McLachlan-projected reduced dynamics for ill-posed Schrödingerized backward diffusion

This paper proposes and analyzes a McLachlan-projected reduced dynamics framework for the ill-posed backward diffusion problem, demonstrating that Schrödingerization combined with projection onto a low-dimensional frame acts as a structured regularizer with provable error bounds, Gram-norm conservation, and competitive performance against classical spectral filtering baselines.

The Gist

The Big Problem: Unraveling a Sweater in Reverse

Imagine you have a perfectly knitted sweater. If you pull a loose thread, the whole thing unravels into a messy pile of yarn. This is easy to do forward in time.

Now, imagine trying to do the reverse: taking that messy pile of yarn and magically knitting it back into a perfect sweater. This is the "Backward Diffusion" problem the paper tackles. In the real world, if you try to reverse a process like heat spreading out or a drop of ink dispersing in water, tiny, invisible specks of noise (like static on an old TV) get amplified exponentially. If you try to calculate this backward on a computer without special help, the noise grows so fast that the answer explodes into nonsense. It is an "ill-posed" problem, meaning it's mathematically unstable.

The Solution: Schrödingerization (The "Magic Elevator")

The authors use a technique called Schrödingerization. Think of this as taking your messy yarn problem and putting it into a "Magic Elevator" (an extended, higher-dimensional space).

In this new space, the rules change. Instead of the yarn unraveling chaotically, the problem transforms into a Hamiltonian system (like a quantum particle moving in a perfectly smooth, energy-conserving landscape). In this "Magic Elevator," the chaos is tamed, and the system evolves smoothly. This is the "Lift."

The New Challenge: The Elevator is Too Big

While the Magic Elevator solves the chaos, it creates a new problem: the elevator is huge. To simulate the full journey, you would need a supercomputer with massive memory to track every single thread of yarn in that high-dimensional space. It's too expensive and slow.

The paper asks: Can we take a shortcut? Can we just watch a few representative threads and guess the rest?

The Shortcut: McLachlan Projection (The "Shadow Puppet" Trick)

The authors propose a method called McLachlan projection. Here is the analogy:

Imagine you are in a dark room with a giant, complex puppet show (the full "Magic Elevator" simulation). You can't see the whole show, but you have a small screen. You want to project the show onto that small screen so you can still understand the story without needing the whole theater.

1. The Frame (The Screen): They pick a small, fixed set of "snapshots" (a few key moments of the yarn's movement) to build their small screen.
2. The Projection: They force the complex, high-dimensional motion to fit onto this small screen. They ask: "What is the best possible version of the story that fits on this small screen?"
3. The Result: This creates a Reduced Dynamics model. It's a smaller, faster version of the simulation that stays stable.

The Safety Net: Measuring the "Gap"

The paper proves that this shortcut isn't just a guess; it's a controlled approximation. They introduce a concept called the Projection Defect.

Think of this as a "leak detector." If you project a 3D object onto a 2D wall, you lose some depth information. The "defect" measures exactly how much information is lost when you squeeze the big simulation into the small screen.

• The Good News: The authors prove that if you know how much information is lost (the defect), you can mathematically guarantee that your small-screen version won't drift too far away from the truth.
• The Trade-off: If you make your screen smaller (fewer snapshots), you lose more detail (bias), but you filter out more noise (stability). If you make the screen bigger, you get more detail but risk letting noise back in. This is a classic "bias-variance trade-off."

The Quantum Twist: Noisy Measurements

Since this is a quantum computing paper, they also tested what happens if the measurements used to build the "screen" are noisy (like trying to take a photo in the dark with a shaky camera).

They found that even if the measurements are a bit fuzzy, the "Magic Elevator" structure protects the final result. The noise doesn't cause the whole thing to explode. However, they warn that if the "screen" is built poorly (mathematically "ill-conditioned"), small measurement errors can get amplified. They showed how to fix this by cleaning up the math before running the simulation.

The Conclusion: A Fair Comparison

Finally, the authors are very careful not to claim their method is a "magic cure-all." They compare their method against standard "low-pass filters" (which are like blurring a photo to remove grain).

They show that:

1. Unfiltered attempts (trying to reverse the yarn without a filter) fail immediately and explode.
2. Their method (Schrödingerization + Projection) produces a stable, accurate result that is comparable to the best classical filters.
3. The value: Their method provides a structured, mathematical way to decide how much detail to keep and how much to throw away, turning a chaotic, unstable problem into a manageable one.

In short: The paper shows how to take a mathematically broken, unstable problem, lift it into a stable quantum-like world, and then compress it into a smaller, faster model without losing the essential story, all while measuring exactly how much detail is being sacrificed.

Technical Summary

Technical Summary: McLachlan-Projected Reduced Dynamics for Ill-Posed Schrödingerized Backward Diffusion

Problem Statement
The paper addresses the numerical challenge of solving backward diffusion equations, which are prototypical ill-posed evolution problems. In these systems, high spatial frequencies grow exponentially over time, causing mesh-based time-marching schemes (such as Crank–Nicolson) without explicit regularization to be rapidly overwhelmed by noise. While the "Schrödingerization" framework has been established to embed non-unitary linear dynamics (including ill-posed problems) into Hermitian dynamics on an extended space, a critical gap remains: it is unclear whether the full lifted amplitude vector must be propagated at every time step, or if a fixed-dimensional reduced representation suffices. Furthermore, existing comparative numerics often fail to separate the effects of lifting, subspace truncation, and statistical estimation of matrix elements.

