⚛️ quantum physics

Accelerating Feedback-based Algorithms for Quantum Optimization Using Gradient Descent

This paper proposes a hybrid method that integrates per-layer gradient estimation into Quantum Lyapunov Control to accelerate the convergence of feedback-based quantum optimization algorithms while maintaining their low training overhead and stability guarantees.

Original authors: Masih Mozakka, Mohsen Heidari

Published 2026-02-16

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Masih Mozakka, Mohsen Heidari

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find the lowest point in a vast, foggy mountain range (the Optimization Problem). Your goal is to reach the deepest valley (the Best Solution) as quickly as possible.

In the world of quantum computing, there are two main ways people have tried to solve this:

1. The Old Way: The "Blind Hiker" (QAOA)

The standard method, called QAOA, is like a hiker who takes a step, stops, asks a guide, "Am I going up or down?", takes another step, and asks again.

The Problem: The guide is slow and expensive to hire (it requires many quantum measurements). Also, as the mountain gets bigger, the fog gets so thick that the hiker can't tell which way is down at all (this is called a "barren plateau"). The hiker gets stuck wandering in circles for a very long time.

2. The Feedback Way: The "Magnetic Compass" (FALQON/QLC)

To fix the "wandering" problem, scientists invented a method called FALQON (or Quantum Lyapunov Control).

How it works: Instead of asking the guide every time, the hiker carries a magical Magnetic Compass. This compass always points slightly downhill. The hiker just takes a tiny step in the direction the compass points, then checks the compass again.
The Good News: It guarantees you never go uphill. You are always moving toward the valley. It's very stable and doesn't get lost in the fog.
The Bad News: The compass only points slightly downhill. To get to the bottom of a deep valley, the hiker has to take thousands of tiny, shuffling steps. It's safe, but it's incredibly slow. It's like trying to empty an ocean with a teaspoon.

3. The New Solution: The "Smart Compass" (GD-QLC)

This paper introduces a new hybrid method called GD-QLC. It keeps the safety of the Magnetic Compass but adds a GPS and a Sprinting Coach.

Here is the analogy of how it works:

The Setup: You still have the Magnetic Compass (the feedback law) that guarantees you don't go uphill.
The Innovation: Before you take your next step, the "Smart Compass" doesn't just point. It quickly runs a mini-simulation (Gradient Descent) to ask: "If I take a slightly bigger step in this direction, or a slightly different angle, will I get to the bottom faster?"
The Result: Instead of taking 1,000 tiny shuffling steps, the hiker takes 100 confident, well-calculated strides.
- It still guarantees you never go uphill (stability).
- But it reaches the bottom much faster (convergence).
- It doesn't require the expensive "guide" to be consulted thousands of times (low training overhead).

Why is this a big deal?

The authors tested this on several "mountain ranges" (math problems like MAX-CUT, which is about dividing a group of people into two teams so that the most arguments happen between the teams).

They found that:

Speed: The "Smart Compass" (GD-QLC) found the solution much faster than the old "Magnetic Compass" (FALQON).
Stability: Unlike other fast methods that might overshoot the valley and get stuck on a cliff, this method stayed smooth and steady.
Robustness: Even if the "steps" were a bit too big or too small (a common problem in quantum computers), the Smart Compass still worked well. The old methods would often crash or fail if the step size wasn't perfect.

The Bottom Line

Think of this paper as upgrading a slow, safe, but tedious walking tour into a fast, safe, and efficient guided hike.

They took a method that was guaranteed to work but was too slow to be useful, and they injected a little bit of "mathematical muscle" (Gradient Descent) into every single step. The result is a quantum algorithm that is fast enough to be practical on today's noisy, imperfect quantum computers, without losing the safety guarantees that make it reliable.

1. Problem Statement

The paper addresses the limitations of current quantum optimization algorithms, specifically focusing on the trade-offs between Quantum Approximate Optimization Algorithm (QAOA) and feedback-based methods like FALQON (Feedback-based Algorithm for Quantum Optimization) and Quantum Lyapunov Control (QLC).

QAOA Challenges: While QAOA is a leading hybrid classical-quantum approach, it suffers from high training overhead due to the need for extensive measurement shots to estimate gradients at every iteration. Furthermore, as circuit depth increases, QAOA faces "barren plateaus" (exponentially vanishing gradients) and non-convex optimization landscapes, making it difficult to train.
Feedback-Based (QLC/FALQON) Challenges: Methods like FALQON use feedback laws derived from Lyapunov functions to deterministically set control parameters, guaranteeing monotonic improvement of the objective function without a global classical optimization loop. However, they often require long, finely discretized control sequences (deep circuits) to achieve high accuracy. For example, FALQON may require circuit depths two orders of magnitude larger than QAOA to solve the same MAX-CUT problem, leading to slow convergence and implementation challenges on near-term hardware.

Core Objective: The authors aim to accelerate the convergence of feedback-based methods and reduce the required circuit depth while preserving their stability guarantees and low training overhead.

