3.2 Sources, Sinks, Saddles, and Spirals

Total Page:16

File Type:pdf, Size:1020Kb

3.2 Sources, Sinks, Saddles, and Spirals 3.2. Sources, Sinks, Saddles, and Spirals 161 3.2 Sources, Sinks, Saddles, and Spirals The pictures in this section show solutions to Ay00 By0 Cy 0. These are linear equations with constant coefficients A; B; and C .C The graphsC showD solutions y on the horizontal axis and their slopes y0 dy=dt on the vertical axis. These pairs .y.t/;y0.t// depend on time, but time is not inD the pictures. The paths show where the solution goes, but they don’t show when. Each specific solution starts at a particular point .y.0/;y0.0// given by the initial conditions. The point moves along its path as the time t moves forward from t 0. D We know that the solutions to Ay00 By0 Cy 0 depend on the two solutions to 2 C C D As Bs C 0 (an ordinary quadratic equation for s). When we find the roots s1 and C C D s2, we have found all possible solutions : s1t s2t s1t s2t y c1e c2e y c1s1e c2s2e (1) D C 0 D C The numbers s1 and s2 tell us which picture we are in. Then the numbers c1 and c2 tell us which path we are on. Since s1 and s2 determine the picture for each equation, it is essential to see the six possibilities. We write all six here in one place, to compare them. Later they will appear in six different places, one with each figure. The first three have real solutions s1 and s2. The last three have complex pairs s a i!. D ˙ Sources Sinks Saddles Spiral out Spiral in Center s1 > s2 > 0 s1 < s2 < 0 s2 <0<s1 a Re s>0 a Re s<0 a Re s 0 D D D D In addition to those six, there will be limiting cases s 0 and s1 s2 (as in resonance). D D Stability This word is important for differential equations. Do solutions decay to zero ? The solutions are controlled by es1t and es2t (and in Chapter 6 by e1t and e2t ). We can identify the two pictures (out of six) that are displaying full stability : the sinks. A center s i! is at the edge of stability (ei!t is neither decaying or growing). D ˙ 2: Sinks are stable s1 <s2 <0 Then y.t/ 0 ! 5: Spiral sinks are stable Re s1 Re s2 <0 Then y.t/ 0 D ! Special note. May I mention here that the same six pictures also apply to a system of two first order equations. Instead of y and y0, the equations have unknowns y1 and y2. Instead of the constant coefficients A;B;C , the equations will have a 2 by 2 matrix. Instead of the roots s1 and s2, that matrix will have eigenvalues 1 and 2. Those 2 eigenvalues are the roots of an equation A B C 0, just like s1 and s2. We will see the same six possibilities forC the ’s,C andD the same six pictures. The eigenvalues of the 2 by 2 matrix give the growth rates or decay rates, in place of s1 and s2. y a b y y .t/ v 10 1 has solutions 1 1 et : y D c d y2 y2.t/ D v2 20 t The eigenvalue is and the eigenvector is v .v1; v2/. The solution is y.t/ ve . D D 162 Chapter 3. Graphical and Numerical Methods The First Three Pictures We are starting with the case of real roots s1 and s2. In the equation Ay00 By0 Cy 0, this means that B2 4AC . Then B is relatively large. The square rootC in theC quadraticD formula produces a real number pB2 4AC . If A;B;C have the same sign, we have overdamping and negative roots and stability. The solutions decay to .0;0/ : a sink. If A and C have opposite sign to B as in y 3y 2y 0, we have negative damping 00 0 C D and positive roots s1; s2. The solutions grow (this is instability : a source at .0;0/). Suppose A and C have different signs, as in y 3y 2y 0. Then s1 and s2 also 00 0 D have different signs and the picture shows a saddle. The moving point .y.t/;y0.t// can start in toward .0;0/ before it turns out to infinity. The positive s gives est . 2 ! 1 Second example for a saddle : y00 4y 0 leads to s 4 .s 2/.s 2/ 0. D D 2t C 2t D The roots s1 2 and s2 2 have opposite signs. Solutions c1e c2e grow D D C unless c1 0. Only that one line with c1 0 has arrows inward. D D In every case with B2 4AC , the roots are real. The solutions y.t/ have growing exponentials or decaying exponentials. We don’t see sines and cosines and oscillation. s1t s2t The first figure shows growth : 0 < s2 < s1. Since e grows faster than e , the larger number s1 will dominate. The solution path for .y;y0/ will approach the straight s1t s1t line of slope s1. That is because the ratio of y c1s1e to y c1e is exactly s1. 