This work addresses an inverse reconstruction task for a time-fractional pseudo-parabolic model with a temporally varying coefficient. By imposing Dirichlet boundary conditions, we aim to recover the unknown initial state from observations collected at the final time. From a theoretical perspective, we derive existence and uniqueness results by proving that, under suitable hypotheses, the problem admits a unique solution. Computationally, we introduce a finite-difference discretisation based on a time-stepping strategy and provide a detailed stability and convergence analysis. Leveraging the resulting forward solver, we then formulate an initial-data identification procedure using Tikhonov regularisation. The proposed approach is validated with numerical simulations, and its resilience is assessed via experiments that incorporate perturbations in the final-time measurements.
Alex: Welcome to another episode of ResearchPod.
Sam: Today, we're looking at a piece of research on figuring out the starting point of certain physical processes when you only see the end result.
Alex: So, Sam, what's the core puzzle here?
Sam: This paper by Arshyn Altybay tackles a backward problem for a time-fractional pseudo-parabolic equation with a variable coefficient. The question is: given measurements of a system's state at the final time, can we uniquely reconstruct what that state was at the very beginning? These equations model processes like water moving through soil or rock, where flow has a kind of memory of past conditions.
Alex: Like running time backwards for diffusion, such as moisture spreading in soil?
Sam: Yes. Normally, you start with initial moisture and predict the end. Here, you measure the final moisture but need to recover the initial state—say, before a pump failed. The backward direction is tricky because small errors in the final measurement can grow huge when rewinding, due to the memory effects and time-changing material properties.
Alex: They've got a proof it works, plus a numerical method. How do they actually rewind without noise wrecking it?
Sam: They expand the solution into sine waves—like breaking a sound into frequencies. This turns the equation per mode into a simpler integral they solve explicitly for the initial coefficients. For noisy data, a regularization trick stabilizes it.
Alex: So each wave mode has its own starting value. What guarantees they can solve for it cleanly?
Sam: For each mode, the value at any time is the initial amount times an accumulation factor A_k(t), plus a term from external forces. At the end time T, they rearrange: initial equals (final minus force term) divided by A_k(T). A lemma proves A_k(T) stays positive, like a rubber band that never fully relaxes.
Alex: With infinitely many modes, how do they ensure the full picture adds up smoothly?
Sam: The series converges uniformly across space and time, because smooth data makes high-frequency coefficients decay fast—like 1 over k squared. This keeps the solution and its second space derivative continuous everywhere. Time derivatives converge away from t=0, where memory effects peak.
Alex: That bridges theory to a reliable reconstruction. But how do they prove only one initial state fits?
Sam: Consider the difference between two solutions—it satisfies the equation with zero final value and no forces. For each mode, that difference is initial chi_k times A_k(t). Since final is zero and A_k(T) never zero, chi_k must be zero everywhere—uniqueness.
Alex: The positive A_k locks it down. And stability for noise?
Sam: Yes—reconstructed initial stays bounded by the final data size plus forces, with proportional changes for small input shifts.
Alex: Unique and stable, even with imperfect data. Now, for computation?
Sam: They discretize with graded time steps near zero for memory singularities, L1 weights for the fractional part, and central space differences. The matrix comes from impulse runs, inverted with regularization.
Alex: Forward solver tuned. But inverse stability?
Sam: Coercivity in the L1 memory approximation keeps energy bounded—like friction stopping runaway speed. The matrix is symmetric positive definite, ensuring unique steps and unconditional stability.
Alex: No grid disasters. How do they build the practical inverse?
Sam: Split final state into initial part—via forward runs from basis vectors—and forced part. Tikhonov minimizes mismatch plus penalty on size, solving stably.
Alex: Penalty grounds wild guesses. They verify by re-running forward?
Sam: Yes. Errors drop with finer grids in noise-free tests using known solutions. With one to five percent Gaussian noise, errors scale linearly to about four to eighteen percent, preserving shapes.
Alex: Proportional, no chaos. Robust across memory strengths.
Sam: Plots show tight overlaps for clean data, shape-faithful under noise. Limits include first-order time accuracy and computation scaling with grid size squared, suiting moderate cases.
Alex: A solid package—theory for uniqueness and stability, numerics that handle noise—for diagnosing starts in soil filtration or similar. Thanks for joining us on ResearchPod.