Abstract
We establish that first-order methods avoid strict saddle points for almost all initializations. Our results apply to a wide variety of first-order methods, including (manifold) gradient descent, block coordinate descent, mirror descent and variants thereof. The connecting thread is that such algorithms can be studied from a dynamical systems perspective in which appropriate instantiations of the Stable Manifold Theorem allow for a global stability analysis. Thus, neither access to second-order derivative information nor randomness beyond initialization is necessary to provably avoid strict saddle points.
Similar content being viewed by others
Notes
This line of work assumes that f is a random function with a specific distribution.
For the purposes of this paper, strict saddle points include local maximizers.
The determinant is invariant under similarity transformations, so is independent of the choice of basis.
References
Absil, P.A., Mahony, R., Andrews, B.: Convergence of the iterates of descent methods for analytic cost functions. SIAM J. Optim. 16(2), 531–547 (2005)
Absil, P.A., Mahony, R., Sepulchre, R.: Optimization Algorithms on Matrix Manifolds. Princeton University Press, Princeton (2009)
Absil, P.A., Mahony, R., Trumpf, J.: An extrinsic look at the Riemannian Hessian. In: Nielsen, F., Barbaresco, F. (eds.) Geometric Science of Information, pp. 361–368. Springer, Berlin (2013)
Absil, P.A., Malick, J.: Projection-like retractions on matrix manifolds. SIAM J. Optim. 22(1), 135–158 (2012)
Adler, R.J., Taylor, J.E.: Random Fields and Geometry. Springer, Berlin (2009)
Arora, S., Ge, R., Ma, T., Moitra, A.: Simple, efficient, and neural algorithms for sparse coding. In: Proceedings of The 28th Conference on Learning Theory, pp. 113–149 (2015)
Attouch, H., Bolte, J., Redont, P., Soubeyran, A.: Proximal alternating minimization and projection methods for nonconvex problems: an approach based on the Kurdyka–Lojasiewicz inequality. Math. Oper. Res. 35(2), 438–457 (2010)
Attouch, H., Bolte, J., Svaiter, B.F.: Convergence of descent methods for semi-algebraic and tame problems: proximal algorithms, forward–backward splitting, and regularized Gauss–Seidel methods. Math. Program. 137(1–2), 91–129 (2013)
Auffinger, A., Arous, G.B., Černỳ, J.: Random matrices and complexity of spin glasses. Commun. Pure Appl. Math. 66(2), 165–201 (2013)
Beck, A.: First-Order Methods in Optimization, vol. 25. SIAM, New Delhi (2017)
Belkin, M., Rademacher, L., Voss, J.: Basis learning as an algorithmic primitive. In: Conference on Learning Theory, pp. 446–487 (2016)
Bhojanapalli, S., Neyshabur, B., Srebro, N.: Global optimality of local search for low rank matrix recovery. In: Advances in Neural Information Processing Systems, pp. 3873–3881 (2016)
Bolte, J., Daniilidis, A., Ley, O., Mazet, L., et al.: Characterizations of Lojasiewicz inequalities: subgradient flows, talweg, convexity. Trans. Am. Math. Soc 362(6), 3319–3363 (2010)
Bolte, J., Sabach, S., Teboulle, M.: Proximal alternating linearized minimization or nonconvex and nonsmooth problems. Math. Program. 146(1–2), 459–494 (2014)
Bolte, J., Sabach, S., Teboulle, M., Vaisbourd, Y.: First order methods beyond convexity and lipschitz gradient continuity with applications to quadratic inverse problems. SIAM J. Optim. 28(3), 2131–2151 (2018)
Brutzkus, A., Globerson, A.: Globally optimal gradient descent for a convnet with Gaussian inputs. arXiv preprint arXiv:1702.07966 (2017)
Cai, T.T., Li, X., Ma, Z., et al.: Optimal rates of convergence for noisy sparse phase retrieval via thresholded wirtinger flow. Ann. Stat. 44(5), 2221–2251 (2016)
Candes, E.J., Li, X., Soltanolkotabi, M.: Phase retrieval via Wirtinger flow: theory and algorithms. IEEE Trans. Inf. Theory 61(4), 1985–2007 (2015)
Choromanska, A., Henaff, M., Mathieu, M., Arous, G.B., LeCun, Y.: The loss surfaces of multilayer networks. In: Artificial Intelligence and Statistics, pp. 192–204 (2015)
Conn, A.R., Gould, N.I., Toint, P.L.: Trust Region Methods, vol. 1. SIAM, New Delhi (2000)
Dauphin, Y.N., Pascanu, R., Gulcehre, C., Cho, K., Ganguli, S., Bengio, Y.: Identifying and attacking the saddle point problem in high-dimensional non-convex optimization. In: Advances in Neural Information Processing Systems, pp. 2933–2941 (2014)
Du, S.S., Jin, C., Lee, J.D., Jordan, M.I., Poczos, B., Singh, A.: Gradient descent can take exponential time to escape saddle points. arXiv preprint arXiv:1705.10412 (2017)
Du, S.S., Lee, J.D., Tian, Y.: When is a convolutional filter easy to learn? arXiv preprint arXiv:1709.06129 (2017)
Ge, R., Huang, F., Jin, C., Yuan, Y.: Escaping from saddle points online stochastic gradient for tensor decomposition. In: Conference on Learning Theory, pp. 797–842 (2015)
