Tikhonov-regularised projected gradient flow for equality-constrained bilinear quantum control

This paper introduces a Tikhonov-regularised projected gradient flow for equality-constrained bilinear quantum control that rigorously ensures objective monotonicity and stability by replacing the ill-conditioned Gram matrix with a regularised version, thereby eliminating heuristic safeguards and significantly reducing constraint drift while maintaining high fidelity in numerical benchmarks.

The Gist

Imagine you are a conductor trying to guide an orchestra (a quantum system) to play a specific, perfect note (a target state). You have a baton (the control field) that you can wave to steer the musicians. However, you have strict rules you must follow while conducting:

1. The "Silence" Rule: Your baton must start and end in the exact same spot (zero net movement).
2. The "Energy" Rule: You cannot wave your baton too wildly; the total energy of your movements is capped.
3. The "Rhythm" Rule: Your movements must sync up with a specific beat in the music.

This is the problem of Quantum Optimal Control. The goal is to find the perfect wave pattern for your baton that gets the orchestra to the right note while obeying all three rules.

The Problem: A Wobbly Ladder

The paper discusses a mathematical method called Projected Gradient Flow. Think of this as a hiker trying to climb a hill (maximizing the quality of the music) while staying on a narrow, winding path (the rules).

In a perfect, continuous world, this hiker moves smoothly up the hill, never slipping off the path. But in the real world, we have to take steps (discretization). When the path gets tricky—specifically, when the rules start to "fight" each other or become very similar—the mathematical map the hiker uses to stay on the path becomes ill-conditioned.

The Analogy: Imagine the map is a ladder. If the rungs of the ladder are very far apart and the wood is rotten, the ladder is "ill-conditioned." If you try to climb it, you might slip, fall, or have to take tiny, hesitant steps. In the paper's specific experiment, this "ladder" was so wobbly that the computer had to take steps so small it was practically crawling, and sometimes it would slip off the path entirely, breaking the rules (like wasting too much energy).

The Solution: Tikhonov Regularization (The "Shock Absorber")

The authors propose a fix called Tikhonov Regularization.

The Metaphor: Imagine adding a shock absorber or a stabilizer to that wobbly ladder.

• Without the stabilizer (The old way): The ladder is pure wood. If the ground is uneven (the math gets messy), the ladder shakes violently. You have to guess how small your steps should be. If you guess wrong, you fall.
• With the stabilizer (The new way): You add a flexible, springy support (represented by a number called $\epsilon$). This doesn't change the destination, but it makes the ladder much sturdier. It allows you to take bigger, safer steps without falling off.

What the Paper Proves

The authors didn't just say "this works"; they proved exactly how it works using five key findings:

1. The Stability Formula: They found a precise mathematical recipe showing that adding the stabilizer ($\epsilon$) makes the "ladder" (the math matrix) much sturdier. The wobbly parts become solid.
2. No Backsliding: Even with the stabilizer, the hiker never goes down the hill. The quality of the music (the objective) always gets better or stays the same; it never gets worse.
3. The Tiny Drift: Because the stabilizer is slightly flexible, the hiker might drift very slightly off the exact path (the rules). However, the authors proved this drift is tiny—specifically, it grows with the square of the stabilizer size. If you make the stabilizer 10 times smaller, the drift becomes 100 times smaller.
4. Convergence: As you make the stabilizer smaller and smaller (approaching zero), the hiker's path becomes identical to the original, perfect path.
5. The Step-Safe Rule: They derived a clear rule for how big your steps can be. Instead of guessing or checking if you fell after every step, you can calculate the perfect step size based on how sturdy your stabilizer is.

The Real-World Test

The authors tested this on a specific scenario: preparing a "Bell State" (a special, entangled connection) between two atoms using light.

• The Old Way: The computer struggled. The "ladder" was so wobbly that the condition number (a measure of instability) was between 1 billion and 100 billion. The computer had to reject many steps, and the energy rule was violated by nearly 40%.
• The New Way: By adding a moderate stabilizer, the computer stopped rejecting steps. The energy violation dropped from 40% to just 3%, and the final result was just as perfect (99.99% fidelity).

Summary

In simple terms, this paper takes a powerful but unstable mathematical tool for controlling quantum systems and adds a "shock absorber" to it. This makes the tool robust enough to handle difficult, real-world constraints without breaking, allowing scientists to design better quantum pulses without the computer getting stuck or making mistakes.

Technical Summary

1. Problem Statement

The paper addresses the constrained optimal control of bilinear quantum systems. The goal is to maximize a terminal fidelity objective $J[E]$ (the probability of transitioning a quantum state $\psi_0$ to a target state $\psi_f$) by optimizing a scalar control field $E(t)$.

The system dynamics are governed by the Schrödinger equation:
$$i \dot{\psi}(t) = (H_0 - \mu E(t)) \psi(t)$$
where $H_0$ is the field-free Hamiltonian and $\mu$ is the transition-dipole operator.

The Challenge: In practical applications (e.g., laser pulse design), the control field must satisfy integral equality constraints (e.g., zero pulse area, fixed fluence/energy, or fixed interaction with a reference frequency). These constraints define a manifold $\mathcal{M}_C$.
Existing methods use a projected gradient flow (specifically the D-MORPH framework) to navigate this manifold. However, this approach relies on inverting a moving Gram matrix $\Gamma(s)$ constructed from the gradients of the objective and the constraints.