0 D D If the initial condition is on the “s1 line” then the solution (y; y0) stays on that line: c2 0. If the initial condition is exactly on the “s2 line” then the solution stays on that D s1t secondary line : c1 0. You can see that if c1 0, the c1e part takes over as t . D ¤ !1 Reverse all the arrows in the left figure: Paths go in toward .0; 0/ 0<s2 <s1 s1 <s2 <0 s2 <0<s1 Source : Unstable Sink : Stable Saddle : Unstable Figure 3.6: Real roots s1 and s2. The paths of the point .y.t/;y0.t// lead out when roots are positive and lead in when roots are negative. With s2 <0<s1, the s2-line leads in but all other paths eventually go out near the s1-line : The picture shows a saddle point. 3.2. Sources, Sinks, Saddles, and Spirals 163 2 Example for a source : y00 3y0 2y 0 leads to s 3s 2 .s 2/.s 1/ 0. The roots 1 and 2 are positive. TheC solutionsD grow and e2t dominates.C D D 2 Example for a sink : y00 3y0 2y 0 leads to s 3s 2 .s 2/.s 1/ 0. The roots 2 and 1 areC negative.C TheD solutions decayC and eCt dominates.D C C D The Second Three Pictures We move to the case of complex roots s1 and s2. In the equation Ay00 By0 Cy 0, this means that B2 < 4AC . Then A and C have the same signs and BCis relativelyC smallD (underdamping). The square root in the quadratic formula (2) is an imaginary number. The exponents s1 and s2 are now a complex pair a i! : ˙ Complex roots of B pB2 4AC s1; s2 a i!: (2) As2 Bs C 0 D 2A ˙ 2A D ˙ C C D at The path of .y;y0/ spirals around the center. Because of e , the spiral goes out if a>0 : spiral source. Solutions spiral in if a<0 : spiral sink. The frequency ! controls how fast the solutions oscillate and how quickly the spirals go around .0;0/. In case a B=2A is zero (no damping), we have a center at (0, 0). The only terms Di!t i!t left in y are e and e , in other words cos !t and sin !t. Those paths are ellipses in the last part of Figure 3.7. The solutions y.t/ are periodic, because increasing t by 2=! will not change cos !t and sin !t. That circling time 2=! is the period. Reverse all the arrows in the left figure: Paths go in toward .0; 0/: a Re s>0 a Re s<0 a Re s 0 Spiral sourceD : Unstable SpiralD sink : Stable Center :D NeutraDlly stable Figure 3.7: Complex roots s1 and s2. The paths go once around .0;0/ when t increases by 2=!. The paths spiral in when A and B have the same signs and a B=2A is negative. They spiral out when a is positive. If B 0 (no damping) andD4AC > 0, we have a center. The simplest center is y sin t; y D cos t (circle) from y y 0. D 0 D 00 C D 164 Chapter 3. Graphical and Numerical Methods First Order Equations for y1 and y2 On the first page of this section, a “Special Note” mentioned another application of the same pictures. Instead of graphing the path of .y.t/;y0.t// for one second order equation, we could followthe path of .y1.t/;y2.t// for two first order equations. The two equations look like this: dy1=dt ay1 by2 First order system y 0 Ay D C (3) D dy2=dt cy1 dy2 D C The starting values y1.0/ and y2.0/ are given. The point .y1; y2/ will move along a path in one of the six figures, depending on the numbers a;b;c;d. Looking ahead, those four numbers will go into a 2 by 2 matrix A. Equation (3) will become dy=dt Ay. The symbol y in boldface stands for the vector y .y1; y2/. D D And most important for the six figures, the exponents s1 and s2 in the solution y.t/ will be the eigenvalues 1 and 2 of the matrix A.
Recommended publications
  • Math 2374: Multivariable Calculus and Vector Analysis
    Math 2374: Multivariable Calculus and Vector Analysis Part 25 Fall 2012 Extreme points Definition If f : U ⊂ Rn ! R is a scalar function, • a point x0 2 U is called a local minimum of f if there is a neighborhood V of x0 such that for all points x 2 V, f (x) ≥ f (x0), • a point x0 2 U is called a local maximum of f if there is a neighborhood V of x0 such that for all points x 2 V, f (x) ≤ f (x0), • a point x0 2 U is called a local, or relative, extremum of f if it is either a local minimum or a local maximum, • a point x0 2 U is called a critical point of f if either f is not differentiable at x0 or if rf (x0) = Df (x0) = 0, • a critical point that is not a local extremum is called a saddle point. First Derivative Test for Local Extremum Theorem If U ⊂ Rn is open, the function f : U ⊂ Rn ! R is differentiable, and x0 2 U is a local extremum, then Df (x0) = 0; that is, x0 is a critical point. Proof. Suppose that f achieves a local maximum at x0, then for all n h 2 R , the function g(t) = f (x0 + th) has a local maximum at t = 0. Thus from one-variable calculus g0(0) = 0. By chain rule 0 g (0) = [Df (x0)] h = 0 8h So Df (x0) = 0. Same proof for a local minimum. Examples Ex-1 Find the maxima and minima of the function f (x; y) = x2 + y 2.
