Numerical Methods I

Solving Nonlinear Equations

Aleksandar Donev
Courant Institute, NYU1
[email protected]
1 Course G63.2010.001 / G22.2420-001, Fall 2010

October 14th, 2010

A. Donev (Courant Institute) Lecture VI 10/14/2010 1 / 31

Outline

1 Basics of Nonlinear Solvers

2 One Dimensional Root Finding

3 Systems of Non-Linear Equations

4 Intro to Unconstrained Optimization

5 Conclusions

A. Donev (Courant Institute) Lecture VI 10/14/2010 2 / 31

Final Presentations

The final project writeup will be due Sunday Dec. 26th by midnight (I
have to start grading by 12/27 due to University deadlines).
You will also need to give a 15 minute presentation in front of me and
other students.
Our last class is officially scheduled for Tuesday 12/14, 5-7pm, and
the final exam Thursday 12/23, 5-7pm. Neither of these are good!
By the end of next week, October 23rd, please let me know the
following:
Are you willing to present early Thursday December 16th during usual
class time?
Do you want to present during the official scheduled last class,
Thursday 12/23, 5-7pm.
If neither of the above, tell me when you cannot present Monday Dec.
20th to Thursday Dec. 23rd (finals week).

A. Donev (Courant Institute) Lecture VI 10/14/2010 3 / 31

 Basics of Nonlinear Solvers

Fundamentals

Simplest problem: Root finding in one dimension:

f (x) = 0 with x [a, b]

Or more generally, solving a square system of nonlinear equations

f(x) = 0 fi (x1 , x2 , . . . , xn ) = 0 for i = 1, . . . , n.

There can be no closed-form answer, so just as for eigenvalues, we

need iterative methods.
Most generally, starting from m 1 initial guesses x 0 , x 1 , . . . , x m ,
iterate:
x k+1 = (x k , x k1 , . . . , x km ).

A. Donev (Courant Institute) Lecture VI 10/14/2010 4 / 31

 Basics of Nonlinear Solvers

Order of convergence

Consider one dimensional root finding and let the actual root be ,
f () = 0.
A sequence of iterates x k that converges to has order of
convergence p > 1 if as k
 k+1   k+1 
x  e 
k p = k p C = const,
|x | |e |

where the constant 0 < C < 1 is the convergence factor.

A method should at least converge linearly, that is, the error should
at least be reduced by a constant factor every iteration, for example,
the number of accurate digits increases by 1 every iteration.
A good method for root finding coverges quadratically, that is, the
number of accurate digits doubles every iteration!

A. Donev (Courant Institute) Lecture VI 10/14/2010 5 / 31

 Basics of Nonlinear Solvers

Local vs. global convergence

A good initial guess is extremely important in nonlinear solvers!

Assume we are looking for a unique root a b starting with an
initial guess a x0 b.
A method has local convergence if it converges to a given root for
any initial guess that is sufficiently close to (in the neighborhood
of a root).
A method has global convergence if it converges to the root for any
initial guess.
General rule: Global convergence requires a slower (careful) method
but is safer.
It is best to combine a global method to first find a good initial guess
close to and then use a faster local method.

A. Donev (Courant Institute) Lecture VI 10/14/2010 6 / 31

 Basics of Nonlinear Solvers

Conditioning of root finding

f ( + ) f () + f 0 () = f

|f | 1
abs = f 0 () .

||
|f 0 ()|

The problem of finding a simple root is well-conditioned when |f 0 ()|

is far from zero.
Finding roots with multiplicity m > 1 is ill-conditioned:
 1/m
 0   
f () = = f (m1) () = 0 |f |
|| m
|f ()|

Note that finding roots of algebraic equations (polynomials) is a

separate subject of its own that we skip.

A. Donev (Courant Institute) Lecture VI 10/14/2010 7 / 31

 One Dimensional Root Finding

The bisection and Newton algorithms

A. Donev (Courant Institute) Lecture VI 10/14/2010 8 / 31

 One Dimensional Root Finding

Bisection
First step is to locate a root by searching for a sign change, i.e.,
finding a0 and b0 such that
f (a0 )f (b0 ) < 0.
The simply bisect the interval,
ak + b k
xx =
2
and choose the half in which the function changes sign by looking at
the sign of f (x k ).
Observe that each step we need one function evaluation, f (x k ), but
only the sign matters.
The convergence is essentially linear because
 k+1 
k x 
x  b
 k 
2.
2k+1 |x k |
A. Donev (Courant Institute) Lecture VI 10/14/2010 9 / 31
 One Dimensional Root Finding

Newtons Method

Bisection is a slow but sure method. It uses no information about the

value of the function or its derivatives.

