Beyond Newton s method Thomas P. Minka

Transcription

1 Beyond Newton s method Thomas P. Minka 2000 (revised 7/21/2017) Abstract Newton s method for optimization is equivalent to iteratively maimizing a local quadratic approimation to the objective function. But some functions are not well-approimated by a quadratic, leading to slow convergence, and some have turning points where the curvature changes sign, leading to failure. To fi this, we can use a more appropriate choice of local approimation than quadratic, based on the type of function we are optimizing. This paper demonstrates three such generalized Newton rules. Like Newton s method, they only involve the first two derivatives of the function, yet converge faster and fail less often. 1 Newton s method Newton s method is usually described in terms of root-finding, but it can also be understood as maimizing a local quadratic approimation to the objective function. Let the objective be f(). A quadratic approimation about 0 has the form g() = k + b 2 ( a)2 (1) g () = b( a) (2) g () = b (3) With three free parameters, we can make g and its first two derivatives match those of f at 0. That is, (k,b,a) are given by whose solution is k + b 2 ( 0 a) 2 = f( 0 ) (4) b( 0 a) = f ( 0 ) (5) b = f ( 0 ) (6) b = f ( 0 ) (7) a = 0 f ( 0 )/f ( 0 ) (8) k = f( 0 ) f ( 0 ) 2 2f ( 0 ) (9) 1

2 If b < 0, then the maimum of g is a, leading to the update rule new = f ()/f () (10) which is eactly Newton s method for optimization. Unfortunately, if b = f () 0, then the method fails and we must use some other optimization technique to get closer to the optimum of f. 2 A non-quadratic variation In maimum-likelihood estimation of a Dirichlet distribution, we encounter the objective f() = log Γ()ep(s) K k=1 Γ(m k) > 0 (11) This objective is conve, but Newton s method may still fail by producing a negative value. We can get a faster and more stable algorithm by using a local approimation of the form Matching f and its derivatives at 0 requires g() = k +alog()+b (12) g () = a/+b (13) g () = a/ 2 (14) a = 2 0f ( 0 ) (15) b = f ( 0 ) a/ 0 (16) Figure 1 demonstrates the high quality of this approimation compared to a quadratic. Since f () 0 for all, we know a 0. If in addition b < 0, then the maimum of g is a/b, giving the update rule 1 = 1 new + f () (17) 2 f () If it happens that b 0, then some other method must be used. Note that this update cannot be achieved by simply changing the parameterization of f and applying Newton s method to the new parameterization. For eample, if we write f() = f(1/y) and use a quadratic approimation in y, we get the Newton update or y new f (1/y)( 1/y 2 ) = y f (1/y)(1/y 4 )+f (1/y)(2/y 3 ) 1 = 1 new + f () 2 f ()+2f () (18) (19) 2

3 f() 7 f() Figure 1: Approimations to the objective (11) at = 1 (left) and = 6 (right). Here K = 3, m = [1/2,1/6,1/3], and s = 1.5. Iteration Newton y = 1/ (19) Newton y = log() (20) method (17) 0 = 6 = 6 = Figure2: Convergenceofthreeiteration schemesstartingfrom = 6ontheobjective infigure1. The proposed method reaches machine precision in 4 iterations. which is as close as you can get to (17) using reparameterization. Since must be positive, it is also tempting to reparameterize with f() = f(e y ). Applying Newton s method to y gives ( ) new f () = ep (20) f ()+f () Figure 2 compares the convergence rate of the three updates (17), (19), and (20). 3

4 3 Another variation In maimum-likelihood estimation of certain Gaussian models, we encounter the objective f() = n i=1 log(+v i ) s i +v i 0 (21) This objective is not conve and has turning points which confound Newton s method. Instead of quadratic g, consider a local approimation of the form Matching f and its derivatives at 0 requires g() = k nlog(+a) b +a (22) g () = n +a + b (+a) 2 (23) g n () = (+a) 2b 2 (+a) 3 (24) b = n( 0 +a)+f ( 0 )( 0 +a) 2 (25) { f ( 0 ) 2 nf ( 0 )+f ( 0 ) a = f ( 0 ) 0 if f ( 0 ) 0 n (26) 2f ( 0 ) 0 if f ( 0 ) = 0 Figure 3 demonstrates the high quality of this approimation compared to a quadratic. If a 0 and b > 0, the unconstrained maimum of g is b/n a, so the update rule is new = ma(0,+ f () n (+a)2 ) (27) If the conditions aren t met, which can happen when we are far from the maimum, then some other scheme must be used, just as with Newton s method. However, the method does not fail when f () crosses zero. Figure 4 compares the convergence rate of the updates (10), (20), and (27). 4