Methodology
The authors propose a structured regularization approach by applying McLachlan's variational principle to the Schrödingerized lifted system. The methodology proceeds in three layers:

1. Lifted Realization: The semidiscrete backward diffusion equation $\dot{u}_h = A_h u_h$ (where $A_h$ has non-negative eigenvalues) is mapped to a Hermitian system on an extended space of dimension $M$. This is formalized via Assumption 1, which posits the existence of a Hermitian generator $H$, an injection $J$, and a recovery map $R$ such that the physical state is approximated by $R e^{-itH} J u_0$ up to a lift error $\varepsilon_{\text{lift}}$.
2. McLachlan Projection: Instead of propagating the full $M$-dimensional state, the authors construct a reduced dynamics on a fixed frame $V$ of $m$ lifted snapshots. By minimizing the residual norm of the Schrödinger equation within the span of $V$, they derive a reduced evolution equation for the coefficient vector $c(t)$:
   $$iS \dot{c} = H_K c$$
   where $S = V^\dagger V$ is the Gram (overlap) matrix and $H_K = V^\dagger H V$ is the projected Hamiltonian.
3. Error Decomposition and Statistical Estimation: The paper rigorously decomposes the total error into spatial discretization, lift/recovery, projection, and sampling components. It specifically analyzes the impact of estimating the pencil $(S, H_K)$ via finite-shot quantum measurements (e.g., Hadamard tests), deriving perturbation bounds that depend on the condition number $\kappa(S)$.

Key Contributions
The paper makes the following specific theoretical and methodological contributions:

• Uniqueness and Conservation: It proves the uniqueness of the reduced flow solution and establishes that the projected Hermitian flow conserves the Gram norm ($d/dt(c^\dagger S c) = 0$).
• Error Bounds: The authors derive a lifted–reduced gap bound (Theorem 2) expressed in terms of the projection defect $\eta_V(t) = \|(I-P_V)H\Psi_m(t)\|_2$. This defect separates the full lifted propagation from the reduced trajectory, providing a quantifiable error budget.
• Perturbation Analysis: Proposition 1 provides bounds on the perturbation of the reduced generator and propagated coefficients when the overlap and Hamiltonian matrices are estimated with statistical noise. This highlights the critical role of the overlap matrix conditioning $\kappa(S)$ in amplifying shot noise.
• Fair Classical Baselines: The paper explicitly constructs fair classical baselines (spectral low-pass filtering or Tikhonov regularization) where the cutoff or ridge strength is matched to the information content (dimension $m$) of the chosen reduced frame. This isolates the performance of the Schrödingerized reduced pipeline from the instability of unregularized solvers.

Results
Numerical experiments on a 1D backward diffusion problem ($N_x=7$, $T=0.3$) using Qiskit Aer with finite-shot estimation ($10^4$ shots) demonstrate:

• Stability vs. Unregularized Methods: An unregularized Crank–Nicolson scheme exhibits a discrete relative $L_2$ error of $\approx 41.99\%$ due to noise blow-up. In contrast, the Schrödingerized lift (full dimension) limits error to $\approx 19.37\%$.
• Reduced Dynamics Performance: The $m=4$ McLachlan projection with classical (exact) matrices yields $\approx 20.02\%$ error. When using Aer-sampled matrices (simulating quantum hardware noise), the error is $\approx 19.82\%$, with a transient peak of $\approx 22.03\%$.
• Noise Masking: Despite a large Frobenius norm discrepancy between the sampled and exact Hamiltonian ($\| \Delta H \|_F / \| H_K \|_F \approx 7.08$) and a highly ill-conditioned overlap matrix ($\kappa(S) \approx 1.7 \times 10^4$), the recovered physical trajectory does not exhibit catastrophic blow-up. The results confirm that the projection layer acts as a structured regularizer, and the error remains within the tens-of-percent regime, consistent with the theoretical bounds derived from the projection defect and sampling perturbations.

Significance and Claims
The authors position this work as a methodological advancement rather than a claim of solving ill-posedness entirely. They explicitly state:

• No Cure for Ill-Posedness: Schrödingerization does not remove the ill-posed nature of the recovered physical variable, nor does projection supersede optimal Tikhonov design.
• Interpretable Regularization: The primary contribution is framing the projection error as an "interpretable knob" analogous to a low-pass bandwidth. This allows for "like-for-like" comparisons between quantum-inspired reduced dynamics and classical regularized methods.
• Separation of Errors: The framework successfully isolates the contributions of lifting, subspace truncation, and statistical estimation, preventing the conflation of these distinct error sources that often occurs in current literature.
• Validation of Structure: The results illustrate that even with significant statistical noise in the generator estimation, the structured nature of the McLachlan projection prevents the exponential amplification of noise that characterizes unregularized backward diffusion.