Abstract
While there has been a significant amount of work studying gradient descent techniques for non-convex optimization problems over the last few years, all existing results establish either local convergence with good rates or global convergence with highly suboptimal rates, for many problems of interest. In this paper, we take the first step in getting the best of both worlds -- establishing global convergence and obtaining a good rate of convergence for the problem of computing squareroot of a positive definite (PD) matrix, which is a widely studied problem in numerical linear algebra with applications in machine learning and statistics among others. Given a PD matrix $M$ and a PD starting point $U_0$, we show that gradient descent with appropriately chosen step-size finds an $\epsilon$-accurate squareroot of $M$ in $O(\alpha \log (\|M-U_0^2\|_F /\epsilon))$ iterations, where $\alpha = (\max\{\|U_0\|_2^2,\|M\|_2\} / \min \{\sigma_{\min}^2(U_0),\sigma_{\min}(M) \} )^{3/2}$. Our result is the first to establish global convergence for this problem and that it is robust to errors in each iteration. A key contribution of our work is the general proof technique which we believe should further excite research in understanding deterministic and stochastic variants of simple non-convex gradient descent algorithms with good global convergence rates for other problems in machine learning and numerical linear algebra.
Introduction
In machine learning and numerical linear algebra, non-convex optimization problems often present a significant challenge due to landscapes filled with saddle points and local minima. A fundamental problem in this domain is computing the matrix square root: given a positive definite (PD) matrix $M$, find a matrix $U$ such that $U^2 = M$. This is typically formulated as minimizing the non-convex objective function $f(U) = \|M - U^2\|_F^2$.
Prior to this work, researchers faced a trade-off between local results (fast convergence but requiring near-perfect initialization) and global results (guaranteed convergence but at practically unusable rates). This paper bridges that gap, proving that standard gradient descent can achieve both global convergence and a geometric rate of convergence.
Methodology
The authors analyze a standard Gradient Descent (GD) algorithm for the matrix square root problem. The update rule is computationally lean, requiring only matrix multiplications in each step.
The core of the theoretical contribution lies in a novel proof technique. To establish global convergence, the authors demonstrate that if the algorithm is initialized with a positive definite matrix (such as the identity matrix), the iterates $U_t$ remain within a "well-behaved" region. Within this region, the objective function satisfies two critical properties:
- Smoothness: The gradient does not change abruptly.
- Gradient Dominance (Polyak-Łojasiewicz condition): The gradient remains large enough to push the algorithm toward the global minimum, effectively avoiding stagnation near saddle points.
Key Insights and Results
The analysis reveals that the matrix square root objective has a clean landscape with no spurious local minima; every stationary point is either the global minimum or a saddle point. By proving that GD naturally stays away from rank-degenerate surfaces where saddle points reside, the authors establish geometric convergence, where error drops exponentially fast.
As shown in the comparison table above, the proposed GD method offers several advantages:
- Convergence Rate: It achieves an $\epsilon$-accurate solution in $O(\alpha \log(1/\epsilon))$ iterations.
- Efficiency: Each iteration is $O(n^\omega)$, where $\omega$ is the matrix multiplication exponent.
- Provable Robustness: Unlike many alternatives, the algorithm is stable under numerical noise or rounding errors (Theorem 3.4).
Conclusion
This work provides the first global convergence guarantee for gradient descent on the matrix square root problem with a fast rate. While the convergence rate currently depends on the condition number $\kappa^{3/2}$, the authors suggest this might be improvable to a linear dependence. More importantly, the "stay in a nice region" proof strategy serves as a potential blueprint for analyzing other non-convex problems in machine learning, such as matrix factorization and sensing.
Summary updated