• Pathology: In many realistic scenarios, the constraint gradients become nearly linearly dependent, causing $\Gamma(s)$ to be severely ill-conditioned (condition numbers $\sim 10^9 - 10^{11}$).
• Consequence: Standard implementations require heuristic step-size reductions (step-halving) to maintain stability, which is computationally expensive and lacks a rigorous mathematical foundation.

2. Methodology

The author proposes replacing the ill-conditioned Gram matrix $\Gamma(s)$ with its Tikhonov-regularized counterpart:
$$\Gamma_\varepsilon(s) = \Gamma(s) + \varepsilon^2 I$$
where $\varepsilon \geq 0$ is a regularization parameter.

The paper analyzes the resulting regularized projected gradient flow:
$$\frac{\partial E_\varepsilon}{\partial s} = S(t) g_0^{(\varepsilon)}(s) \sum_{\ell=0}^M [\Gamma_\varepsilon(s)^{-1}]_{0\ell} c_\ell(E_\varepsilon(s, \cdot); t)$$
where $S(t)$ is a gating envelope, $c_\ell$ are the gradients, and $g_0$ is a scalar scaling factor.

The analysis covers:

1. Continuous-time properties: Spectral conditioning, monotonicity of the objective, and constraint drift.
2. Discrete-time stability: Deriving a Courant–Friedrichs–Lewy (CFL) criterion for the forward-Euler discretization.
3. Convergence: Proving the rate at which the regularized solution converges to the unregularized solution as $\varepsilon \to 0$.

3. Key Contributions & Theoretical Results

The paper establishes five rigorous mathematical results:

• R1: Exact Spectral Identity & Conditioning:
  The spectrum of $\Gamma_\varepsilon$ is exactly $\{\sigma_k^2 + \varepsilon^2\}$, where $\sigma_k$ are singular values of the assembly operator. This yields an explicit conditioning formula:
  $$\text{cond}(\Gamma_\varepsilon) = \frac{\sigma_{\max}^2 + \varepsilon^2}{\sigma_{\min}^2 + \varepsilon^2}$$
  This proves that even if $\sigma_{\min} = 0$ (rank deficiency), $\Gamma_\varepsilon$ remains invertible for any $\varepsilon > 0$.

• R2: Persistence of Monotonicity:
  The objective function $J$ remains non-decreasing ($dJ/ds \geq 0$) along the regularized flow for any $\varepsilon \geq 0$. The regularization only scales the ascent rate but never reverses the direction.

• R3: Exact Constraint-Drift Identity:
  The violation of constraints is not arbitrary; it follows an exact identity:
  $$|h_m[E_\varepsilon(s)] - C_m| = O(\varepsilon^2)$$
  The paper provides a computable prefactor for this drift, showing it is quadratic in $\varepsilon$.

• R4: $L^2$-Convergence Rate:
  Under the assumption that the unregularized Gram matrix is uniformly invertible, the regularized trajectory $E_\varepsilon$ converges to the unregularized trajectory $E_0$ in the $L^2$ norm at a rate of $O(\varepsilon^2)$.

• R5: Discrete CFL Stability Criterion:
  For the forward-Euler discretization, the paper derives a stability condition:
  $$\Delta s \cdot G \cdot \|\Gamma_\varepsilon^{-1}\| \leq \alpha < 2$$
  where $G$ bounds the second variation of the objective. This transforms the heuristic "step-halving" into a transparent, calculable stability mechanism: increasing $\varepsilon$ reduces $\|\Gamma_\varepsilon^{-1}\|$, thereby allowing larger stable step sizes $\Delta s$.

4. Numerical Results

The theory is validated on a three-level bilinear benchmark modeling the all-optical preparation of a Bell state in two dipole-dipole coupled atoms.

• Ill-Conditioning: The unregularized Gram matrix consistently exhibited condition numbers between $10^9$ and $10^{11}$.
• Drift vs. Stability:
  • Unregularized ($\varepsilon=0$): The fluence constraint (nonlinear) drifted by 20–38%, and the algorithm frequently rejected steps due to instability.
  • Regularized ($\varepsilon \sim 10^{-2}$):
    • Eliminated step rejections entirely.
    • Reduced fluence drift to ~3%.
    • Maintained final fidelity $J \geq 0.9999$ (unchanged from the unregularized case).
• Convergence Verification: Numerical experiments confirmed the $O(\varepsilon^2)$ convergence rate over eight decades of $\varepsilon$, matching the theoretical prediction of Theorem 4.6.
• CFL Validation: The regularized scheme allowed for larger time steps without violating monotonicity, confirming the derived stability bound.

5. Significance and Impact

• Theoretical: The paper resolves the mathematical gap regarding the stability of projected gradient flows in ill-conditioned settings. It provides the first rigorous error analysis linking Tikhonov regularization to constraint drift and step-size stability in quantum control.
• Practical: It offers a robust alternative to heuristic step-size tuning. By choosing a moderate $\varepsilon$ (e.g., $10^{-3}$ to $10^{-2}$), engineers can:
  1. Stabilize the optimization process.
  2. Avoid expensive step-rejection loops.
  3. Control constraint violations within a predictable $O(\varepsilon^2)$ bound.
• Applicability: While demonstrated on quantum control, the framework applies to any optimal control problem with integral equality constraints where gradients become nearly collinear, including molecular dynamics and superconducting qubit control.

In summary, the paper transforms a heuristic fix for ill-conditioning into a principled, mathematically rigorous method that improves both the stability and efficiency of constrained quantum optimal control algorithms.