5 11 11 f() 12 f() Figure 3: Approimations to the objective (21) at = 2 (left) and = 40 (right). Here n = 3 and v = [1,1/2,1/3], s = [2,5,10]. In both cases, the proposed approimation (22) is nearly identical to the true objective. Iteration Newton (10) Newton y = log() (20) method (27) 0 = 2 = 2 = Figure 4: Convergence of three iteration schemes starting from = 2 on the objective in figure 3. Update (19) fails and is not shown. The proposed method reaches machine precision in 3 iterations. Starting from = 40, the proposed method converges equally fast, while the updates (10) and (20) fail. 5

6 4 Cauchy approimation In certain maimum-likelihood and robust estimation problems, we encounter the objective f() = n log(1+a i ( b i ) 2 ) (28) i=1 For consistency with previous sections, it is written as a maimization problem. Following the eample of the last section, consider a local approimation which has the form of one term: Matching f and its derivatives at 0 requires g() = k log(1+a( b) 2 ) (29) g () = 2a( b) 1+a( b) 2 (30) g () = 2a 1 a( b)2 (1+a( b) 2 ) 2 (31) b = 0 f ( 0 ) f ( 0 ) f ( 0 ) 2 (32) a = (f ( 0 ) f ( 0 ) 2 ) 2 f ( 0 ) 2 2f ( 0 ) Figure 5 compares this approimation with a quadratic. It doesn t fit particularly well, but it is robust. The approimation has a maimum when a > 0, i.e. 2f ( 0 ) < f ( 0 ) 2, which is a weaker constraint than f ( 0 ) < 0 required by Newton s method. When the constraint is satisfied, the maimum of g() is b, so the update rule is new = (33) f () f () f () 2 (34) This can be used as a general replacement for Newton s method, when maimizing any function. For the function in figure 5, the convergence rate is comparable to Newton when started at a point where Newton doesn t fail (points near the maimum). 6

7 f() f() Figure 5: Approimations to the objective (28) at = 2 (left) and = 0.6 (right). Here n = 3 and a = [10,20,30], b = [2,1.5,1]. 7

8 5 Multivariate eample The approach readily generalizes to multivariate functions. For eample, suppose (28) was a function of a vector : f() = n log(1+( b i ) T A i ( b i )) (35) i=1 Instead of using a univariate Cauchy-type approimation, we can use a multivariate Cauchy: The gradient and Hessian of g are: g() = k log(1+( b) T A( b)) (36) g() = g() = 2A( b) 1+( b) T A( b) (37) 2A 1+( b) T A( b) + 4A( b)( b)t A (1+( b) T A( b)) 2 (38) We set A and b by matching the gradient and Hessian of f at 0. This gives: H( 0 ) = f( 0 ) (39) b = 0 ( H( 0 ) f( 0 ) f( 0 ) T) 1 f(0 ) (40) s = f( 0 ) T H( 0 ) 1 f( 0 ) (41) A = (H( 0 ) f( 0 ) f( 0 ) T ) s 1 2 s (42) The approimation has a maimum when A is positive definite, which is a weaker constraint than H( 0 ) being negative definite as required by Newton s method. When the constraint is satisfied, the maimum of g() is b, so the update rule is new = ( H( 0 ) f( 0 ) f( 0 ) T) 1 f(0 ) (43) = H( 0 ) 1 f( 0 ) 1 f( 0 ) T H( 0 ) 1 f( 0 ) This can be used as a general replacement for Newton s method, when maimizing any multivariate function. From (44) we see that it moves in the same direction that Newton would, but with a different stepsize. This stepsize adjustment allows it to work even when Newton would fail. (44) 8