 3 derives the finite nonlinear system for the spatial modal coefficients by time-dimensional reduction. Section 4 introduces the Carleman-weighted Picard map and proves its contraction property and global convergence within the admissible set. Section 5 establishes stability of the reduced reconstruction with respect to noisy outflow data. Section 6 describes the numerical implementation and presents the reconstruction experiments, including the study of the Carleman parameter. Finally, Section 7 summarizes the main results.

The governing equation is:

∂ t u + c(x, u)H(x) · ∇u + f(x, u) + ∫ 0 t α(x, t − t′)u(x, t′) dt′ = 0, (x, t) ∈ Q T,

with initial condition u(x, 0) = u 0(x) and inflow boundary condition u(x, t) = g(x, t) on Γ− × (0, T).

The inverse problem (Problem 1.1) is: Given the outflow boundary observation h(x, t):= u(x, t) on Γ+ × (0, T), reconstruct the initial state u 0(x) for all x ∈ omega.

The paper addresses two principal difficulties: the nonlinearity of the forward equation (from the dependence of c and f on u, and the Volterra memory term) and the backward, ill-posed character of the reconstruction. The authors combine regularization with a Carleman-based globally convergent technique, specifically the Carleman contraction principle, because it leads to a computationally inexpensive iteration, converges rapidly in practice, and is comparatively simple to implement.

The geometric propagation condition requires that every forward integral curve of H starting in omega reaches the outflow boundary Γ+ in finite time, with the uniform nontrapping condition τ H(x) 0 is a lower bound for c(x, s).

The Carleman estimate (Proposition 2.1) states: Let φ* satisfy H(x) · ∇φ*(x) ≥ µ 0 > 0. Then there exist constants λ 0 > 0 and C > 0 such that, for all λ ≥ λ 0 and all v ∈ H 1(omega),

λ ∫ omega e 2λφv2 dx ≤ C ∫ omega e 2λφH · ∇v2 dx + Cλ ∫∂omega e 2λφ*v2 dS.

The Legendre–exponential basis is defined as Ψ n(t) = sqrt((2n+1)/T) e t P n(2t/T − 1), n ≥ 0, where P n is the Legendre polynomial. This basis is orthonormal in L2 e−2t(0, T). A key property is that none of its basis functions has an identically vanishing derivative, so no retained spatial coefficient is eliminated when the truncated temporal expansion is differentiated.

The time-dimensional reduction proceeds by dividing both sides of (1.1) by c(x, u), expanding u and its time derivative in the basis, and projecting onto the basis functions. This yields the truncated system (3.8):

H(x) · ∇u m(x) + Σ n=0 N S mn(x, u j j=0 N) u n(x) + Σ n=0 N M mn(x, u j j=0 N) u n(x) + F m(x, u j j=0 N) = 0,

for m = 0,..., N, with boundary conditions u m = g m on Γ− and u m = h m on Γ+.

The Carleman-weighted functional (4.2) is defined as:

J Φ λ,ε (U):= ∫ omega e 2λφH(x) · ∇U(x) + N(x; Φ, U)2 dx + λ2 ∫ Γ− e 2λφU − G2 dS + λ2 ∫ Γ+ e 2λφ*U − D obs2 dS + ε∥U∥2[H s(omega)] N+1,

where N(x; Φ, V) collects the frozen nonlinear terms. The Picard map is T λ,ε(Φ):= argmin U∈B M J Φ λ,ε (U).

The main convergence result (Theorem 4.1) states that there exist constants λ 1 > 0 and C > 0 such that, for every fixed ε > 0, all λ ≥ λ 1, and all Φ, Ψ ∈ B M,

∥T λ,ε(Φ) − T λ,ε(Ψ)∥ λ,ε,φ* ≤ sqrt(C/λ) ∥Φ − Ψ∥ λ,ε,φ*.

In particular, if λ > max λ 1, C, then T λ,ε is a strict contraction on B M.

Corollary 4.1 states that for arbitrary U(0) ∈ B M, the Picard sequence U(k+1) = T λ,ε(U(k)) converges in ∥·∥ λ,ε,φ* to the unique fixed point U λ,ε ∈ B M, with geometric convergence rate µ k where µ = sqrt(C/λ) < 1.

The stability result (Theorem 5.1) establishes that if the measurement error satisfies ∥D obs − D*∥[L2(Γ+)] N+1 ≤ δ∥D*∥[L2(Γ+)] N+1 with 0 < δ < 1, then the fixed point U λ,ε,δ satisfies:

∫ omega e 2λφU λ,ε,δ − U2 dx + λ ∫∂omega e 2λφU λ,ε,δ − U2 dS + (ε/λ)∥U λ,ε,δ − U*∥2[H s(omega)] N+1 ≤ (Cε/λ)∥U*∥2[H s(omega)] N+1 + Cλ ∫ Γ+ e 2λφD obs − D2 dS.

The numerical experiments use the domain omega = (−1, 1)2, with advection field H(x, y) = (1 + 0.25 sin(πy) cos(πx), 0.5 + 0.20 cos(πx) cos(πy)) T, c(x, y, s) = 1 + 0.15 sin(πx) cos(πy) + 0.2s2, f(x, y, s) = 0.02(1 + 0.25 cos(πx) sin(πy))s, and α(x, y, s) = 0.02(1 + 0.20 sin(πx) cos(πy))e−s. The computations use T = 1.5, λ = 4, ε = 10−6, and the Carleman weight φ*(x, y) = x + 0.5y.