Better convergence, of order p = (1 + 5)/2 1.63 (the golden
ratio), can be achieved by using the value of the function at two
points, as in the secant method.
Achieving second-order convergence requires also evaluating the
function derivative.
Linearize the function around the current guess using Taylor series:

f (x k+1 ) f (x k ) + (x k+1 x k )f 0 (x k ) = 0

f (x k )
x k+1 = x k
f 0 (x k )

A. Donev (Courant Institute) Lecture VI 10/14/2010 10 / 31

 One Dimensional Root Finding

Convergence of Newtons method

Taylor series with remainder:
1
f () = 0 = f (x k )+(x k )f 0 (x k )+ (x k )2 f 00 () = 0, for some [xn , ]
2
After dividing by f 0 (x k ) 6= 0 we get

f (x k ) f 00 ()
 
1
k
x 0 k = ( x k )2 0 k
f (x ) 2 f (x )

1 k
2 f 00 ()
x k+1 = e k+1 = e
2 f 0 (x k )
which shows second-order convergence
 k+1   k+1  
x  e   f 00 ()   00
 f () 

2
= 2
= 0 k  0
   
|x k | |e k | 2f (x ) 2f () 

A. Donev (Courant Institute) Lecture VI 10/14/2010 11 / 31

 One Dimensional Root Finding

Proof of Local Convergence

 k+1   00  00

x  
 f ()   f (x) 
2
=  0 k  M = sup 
 0

|x k | 2f (x ) |e 0 |x,y +|e 0 | 2f (y )


2
2
M x k+1  = E k+1 M x k  = E k
  

which will converge if E 0 < 1, i.e., if

x  = e 0  < M 1
 0   

Newtons method thus always converges quadratically if we start

sufficiently close to a simple root.

A. Donev (Courant Institute) Lecture VI 10/14/2010 12 / 31

 One Dimensional Root Finding

Fixed-Point Iteration

Another way to devise iterative root finding is to rewrite f (x) in an

equivalent form
x = (x)
Then we can use fixed-point iteration

x k+1 = (x k )

whose fixed point (limit), if it converges, is x .

For example, recall from first lecture solving x 2 = c via the
Babylonian method for square roots
1 c 
xn+1 = (xn ) = +x ,
2 x
which converges (quadratically) for any non-zero initial guess.

A. Donev (Courant Institute) Lecture VI 10/14/2010 13 / 31

 One Dimensional Root Finding

Convergence theory

It can be proven that the fixed-point iteration x k+1 = (x k )

converges if (x) is a contraction mapping:
 0 
 (x) K < 1 x [a, b]

x k+1 = (x k )() = 0 () x k by the mean value theorem

 k+1
 < K x k 
  
x

If 0 () 6= 0 near the root we have linear convergence

 k+1 
x 
0 ().
|x k |

If 0 () = 0 we have second-order convergence if 00 () 6= 0, etc.

A. Donev (Courant Institute) Lecture VI 10/14/2010 14 / 31

 One Dimensional Root Finding

Applications of general convergence theory

Think of Newtons method

f (x k )
x k+1 = x k
f 0 (x k )

as a fixed-point iteration method x k+1 = (x k ) with iteration

function:
f (x)
(x) = x 0 .
f (x)
We can directly show quadratic convergence because (also see
homework)
f (x)f 00 (x)
0 (x) = 0 () = 0
[f 0 (x)]2
f 00 ()
00 () = 6= 0
f 0 ()

A. Donev (Courant Institute) Lecture VI 10/14/2010 15 / 31

 One Dimensional Root Finding

Stopping Criteria
A good library function for root finding has to implement careful
termination criteria.
An obvious option is to terminate when the residual becomes small
 k 
f (x ) < ,

which is only good for very well-conditioned problems, |f 0 ()| 1.

