Abstract
Rotolasso is a novel optimization framework designed to tune parameterized quantum circuits (PQCs) using sparsity-inducing coordinate descent. By integrating L1 regularization with the Rotosolve algorithm, Rotolasso identifies and prunes redundant quantum gates during training. This approach addresses the "barren plateau" problem and reduces the high cost of circuit executions on NISQ hardware. Empirical evaluations on MaxCut and Low-Autocorrelation Binary Sequences (LABS) demonstrate that Rotolasso produces simpler, noise-resilient circuits with fewer active parameters while maintaining or exceeding the performance of state-of-the-art optimizers.
Introduction
In the current "Noisy Intermediate-Scale Quantum" (NISQ) era, Parameterized Quantum Circuits (PQCs) are central to quantum applications in chemistry and optimization. However, optimizing these circuits faces two significant hurdles: the "barren plateau" phenomenon, where the optimization landscape becomes exponentially flat, and the slow "quantum clock speed," which makes high-iteration classical optimization (like gradient descent) prohibitively expensive. Furthermore, deeper circuits with more parameters are more susceptible to decoherence and environmental noise, which can negate any potential quantum advantage.
Methods
Rotolasso addresses these challenges by combining Coordinate Descent (CD) with L1 regularization (Lasso). Unlike gradient-based methods that update all parameters simultaneously, CD updates parameters individually. Since quantum gate outputs are periodic trigonometric functions, the global minimum for a single parameter can often be found analytically—a technique known as Rotosolve. Rotolasso adds an L1 penalty to this process, forcing unimportant parameters to zero and effectively pruning the circuit.
A key innovation in Rotolasso is the use of compressive sensing. By recognizing that the objective functions of these circuits are sparse in the frequency domain, the authors can reconstruct the optimization landscape using significantly fewer measurements. This drastically reduces the number of circuit executions required on quantum hardware, making the optimization process more efficient.
Results
The framework was evaluated on two notorious combinatorial optimization problems: MaxCut and Low-Autocorrelation Binary Sequences (LABS). LABS is a particularly complex problem from communication engineering, while MaxCut is a standard graph theory benchmark.
Key findings include:
- Enhanced Solution Quality: On the LABS problem, Rotolasso achieved higher "Merit Factors" than the state-of-the-art Rotosolve method, particularly as the number of circuit layers ($p$) increased.
- Circuit Compression: Rotolasso maintained high accuracy while keeping only 30% to 50% of the parameters active. In quantum computing, a gate with a zero-degree rotation is a gate that does not need to exist, leading to shorter, more noise-resilient circuits.
- Robustness: The method proved less sensitive to the initial "guess" of parameters compared to standard methods, helping it avoid the dreaded barren plateau problem.
- MaxCut Success: For MaxCut, Rotolasso consistently achieved a better Approximation Ratio (AR) across different graph sizes ($n=10$ to $16$).
Conclusion
Rotolasso offers a promising path toward efficient PQC optimization by exploiting the natural periodicity of quantum gates and inducing sparsity. While currently validated on classical simulators, future work will focus on testing the framework on actual noisy quantum hardware and extending the analytical updates to a broader range of universal gate sets. As quantum problems scale, managing the classical overhead of coordinate descent will be a critical area for further research.
Summary updated