Three test profiles are considered: a disk inclusion, two Gaussian inclusions, and a Y-shaped inclusion. For the disk inclusion, the plateau criterion selects N = 13 for both 5% and 10% noise, with maximum reconstructed values of 0.889461 (11.05% error) and 0.883682 (11.63% error), respectively. For the two Gaussian inclusions, N = 10 for 5% noise and N = 8 for 10% noise, with relative maximum-value errors of 10.28% and 10.24% for the lower and upper inclusions at 5% noise, and 13.64% and 16.68% at 10% noise. For the Y-shaped inclusion, N = 14 for both noise levels, with relative maximum-value errors of 3.43% and 3.37% for 5% and 10% noise, respectively.

The Carleman parameter study shows that λ = 0 (unweighted Picard) produces unsatisfactory reconstructions, while moderate positive values (λ = 1, 2, 3, 4) improve localization and separation of inclusions. However, λ = 7 deteriorates due to numerical conditioning, as the weight e 2λφ* ranges from e−21 to e 21 (ratio approximately 1.7 × 10 18), making the discrete system severely ill-conditioned.

The paper concludes that the theoretical results concern the truncated and regularized reduced problem for fixed values of the truncation index and Tikhonov regularization parameter. The temporal truncation serves as an additional layer of regularization by suppressing high-order temporal modes, and the Tikhonov term provides a second regularization mechanism for the resulting spatial system.

Based on this paper, here are specific improvements for AI systems:

1. Carleman-Weighted Contraction-Based Inverse Solvers

• Improvement: Implement the Carleman-weighted Picard iteration (Theorem 4.1) as a novel optimization strategy for ill-posed inverse problems. The key innovation is using a Carleman weight e 2 lambda phi* to enforce contraction in a weighted norm, guaranteeing global convergence from arbitrary initial guesses—unlike standard gradient descent or Newton methods that require good initialization.

• Capability: An AI system can now solve nonlinear inverse problems (e.g., medical imaging, geophysical inversion) with provable convergence guarantees, even when starting from poor initial estimates, by automatically selecting lambda to ensure the contraction factor sqrt C/lambda < 1.

2. Adaptive Regularization via Temporal Dimensional Reduction

• Improvement: Use the Legendre–exponential basis (orthonormal in L 2 e-2t) to reduce time-dependent PDEs to finite spatial systems. The property that no basis function has vanishing derivative prevents information loss during differentiation—a critical advantage over standard Fourier or polynomial bases.

• Capability: An AI system can handle time-history-dependent (memory) effects in dynamical systems (e.g., viscoelastic materials, delayed feedback control) by automatically truncating temporal modes, with the truncation index N selected via a plateau criterion to balance accuracy and computational cost.

3. Noise-Robust Reconstruction with Explicit Error Bounds

• Improvement: Integrate the stability result (Theorem 5.1) into AI pipelines to provide quantitative uncertainty quantification. The bound shows reconstruction error scales as O(delta) for measurement noise delta, with explicit dependence on the Carleman parameter lambda and Tikhonov parameter epsilon.

• Capability: An AI system can output not just reconstructions but also rigorous error bars, enabling safe deployment in high-stakes applications (e.g., nondestructive testing, climate modeling) where noisy sensor data is unavoidable.

4. Hybrid Nonlinearity Handling via Frozen-Coefficient Linearization

• Improvement: Adopt the freeze nonlinear terms at previous iterate strategy from the Picard map. This converts a nonlinear problem into a sequence of linear minimization problems, each with a unique solution—avoiding the need for expensive Jacobian computations or Hessian approximations.

• Capability: An AI system can solve strongly nonlinear PDE-constrained inverse problems (e.g., parameter identification in reaction-diffusion systems) with significantly lower per-iteration cost, while retaining rapid convergence (observed numerically in 2–5 iterations).

5. Carleman Parameter Auto-Tuning

• Improvement: Implement the numerical finding that moderate lambda values (1–4) improve reconstruction, while lambda=0 fails and lambda=7 degrades due to conditioning. This suggests an adaptive scheme: start with lambda=0, increase until contraction is achieved, but monitor the condition number of the discrete system (weight ratio e 2 lambda(phi* max-phi* min)).

• Capability: An AI system can autonomously select the optimal Carleman weight for each problem instance, balancing convergence speed against numerical stability—useful for real-time inverse problems where manual tuning is impractical.

6. Multi-Inclusion Reconstruction with Geometric Propagation Constraints

• Improvement: Leverage the geometric nontrapping condition (every integral curve of H reaches the outflow boundary in finite time) to design AI architectures that respect physical propagation directions. This prevents spurious reconstructions in shadow regions.

• Capability: An AI system can accurately reconstruct multiple disjoint inclusions (as demonstrated with disk, Gaussian, and Y-shaped profiles) even in advection-dominated regimes, by enforcing that information flows along characteristic curves—critical for flow diagnostics or pollution source identification.

7. Two-Layer Regularization for Ill-Posedness

• Improvement: Combine temporal truncation (suppressing high-order modes) with Tikhonov regularization (spatial smoothing) as a dual mechanism. The paper shows this yields stable reconstructions with relative errors below 12% even at 10% noise.

• Capability: An AI system can handle severely ill-posed problems (e.g., backward heat conduction, wave source reconstruction) by automatically balancing these two regularization layers, avoiding over-smoothing while maintaining numerical stability.