    [Show full text]
  • A Stability Boundary Based Method for Finding Saddle Points on Potential Energy Surfaces
    JOURNAL OF COMPUTATIONAL BIOLOGY Volume 13, Number 3, 2006 © Mary Ann Liebert, Inc. Pp. 745–766 A Stability Boundary Based Method for Finding Saddle Points on Potential Energy Surfaces CHANDAN K. REDDY and HSIAO-DONG CHIANG ABSTRACT The task of finding saddle points on potential energy surfaces plays a crucial role in under- standing the dynamics of a micromolecule as well as in studying the folding pathways of macromolecules like proteins. The problem of finding the saddle points on a high dimen- sional potential energy surface is transformed into the problem of finding decomposition points of its corresponding nonlinear dynamical system. This paper introduces a new method based on TRUST-TECH (TRansformation Under STability reTained Equilibria CHaracter- ization) to compute saddle points on potential energy surfaces using stability boundaries. Our method explores the dynamic and geometric characteristics of stability boundaries of a nonlinear dynamical system. A novel trajectory adjustment procedure is used to trace the stability boundary. Our method was successful in finding the saddle points on different potential energy surfaces of various dimensions. A simplified version of the algorithm has also been used to find the saddle points of symmetric systems with the help of some analyt- ical knowledge. The main advantages and effectiveness of the method are clearly illustrated with some examples. Promising results of our method are shown on various problems with varied degrees of freedom. Key words: potential energy surfaces, saddle points, stability boundary, minimum gradient point, computational chemistry. I. INTRODUCTION ecently, there has been a lot of interest across various disciplines to understand a wide Rvariety of problems related to bioinformatics.
    [Show full text]
  • Lecture 8: Maxima and Minima
    Maxima and minima Lecture 8: Maxima and Minima Rafikul Alam Department of Mathematics IIT Guwahati Rafikul Alam IITG: MA-102 (2013) Maxima and minima Local extremum of f : Rn ! R Let f : U ⊂ Rn ! R be continuous, where U is open. Then • f has a local maximum at p if there exists r > 0 such that f (x) ≤ f (p) for x 2 B(p; r): • f has a local minimum at a if there exists > 0 such that f (x) ≥ f (p) for x 2 B(p; ): A local maximum or a local minimum is called a local extremum. Rafikul Alam IITG: MA-102 (2013) Maxima and minima Figure: Local extremum of z = f (x; y) Rafikul Alam IITG: MA-102 (2013) Maxima and minima Necessary condition for extremum of Rn ! R Critical point: A point p 2 U is a critical point of f if rf (p) = 0: Thus, when f is differentiable, the tangent plane to z = f (x) at (p; f (p)) is horizontal. Theorem: Suppose that f has a local extremum at p and that rf (p) exists. Then p is a critical point of f , i.e, rf (p) = 0: 2 2 Example: Consider f (x; y) = x − y : Then fx = 2x = 0 and fy = −2y = 0 show that (0; 0) is the only critical point of f : But (0; 0) is not a local extremum of f : Rafikul Alam IITG: MA-102 (2013) Maxima and minima Saddle point Saddle point: A critical point of f that is not a local extremum is called a saddle point of f : Examples: 2 2 • The point (0; 0) is a saddle point of f (x; y) = x − y : 2 2 2 • Consider f (x; y) = x y + y x: Then fx = 2xy + y = 0 2 and fy = 2xy + x = 0 show that (0; 0) is the only critical point of f : But (0; 0) is a saddle point.