2. Methodology: GD-QLC

The authors propose a hybrid framework called GD-QLC (Gradient Descent - Quantum Lyapunov Control). This method integrates per-layer gradient descent into the standard FALQON/QLC workflow.

Key Mechanisms:

Layer-wise Optimization: Instead of determining the control parameter $\beta_k$ for layer $k$ using a single measurement-based feedback step (as in standard FALQON), GD-QLC performs $L$ iterations of gradient descent on $\beta_k$ before moving to the next layer.
Objective Function: The method minimizes the time derivative of the energy function, $\dot{E}_p(\beta_k)$ $\dot{E}_{p} (β_{k})$ , at each layer.
- The energy derivative is given by $\dot{E}_p = A(t)\beta(t)$ , where $A(t) = \langle \psi | i[H_d, H_p] | \psi \rangle$ .
- To perform gradient descent, the algorithm computes the derivative of this objective with respect to $\beta$ $β$ . This requires estimating two observables:
  - $A_k = \langle \psi_k | i[H_d, H_p] | \psi_k \rangle$
  - $B_k = \langle \psi_k | [H_d, [H_d, H_p]] | \psi_k \rangle$ (the second-order commutator term).
Update Rule: The control parameter $\beta_k$ is updated iteratively using the rule:
$\beta_k^{(l+1)} = \beta_k^{(l)} (1 + \eta \Delta t B_k) - \eta A_k$
where $\eta$ is a learning rate and $l$ is the iteration index within the layer.
Selection: After $L$ iterations, the best $\beta_k$ (minimizing $\dot{E}_p$ ) is selected, and the algorithm proceeds to the next layer.

Overhead Analysis:

FALQON: Requires $O(K \cdot N_{shots})$ measurements for $K$ layers.
GD-QLC: Requires $O(K \cdot L \cdot N_{shots})$ measurements.
Trade-off: While GD-QLC increases the measurement cost per layer by a factor of $L$ , the authors argue this is justified because the total number of layers $K$ required to reach a target accuracy is significantly reduced, leading to a net gain in efficiency and faster convergence.

3. Key Contributions

Hybrid Framework: Introduction of GD-QLC, which combines the monotonic stability of Lyapunov control with the acceleration capabilities of gradient descent.
Convergence Acceleration: The method significantly reduces the number of circuit layers (depth) required to achieve high approximation ratios compared to pure feedback methods.
Robustness to Timestep ( $\Delta t$ ): Unlike FALQON, which is highly sensitive to the choice of timestep (requiring very small $\Delta t$ for stability), GD-QLC demonstrates strong robustness across a wide range of timestep values.
Control Stability: The method produces smoother control parameter trajectories ( $\beta_k$ ), avoiding the extreme spikes (e.g., values $>4000$ ) observed in second-order feedback methods (SO-FALQON) when using larger timesteps.

4. Experimental Results

The authors validated GD-QLC through classical simulations on four combinatorial optimization problems: MAX-CUT, Weighted MAX-CUT, MAX-CLIQUE, and MIN-COVER.

Performance vs. FALQON & SO-FALQON:
- GD-QLC consistently outperformed standard FALQON in both approximation ratio ( $r_A$ ) and success probability ( $p(t)$ ) across all problem sizes (10 to 20 qubits).
- In small timestep regimes ( $\Delta t = 0.01$ ), GD-QLC outperformed even the second-order method (SO-FALQON).
- In large timestep regimes ( $\Delta t = 0.1$ ), GD-QLC achieved performance comparable to SO-FALQON but with much more stable control parameters.
Robustness to $\Delta t$ :
- FALQON performance degraded significantly as $\Delta t$ increased due to the breakdown of linear approximations.
- GD-QLC maintained rapid convergence and high accuracy even with larger timesteps ( $\Delta t = 0.07$ ), reducing the need for fine-tuning.
Effect of Iterations ( $L$ ):
- Increasing the number of GD iterations per layer ( $L$ ) improved convergence speed and solution quality.
- Diminishing returns were observed beyond $L \approx 6-8$ , suggesting a practical sweet spot that balances overhead and performance.
- Higher $L$ values also helped regulate control amplitudes, preventing sharp transients in $\beta_k$ .

5. Significance and Future Directions

Practical Impact: GD-QLC offers a viable path for solving combinatorial optimization problems on near-term quantum hardware. By reducing circuit depth and mitigating the need for extremely fine timestep control, it addresses hardware limitations regarding coherence times and control precision.
Training Efficiency: It avoids the barren plateau issues common in QAOA by utilizing local gradient information within a Lyapunov framework, ensuring monotonic progress without a global optimization loop.
Future Work: The authors suggest establishing formal convergence rate guarantees, exploring second-order optimization techniques (e.g., natural gradient), and validating the method on actual quantum hardware.

In summary, the paper demonstrates that integrating gradient descent into feedback-based quantum control creates a robust, fast-converging algorithm that overcomes the depth and stability limitations of existing feedback methods while retaining their training efficiency.