Ge, R., Jin, C., Zheng, Y.: No spurious local minima in nonconvex low rank problems: a unified geometric analysis. arXiv preprint arXiv:1704.00708 (2017)
Ge, R., Lee, J.D., Ma, T.: Matrix completion has no spurious local minimum. In: Advances in Neural Information Processing Systems, pp. 2973–2981 (2016)
Gill, P.E., Murray, W.: Newton-type methods for unconstrained and linearly constrained optimization. Math. Program. 7(1), 311–350 (1974)
Jin, C., Ge, R., Netrapalli, P., Kakade, S.M., Jordan, M.I.: How to escape saddle points efficiently. arXiv preprint arXiv:1703.00887 (2017)
Keshavan, R.H., Montanari, A., Oh, S.: Matrix completion from a few entries. IEEE Trans. Inf. Theory 56(6), 2980–2998 (2009)
Kleinberg, R., Piliouras, G., Tardos, E.: Multiplicative updates outperform generic no-regret learning in congestion games. In: Proceedings of the Forty-First Annual ACM Symposium on Theory of Computing, pp. 533–542. ACM (2009)
Lange, K.: Optimization, vol. 95. Springer (2013)
Lee, J.D., Simchowitz, M., Jordan, M.I., Recht, B.: Gradient descent only converges to minimizers. In: Conference on Learning Theory, pp. 1246–1257 (2016)
Lewis, A.S., Malick, J.: Alternating projections on manifolds. Math. Oper. Res. 33(1), 216–234 (2008)
Liu, M., Yang, T.: On noisy negative curvature descent: competing with gradient descent for faster non-convex optimization. arXiv preprint arXiv:1709.08571 (2017)
Losert, V., Akin, E.: Dynamics of games and genes: discrete versus continuous time. J. Math. Biol. 17, 241–251 (1983)
Mikusinski, P., Taylor, M.: An Introduction to Multivariable Analysis From Vector to Manifold. Springer, Berlin (2012)
Moré, J.J., Sorensen, D.C.: On the use of directions of negative curvature in a modified Newton method. Math. Program. 16(1), 1–20 (1979)
Murty, K.G., Kabadi, S.N.: Some NP-complete problems in quadratic and nonlinear programming. Math. Program. 39(2), 117–129 (1987)
Nesterov, Y.: Introductory Lectures on Convex Optimization, vol. 87. Springer, Berlin (2004)
Nesterov, Y., Polyak, B.T.: Cubic regularization of Newton method and its global performance. Math. Program. 108(1), 177–205 (2006)
ONeill, M., Wright, S.: Behavior of accelerated gradient methods near critical points of nonconvex problems. arXiv preprint arXiv:1706.07993 (2017)
Panageas, I., Piliouras, G.: Gradient descent only converges to minimizers: non-isolated critical points and invariant regions. In: Innovations of Theoretical Computer Science (ITCS) (2017)
Pascanu, R., Dauphin, Y.N., Ganguli, S., Bengio, Y.: On the saddle point problem for non-convex optimization. arXiv:1405.4604 (2014)
Pemantle, R.: Nonconvergence to unstable points in urn models and stochastic approximations. Ann. Probab. 18, 698–712 (1990)
Reddi, S.J., Zaheer, M., Sra, S., Poczos, B., Bach, F., Salakhutdinov, R., Smola, A.J.: A generic approach for escaping saddle points. arXiv preprint arXiv:1709.01434 (2017)
Richtárik, P., Takáč, M.: Iteration complexity of randomized block-coordinate descent methods for minimizing a composite function. arXiv preprint arXiv:1107.2848 (2011)
Royer, C.W., Wright, S.J.: Complexity analysis of second-order line-search algorithms for smooth nonconvex optimization. arXiv preprint arXiv:1706.03131 (2017)
Shub, M.: Global Stability of Dynamical Systems. Springer, Berlin (1987)
Soltanolkotabi, M., Javanmard, A., Lee, J.D.: Theoretical insights into the optimization landscape of over-parameterized shallow neural networks. arXiv preprint arXiv:1707.04926 (2017)
Sun, J., Qu, Q., Wright, J.: When are nonconvex problems not scary? arXiv preprint arXiv:1510.06096 (2015)
Sun, J., Qu, Q., Wright, J.: A geometric analysis of phase retrieval. In: 2016 IEEE International Symposium on Information Theory (ISIT), pp. 2379–2383. IEEE (2016)
Sun, J., Qu, Q., Wright, J.: Complete dictionary recovery over the sphere I: overview and the geometric picture. IEEE Trans. Inf. Theory 63(2), 853–884 (2017)
Sun, J., Qu, Q., Wright, J.: Complete dictionary recovery over the sphere II: recovery by Riemannian trust-region method. IEEE Tran. Inf. Theory 63(2), 885–914 (2017)
Zhang, Y., Chen, X., Zhou, D., Jordan, M.I.: Spectral methods meet EM: a provably optimal algorithm for crowdsourcing. In: Advances in Neural Information Processing Systems, pp. 1260–1268 (2014)
Zhao, T., Wang, Z., Liu, H.: Nonconvex low rank matrix factorization via inexact first order oracle. Adv. Neural Inf. Process. Syst. 559–567 (2015)
Author information
Authors and Affiliations
Corresponding author
Electronic supplementary material
Below is the link to the electronic supplementary material.