    [Show full text]
  • Calculus and Differential Equations II
    Calculus and Differential Equations II MATH 250 B Linear systems of differential equations Linear systems of differential equations Calculus and Differential Equations II Second order autonomous linear systems We are mostly interested with2 × 2 first order autonomous systems of the form x0 = a x + b y y 0 = c x + d y where x and y are functions of t and a, b, c, and d are real constants. Such a system may be re-written in matrix form as d x x a b = M ; M = : dt y y c d The purpose of this section is to classify the dynamics of the solutions of the above system, in terms of the properties of the matrix M. Linear systems of differential equations Calculus and Differential Equations II Existence and uniqueness (general statement) Consider a linear system of the form dY = M(t)Y + F (t); dt where Y and F (t) are n × 1 column vectors, and M(t) is an n × n matrix whose entries may depend on t. Existence and uniqueness theorem: If the entries of the matrix M(t) and of the vector F (t) are continuous on some open interval I containing t0, then the initial value problem dY = M(t)Y + F (t); Y (t ) = Y dt 0 0 has a unique solution on I . In particular, this means that trajectories in the phase space do not cross. Linear systems of differential equations Calculus and Differential Equations II General solution The general solution to Y 0 = M(t)Y + F (t) reads Y (t) = C1 Y1(t) + C2 Y2(t) + ··· + Cn Yn(t) + Yp(t); = U(t) C + Yp(t); where 0 Yp(t) is a particular solution to Y = M(t)Y + F (t).
    [Show full text]
  • Maxima, Minima, and Saddle Points
    Real Analysis III (MAT312β) Department of Mathematics University of Ruhuna A.W.L. Pubudu Thilan Department of Mathematics University of Ruhuna | Real Analysis III(MAT312β) 1/80 Chapter 12 Maxima, Minima, and Saddle Points Department of Mathematics University of Ruhuna | Real Analysis III(MAT312β) 2/80 Introduction A scientist or engineer will be interested in the ups and downs of a function, its maximum and minimum values, its turning points. For instance, locating extreme values is the basic objective of optimization. In the simplest case, an optimization problem consists of maximizing or minimizing a real function by systematically choosing input values from within an allowed set and computing the value of the function. Department of Mathematics University of Ruhuna | Real Analysis III(MAT312β) 3/80 Introduction Cont... Drawing a graph of a function using a computer graph plotting package will reveal behavior of the function. But if we want to know the precise location of maximum and minimum points, we need to turn to algebra and differential calculus. In this Chapter we look at how we can find maximum and minimum points in this way. Department of Mathematics University of Ruhuna | Real Analysis III(MAT312β) 4/80 Chapter 12 Section 12.1 Single Variable Functions Department of Mathematics University of Ruhuna | Real Analysis III(MAT312β) 5/80 Local maximum and local minimum The local maximum and local minimum (plural: maxima and minima) of a function, are the largest and smallest value that the function takes at a point within a given interval. It may not be the minimum or maximum for the whole function, but locally it is.
    [Show full text]
  • Analysis of Asymptotic Escape of Strict Saddle Sets in Manifold Optimization
    Analysis of Asymptotic Escape of Strict Saddle Sets in Manifold Optimization Thomas Y. Hou∗ , Zhenzhen Li y , and Ziyun Zhang z Abstract. In this paper, we provide some analysis on the asymptotic escape of strict saddles in manifold optimization using the projected gradient descent (PGD) algorithm. One of our main contributions is that we extend the current analysis to include non-isolated 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 and there is a uniform escape direction. We provide a counterexample to illustrate this important point. 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 they are strict saddles with high probability. We also show the potential application of our analysis for a broader range of manifold optimization problems. Key words. Manifold optimization, projected gradient descent, strict saddles, stable manifold theorem, phase retrieval AMS subject classifications. 58D17, 37D10, 65F10, 90C26 1. Introduction. Manifold optimization has long been studied in mathematics. From signal and imaging science [38][7][36], to computer vision [47][48] and quantum informa- tion [40][16][29], it finds applications in various disciplines.
    [Show full text]
  • Instability Analysis of Saddle Points by a Local Minimax Method
    INSTABILITY ANALYSIS OF SADDLE POINTS BY A LOCAL MINIMAX METHOD JIANXIN ZHOU Abstract. The objective of this work is to develop some tools for local insta- bility analysis of multiple critical points, which can be computationally carried out. The Morse index can be used to measure local instability of a nondegen- erate saddle point. However, it is very expensive to compute numerically and is ineffective for degenerate critical points. A local (weak) linking index can also be defined to measure local instability of a (degenerate) saddle point. But it is still too difficult to compute. In this paper, a local instability index, called a local minimax index, is defined by using a local minimax method. This new instability index is known beforehand and can help in finding a saddle point numerically. Relations between the local minimax index and other local insta- bility indices are established. Those relations also provide ways to numerically compute the Morse, local linking indices. In particular, the local minimax index can be used to define a local instability index of a saddle point relative to a reference (trivial) critical point even in a Banach space while others failed to do so. 1. Introduction Multiple solutions with different performance, maneuverability and instability in- dices exist in many nonlinear problems in natural or social sciences [1,7, 9,24,29,31, 33,35]. When cases are variational, the problems reduce to solving the Euler- Lagrange equation (1.1) J 0(u) = 0; where J is a C1 generic energy functional on a Banach space H and J 0 its Frechet derivative.
