Abstract:The problem of finding a vector x which obeys a set of quadratic equations |a ⊤ k x| 2 = y k , k = 1, · · · , m, plays an important role in many applications. In this paper we consider the case when both x and a k are real-valued vectors of length n. A new loss function is constructed for this problem, which combines the smooth quadratic loss function with an activation function. Under the Gaussian measurement model, we establish that with high probability the target solution x is the unique local minimizer (u… Show more
“…Surprisingly, we are able to prove that there are only a finite number of saddles and they are all strict saddles with very high probability. Our analysis provides a rigorous explanation for the robust performance of the PGD in solving the phase retrieval problem as a low-rank optimization problem, a phenomenon that has been observed in previous applications reported in the literature [41]. Although our primary focus is the low-rank matrix manifold, the asymptotic convergence to the minimum and the escape of strict saddles (strict critical submanifolds) are also valid on any arbitrary finite dimensional Riemannian manifold.…”
supporting
confidence: 53%
“…To ensure asymptotic convergence of the PGD to the global minimizer, it remains to rule out local minimizers and identify other critical points as strict saddles. Previous works [41], [57] have shown that phase retrieval has no spurious local minimum at least with high probability in the Euclidean setting. The analysis of saddles has been more complicated because of the stochasticity and Euclidean space parameterization.…”
mentioning
confidence: 94%
“…, [41]. More importantly, they point out that the saddle points do not interfere with the convergence of various first-order algorithms towards a minimum, as long as they are``strict saddles,"" even though saddles are first-order stationary points [17], [30].…”
mentioning
confidence: 99%
“…As an application of our asymptotic escape analysis for strict saddles, we consider the phase retrieval problem [15], [19], [29], [52] that has received considerable attention in recent years. We combine the perspectives of Riemannian manifold optimization [7] and landscape analysis [41], [57] and derive new results. We analyze the saddle points of the phase retrieval problem on the low-rank matrix manifold.…”
mentioning
confidence: 99%
“…Asymptotic escape on the low-rank matrix manifold. In this section, we consider the phase retrieval problem [7], [41], [57] on the rank-1 matrix manifold. This serves both as an application of our asymptotic escape analysis for strict saddles and as a demonstration of the possibility of treating such problems rigorously on the manifold as opposed to the Euclidean space.…”
In this paper, we provide some analysis on the asymptotic escape of strict saddles in manifold optimization using the projected gradient descent algorithm (PGD). One of our main contributions is that we extend the current analysis to include nonisolated and possibly continuous saddle sets with complicated geometry. We prove that the PGD is able to escape strict critical submanifolds under certain conditions on the geometry and the distribution of the saddle point sets. We also show that the PGD may fail to escape strict saddles under weaker assumptions even if the saddle point set has zero measure. We apply this saddle analysis to the phase retrieval problem on the low-rank matrix manifold, prove that there are only a finite number of saddles, and that in a specific region, they are strict saddles with high probability. We also show the potential application of our analysis for a broader range of manifold optimization problems.
“…Surprisingly, we are able to prove that there are only a finite number of saddles and they are all strict saddles with very high probability. Our analysis provides a rigorous explanation for the robust performance of the PGD in solving the phase retrieval problem as a low-rank optimization problem, a phenomenon that has been observed in previous applications reported in the literature [41]. Although our primary focus is the low-rank matrix manifold, the asymptotic convergence to the minimum and the escape of strict saddles (strict critical submanifolds) are also valid on any arbitrary finite dimensional Riemannian manifold.…”
supporting
confidence: 53%
“…To ensure asymptotic convergence of the PGD to the global minimizer, it remains to rule out local minimizers and identify other critical points as strict saddles. Previous works [41], [57] have shown that phase retrieval has no spurious local minimum at least with high probability in the Euclidean setting. The analysis of saddles has been more complicated because of the stochasticity and Euclidean space parameterization.…”
mentioning
confidence: 94%
“…, [41]. More importantly, they point out that the saddle points do not interfere with the convergence of various first-order algorithms towards a minimum, as long as they are``strict saddles,"" even though saddles are first-order stationary points [17], [30].…”
mentioning
confidence: 99%
“…As an application of our asymptotic escape analysis for strict saddles, we consider the phase retrieval problem [15], [19], [29], [52] that has received considerable attention in recent years. We combine the perspectives of Riemannian manifold optimization [7] and landscape analysis [41], [57] and derive new results. We analyze the saddle points of the phase retrieval problem on the low-rank matrix manifold.…”
mentioning
confidence: 99%
“…Asymptotic escape on the low-rank matrix manifold. In this section, we consider the phase retrieval problem [7], [41], [57] on the rank-1 matrix manifold. This serves both as an application of our asymptotic escape analysis for strict saddles and as a demonstration of the possibility of treating such problems rigorously on the manifold as opposed to the Euclidean space.…”
In this paper, we provide some analysis on the asymptotic escape of strict saddles in manifold optimization using the projected gradient descent algorithm (PGD). One of our main contributions is that we extend the current analysis to include nonisolated and possibly continuous saddle sets with complicated geometry. We prove that the PGD is able to escape strict critical submanifolds under certain conditions on the geometry and the distribution of the saddle point sets. We also show that the PGD may fail to escape strict saddles under weaker assumptions even if the saddle point set has zero measure. We apply this saddle analysis to the phase retrieval problem on the low-rank matrix manifold, prove that there are only a finite number of saddles, and that in a specific region, they are strict saddles with high probability. We also show the potential application of our analysis for a broader range of manifold optimization problems.
We aim to find a solution x ∈ C n to a system of quadratic equations of the form b i = |a * i x| 2 , i = 1, 2, . . . , m, e.g., the well-known NP-hard phase retrieval problem. As opposed to recently proposed state-of-the-art nonconvex methods, we revert to the semidefinite relaxation (SDR) PhaseLift convex formulation and propose a successive and incremental nonconvex optimization algorithm, termed as IncrePR, to indirectly minimize the resulting convex problem on the cone of positive semidefinite matrices. Our proposed method overcomes the excessive computational cost of typical SDP solvers as well as the need of a good initialization for typical nonconvex methods. For Gaussian measurements, which is usually needed for provable convergence of nonconvex methods, restart-IncrePR solving three consecutive PhaseLift problems outperforms state-of-the-art nonconvex gradient flow based solvers with a sharper phase transition of perfect recovery and typical convex solvers in terms of computational cost and storage. For more challenging structured (non-Gaussian) measurements often occurred in real applications, such as transmission matrix and oversampling Fourier transform, IncrePR with several consecutive repeats can be used to find a good initial guess. With further refinement by local nonconvex solvers, one can achieve a better solution than that obtained by applying nonconvex gradient flow based solvers directly when the number of measurements is relatively small. Extensive numerical tests are performed to demonstrate the effectiveness of the proposed method.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.