Appendices
Proof of Claim 4
Proof
We assume that \(\alpha |e_j^{T}Hz_j| < \delta \left\| z_j \right\| _2\) for all \(j \in \{1,\ldots ,n\}\), for some \(\delta \) to be chosen later. For the base case \(j=2\), it holds that \(\left\| y_t-z_2 \right\| _2 = \left\| z_1-z_2 \right\| _2 = \alpha |e_1^{T}Hz_1|< \delta \left\| z_1 \right\| _2 < 2\delta \left\| y_t \right\| _2\) and \(\left\| z_2 \right\| _2 < (1+2\delta )\left\| y_t \right\| _2\). Suppose for \(j \ge 2\) that \(\left\| y_t-z_{j} \right\| _2 < 2(j-1)\delta \left\| y_t \right\| _2\) and thus \(\left\| z_j \right\| _2 < [1+2(j-1)\delta ] \left\| y_t \right\| _2.\) Using induction and triangle inequality we get
where we assume \(\delta < \frac{1}{2n}\) so that \(2(j-1)\delta <1\) for all \(j \in [n]\). Using the above calculation,
Thus \(\alpha \left\| Hy_t \right\| _2 < \sqrt{n}\delta \big ( 1+2n \delta + 2n \alpha L\big ) \left\| y_t \right\| _2\), and
where \(\sigma _{\min ^+}\) is the smallest non-zero singular value of H. Thus by choosing \(\delta \) small enough such that
we have obtained a contradiction. \(\square \)
Proof of Proposition 7
Proof
Let \(H= \nabla ^2 f(x^*)\), \(J =\mathrm {D}g(x^*) = \prod _{i=1}^b (I- \alpha P_{S_{b-i+1}} H)\), and \(y_0\) be an eigenvector of the Hessian at \(x^*\).
The proof technique is very similar to that of the proof of Proposition 5. We shall prove that \(\left\| J^t y_0 \right\| _2 \ge c(1+\eta )^t \). Hence by Gelfand’s theorem J must have at least one eigenvalue with magnitude greater than one.
We fix some arbitrary iteration t and let \(y_t = J^t y_0\). We will first show that there exists an \(\epsilon >0\),
for all \(t\in {\mathbb {N}}\). Let \(z_1 = y_t\) and \(z_{i+1} = (I- \alpha P_{S_i} H)z_i = z_i - \alpha \sum _{j \in S_i}(e_j^{T}Hz_i)e_j \), so that \(y_{t+1} = Jy_t = z_{b+1}\). We get that
Thus \(z_{i}^T H z_i \) is a decreasing (non-increasing) sequence.
We shall prove that there exists an \(i \in [b]\) so that \(z_{i+1}^{T}Hz_{i+1} \le (1+\delta ) z_{i}^{T}Hz_{i}\) for some global constant \(\delta \) to be chosen later.
Claim
Let \(y_t\) be in the range of H. There exists an \(i \in [b]\) so that \(\alpha \sum _{j \in S_i} \left| e_j^{T}Hz_i\right| \ge \delta \left\| z_i \right\| _2\) for some \(\delta >0\).
To finish the proof of the lemma, suppose that Claim B applies. Then by Cauchy–Schwarz, there exists an index i such that
However, \(w^{T}Hw \ge \lambda _{\min }(H)\left\| w \right\| ^2_2 \ge - L \left\| w \right\| ^2_2\), hence we get that
By choosing \(\epsilon =\frac{\delta ^2}{\alpha L n}\) we showed that \(y_{t+1}^{T}Hy_{t+1} \le (1+\epsilon )y_t^{T}Hy_t\) as long as \(y_t\) is in the range of H.