    [Show full text]
  • Gradient Descent Can Take Exponential Time to Escape Saddle Points
    Gradient Descent Can Take Exponential Time to Escape Saddle Points Simon S. Du Chi Jin Carnegie Mellon University University of California, Berkeley [email protected] [email protected] Jason D. Lee Michael I. Jordan University of Southern California University of California, Berkeley [email protected] [email protected] Barnabás Póczos Aarti Singh Carnegie Mellon University Carnegie Mellon University [email protected] [email protected] Abstract Although gradient descent (GD) almost always escapes saddle points asymptot- ically [Lee et al., 2016], this paper shows that even with fairly natural random initialization schemes and non-pathological functions, GD can be significantly slowed down by saddle points, taking exponential time to escape. On the other hand, gradient descent with perturbations [Ge et al., 2015, Jin et al., 2017] is not slowed down by saddle points—it canfind an approximate local minimizer in polynomial time. This result implies that GD is inherently slower than perturbed GD, and justifies the importance of adding perturbations for efficient non-convex optimization. While our focus is theoretical, we also present experiments that illustrate our theoreticalfindings. 1 Introduction Gradient Descent (GD) and its myriad variants provide the core optimization methodology in machine learning problems. Given a functionf(x), the basic GD method can be written as: x(t+1) x (t) η f x(t) , (1) ← − ∇ where η is a step size, assumedfixed in the current paper.� While� precise characterizations of the rate of convergence GD are available for convex problems, there is far less understanding of GD for non-convex problems. Indeed, for general non-convex problems, GD is only known tofind a stationary point (i.e., a point where the gradient equals zero) in polynomial time [Nesterov, 2013].
    [Show full text]
  • Escaping from Saddle Points on Riemannian Manifolds
    Escaping from saddle points on Riemannian manifolds Yue Suny, Nicolas Flammarionz, Maryam Fazely y Department of Electrical and Computer Engineering, University of Washington, Seattle z School of Computer and Communication Sciences, EPFL Abstract Taylor series and smoothness assumptions Example – Burer-Monteiro factorization. Notation: Expx(y) denotes the exponential map, gradf(x) and H(x) d×d Let A 2 S , the problem 6 We consider minimizing a nonconvex, smooth function f(x) on a are Riemannian gradient and Hessian of f(x); x is a saddle point, and smooth manifold x 2 M. We show that a perturbed Riemannian −1 max trace(AX); 5 u = Exp (u). 4 x X2 d×d gradient algorithm converges to a second-order stationary point in a S 3 • Riemannian gradient descent. Let f have a β-Lipschitz gradient. Function value number of iterations that is polynomial in appropriate smoothness s:t: diag(X) = 1;X 0; rank(X) ≤ r: 2 q There exists η = Θ(1/β) such that Riemannian gradient descent step 1 parameters of f and M, and polylog in dimension. This matches the can be factorized as + + 0 0 5 10 15 20 25 30 35 40 45 best known rate for unconstrained smooth minimization. u = Exp (−ηgradf(u)) (c.f. Euclidean case: u = u − ηrf(u)) Iterations u max trace(AY Y T ); s:t: diag(YY T ) = 1: η Iteration versus function value. The iterations 2 Y 2 d×p monotonically decreases f by 2kgradf(u)k . R start from a saddle point, is perturbed by the Background and motivation ^ when r(r + 1)=2 ≤ d, p(p + 1)=2 ≥ d.
    [Show full text]
  • Escaping from Saddle Points on Riemannian Manifolds
    Escaping from saddle points on Riemannian manifolds Yue Suny, Nicolas Flammarionz, Maryam Fazely y Department of Electrical and Computer Engineering, University of Washington, Seattle z School of Computer and Communication Sciences, EPFL, Lausanne, Switzerland November 8, 2019 1 / 24 Manifold constrained optimization We consider the problem minimize f (x); subject to x 2 M x Same as Euclidean space, generally we cannot find global optimum in polynomial time, so we want to find an approximate local minimum. Plot of saddle point in Euclidean Contour of function value on space. sphere. 2 / 24 Example of manifolds d Pd 2 2 1. Sphere. fx 2 R : i=1 xi = r g. m×n T 2. Stiefel manifold. fX 2 R : X X = I g. 3. Grassmannian manifold. Grass(p; n) is set of p dimensional n subspaces in R . m×n T 4. Bruer-Monteiro relaxation. fX 2 R : diag(X X ) = 1g. 3 / 24 Curve A curve in a continuous map γ : t !M. Usually t 2 [0; 1], where γ(0) and γ(1) are start and end points of the curve. 4 / 24 Tangent vector and tangent space We use x 2 M can be start point of many γ(t + τ) − γ(t) γ_ (t) = lim curves, and a tangent space Tx M is τ!0 τ the set of tangent vectors at x. as the velocity of the curve, γ_ (t) is a Tangent space is a metric space. tangent vector at γ(t) 2 M. 5 / 24 Gradient of a function Let f : M! R be a function defined on M, and γ be a curve on M.
    [Show full text]
  • Qualitative Analysis of Differential Equations
    QUALITATIVE ANALYSIS OF DIFFERENTIAL EQUATIONS Alexander Panfilov arXiv:1803.05291v1 [math.GM] 10 Mar 2018 det A D=0 stable spiral non−stable spiral 6 3 5 4 2 stable center non−stable node node tr A 1 saddle 2 QUALITATIVE ANALYSIS OF DIFFERENTIAL EQUATIONS Alexander Panfilov Theoretical Biology, Utrecht University, Utrecht c 2010 2010 2 Contents 1 Preliminaries 5 1.1 Basic algebra . 5 1.1.1 Algebraic expressions . 5 1.1.2 Limits . 6 1.1.3 Equations . 7 1.1.4 Systems of equations . 8 1.2 Functions of one variable . 9 1.3 Graphs of functions of one variable . 11 1.4 Implicit function graphs . 18 1.5 Exercises . 19 2 Selected topics of calculus 23 2.1 Complex numbers . 23 2.2 Matrices . 25 2.3 Eigenvalues and eigenvectors . 27 2.4 Functions of two variables . 32 2.5 Exercises . 35 3 Differential equations of one variable 37 3.1 Differential equations of one variable and their solutions . 37 3.1.1 Definitions . 37 3.1.2 Solution of a differential equation . 39 3.2 Qualitative methods of analysis of differential equations of one variable . 42 3.2.1 Phase portrait . 42 3.2.2 Equilibria, stability, global plan . 43 3.3 Systems with parameters. Bifurcations. 46 3.4 Exercises . 48 4 System of two linear differential equations 51 4.1 Phase portraits and equilibria . 51 4.2 General solution of linear system . 53 4.3 Real eigen values. Saddle, node. 56 4.3.1 Saddle; l1 < 0;l2 > 0, or l1 > 0;l2 < 0 ....................
    [Show full text]
  • How to Escape Saddle Points Efficiently
    How to Escape Saddle Points Efficiently Chi Jin 1 Rong Ge 2 Praneeth Netrapalli 3 Sham M. Kakade 4 Michael I. Jordan 1 Abstract a point with small gradient is independent of the dimension This paper shows that a perturbed form of gradi- (“dimension-free”). More precisely, for a function that is `- ent descent converges to a second-order station- gradient Lipschitz (see Definition1), it is well known that gradient descent finds an -first-order stationary point (i.e., ary point in a number iterations which depends ? 2 only poly-logarithmically on dimension (i.e., it is a point x with krf(x)k ≤ ) within `(f(x0) − f )/ it- erations (Nesterov, 1998), where x0 is the initial point and almost “dimension-free”). The convergence rate ? of this procedure matches the well-known con- f is the optimal value of f. This bound does not depend vergence rate of gradient descent to first-order on the dimension of x. In convex optimization, finding an stationary points, up to log factors. When all sad- -first-order stationary point is equivalent to finding an ap- dle points are non-degenerate, all second-order proximate global optimum. stationary points are local minima, and our result In non-convex settings, however, convergence to first-order thus shows that perturbed gradient descent can stationary points is not satisfactory. For non-convex func- escape saddle points almost for free. Our results tions, first-order stationary points can be global minima, can be directly applied to many machine learning local minima, saddle points or even local maxima. Find- applications, including deep learning.
    [Show full text]