Assume that \(y_t = y_{{\mathcal {N}}} + y_{{\mathcal {R}}}\). It is easy to see \(y_t^{T}Hy_t = y_{{\mathcal {R}}}^{T}Hy_{{\mathcal {R}}}\) and also \(y_{t+1} = Jy_t = y_{{\mathcal {N}}} + Jy_{{\mathcal {R}}}\), hence \(y_{t+1}^{T}Hy_{t+1} =(Jy_{{\mathcal {R}}})^{T} H (Jy_{{\mathcal {R}}})\). Therefore from Inequality (14) proved above, if the starting vector is \(y_{{\mathcal {R}}}\), which Claim B applies too, then \((Jy_{{\mathcal {R}}})^{T}HJy_{{\mathcal {R}}} \le (1+\epsilon )y_{{\mathcal {R}}}^{T}H y_{{\mathcal {R}}} = (1+\epsilon )y_t^{T}Hy_t\).
To sum up, we showed that \(y_t^{T}Hy_t \le (1+\epsilon )^t y_0^{T} H y_0\) and since \(y_0\) is an eigenvector of H (of norm one) with corresponding negative eigenvalue \(\lambda \), it follows that \(y_t^{T}Hy_t \le \lambda (1+\epsilon )^t.\) Finally using \(y_t^{T}Hy_t \ge \lambda _{\min }(H)\left\| y_t \right\| ^2_2\), we get \(\left\| y_t \right\| _2 \ge (1+\epsilon )^{t/2} \frac{\lambda }{\lambda _{\min }(H)}\). Observe that \(\frac{\lambda }{\lambda _{\min }(H)}>0\) is a positive constant, \((1+\epsilon )^{t/2} \ge (1+\epsilon /4)^t\) (since \(\epsilon \le 1/2\)) and the proof follows (the parameters as claimed in the beginning will be \(c = \frac{\lambda }{\lambda _{\min }(H)}\) and \(\eta = \epsilon /4\)). \(\square \)
Proof of Claim B
We assume that \(\alpha \sum _{j \in S_i}\left| e_j^{T}Hz_i \right| < \delta \left\| z_i \right\| _2\) for all \(i \in [b]\). For base case \(i=2\), it holds that \(\left\| y_t-z_2 \right\| _2 = \left\| z_1-z_2 \right\| _2 = \alpha |\sum _{j \in S_1}e_j^{T}Hz_1|< \alpha \sum _{j \in S_1} |e_j^{T}Hz_1|< \delta \left\| z_1 \right\| _2 < 2\delta \left\| y_t \right\| _2\) and \(\left\| z_2 \right\| _2 < (1+2\delta )\left\| y_t \right\| _2\). Suppose for \(i \ge 2\) that \(\left\| y_t-z_{i} \right\| _2 < 2(i-1)\delta \left\| y_t \right\| _2\) and thus \(\left\| z_i \right\| _2 < [1+2(i-1)\delta ] \left\| y_t \right\| _2.\) Using induction and triangle inequality we obtain
where we assume \(\delta < \frac{1}{2b}\) so that \(2(i-1)\delta <1\) for all \(i \in [b]\). Using the above,
Since \(\left\| He_i \right\| _2 < \sigma _{\max }(H) \le L\), we get that \(\alpha \sum _{j \in S_i} \left\| He_j \right\| _2 < |S_i| \le n\) and we conclude
Finally, using Inequality (15) it follows that \(\alpha \left\| Hy_t \right\| _2 < 2n^2 \delta \sqrt{n} \left\| y_t \right\| _2\). Let \(w \in \text {Im}(H)\) be a vector that is orthogonal to \(\text {null}(H)\) (since H is symmetric). Then it holds that \(\left\| Hw \right\| _2 \ge \sigma _{\min ^+}(H) \left\| w \right\| _2\) where \(\sigma _{\min ^+}(H)\) denotes the smallest positive singular value of H (greater than zero). Assume that \(y_t \in \text {Im}(H)\) and we get \(\left\| Hy_t \right\| _2 < \frac{2n^2 \delta \sqrt{n}}{\alpha } \left\| y_t \right\| _2\). However, \(\left\| Hy_t \right\| _2 \ge \sigma _{\min ^+}(H) \left\| y_t \right\| _2\) thus by choosing \(\frac{2n^2 \sqrt{n}\delta }{\alpha } < \sigma _{\min ^+}(H)\) we reach a contradiction. \(\square \)
Rights and permissions
About this article
Cite this article
Lee, J.D., Panageas, I., Piliouras, G. et al. First-order methods almost always avoid strict saddle points. Math. Program. 176, 311–337 (2019). https://doi.org/10.1007/s10107-019-01374-3
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10107-019-01374-3