Another option is to terminate when the increment becomes small
 k+1
x k  < .

x

For fixed-point iteration

x k+1 x k = e k+1 e k 1 0 () e k
   k
e  ,
[1 0 ()]
so we see that the increment test works for rapidly converging
iterations (0 ()  1).
A. Donev (Courant Institute) Lecture VI 10/14/2010 16 / 31
 One Dimensional Root Finding

In practice

A robust but fast algorithm for root finding would combine bisection
with Newtons method.
Specifically, a method like Newtons that can easily take huge steps in
the wrong direction and lead far from the current point must be
safeguarded by a method that ensures one does not leave the search
interval and that the zero is not missed.
Once x k is close to , the safeguard will not be used and quadratic or
faster convergence will be achieved.
Newtons method requires first-order derivatives so often other
methods are preferred that require function evaluation only.
Matlabs function fzero combines bisection, secant and inverse
quadratic interpolation and is fail-safe.

A. Donev (Courant Institute) Lecture VI 10/14/2010 17 / 31

 One Dimensional Root Finding

Find zeros of a sin(x) + b exp(x 2 /2) in MATLAB

% f=@ m f i l e u s e s a f u n c t i o n i n an m f i l e

% Parameterized f u n c t i o n s are created with :

a = 1; b = 2;
f = @( x ) a s i n ( x ) + b exp(x 2 / 2 ) ; % Handle

figure (1)
ezplot ( f ,[ 5 ,5]); grid

x1=f z e r o ( f , [ 2 , 0 ] )
[ x2 , f 2 ]= f z e r o ( f , 2 . 0 )

x1 = 1.227430849357917
x2 = 3.155366415494801
f2 = 2.116362640691705 e 16

A. Donev (Courant Institute) Lecture VI 10/14/2010 18 / 31

 One Dimensional Root Finding

Figure of f (x)

a sin(x)+b exp(x2/2)
2.5

1.5

0.5

0.5

5 4 3 2 1 0 1 2 3 4 5
x

A. Donev (Courant Institute) Lecture VI 10/14/2010 19 / 31

 Systems of Non-Linear Equations

Multi-Variable Taylor Expansion

We are after solving a square system of nonlinear equations for
some variables x:
f(x) = 0 fi (x1 , x2 , . . . , xn ) = 0 for i = 1, . . . , n.
It is convenient to focus on one of the equations, i.e., consider a
scalar function f (x).
The usual Taylor series is replaced by
1
f (x + x) = f (x) + gT (x) + (x)T H (x)
2
where the gradient vector is
f T
 
f f
g = x f = , , ,
x1 x2 xn
and the Hessian matrix is
2f
 
H= 2x f =
xi xj ij

A. Donev (Courant Institute) Lecture VI 10/14/2010 20 / 31

 Systems of Non-Linear Equations

Newtons Method for Systems of Equations

It is much harder if not impossible to do globally convergent methods
like bisection in higher dimensions!
A good initial guess is therefore a must when solving systems, and
Newtons method can be used to refine the guess.
The first-order Taylor series is
f xk + x f xk + J xk x = 0





where the Jacobian J has the gradients of fi (x) as rows:

fi
[J (x)]ij =
xj
So taking a Newton step requires solving a linear system:
J xk x = f xk but denote J J xk



xk+1 = xk + x = xk J1 f xk .

A. Donev (Courant Institute) Lecture VI 10/14/2010 21 / 31

 Systems of Non-Linear Equations

Convergence of Newtons method

Newtons method converges quadratically if started sufficiently close

to a root x? at which the Jacobian is not singular.

x? = ek+1 = xk J1 f xk x? = ek J1 f xk
k+1


x

but using second-order Taylor series

 
1 k
T
J1 f xk J1 f (x? ) + Jek + H ek



e
2
1
J
T
= ek + ek H ek


2

k+1 J1 k
T
J1 kHk 2

H ek ek



e =
2 e 2
Fixed point iteration theory generalizes to multiple variables, e.g.,
replace f 0 () < 1 with (J(x? )) < 1.
A. Donev (Courant Institute) Lecture VI 10/14/2010 22 / 31
 Systems of Non-Linear Equations

Problems with Newtons method

Newtons method requires solving many linear systems, which can

become complicated when there are many variables.
It also requires computing a whole matrix of derivatives, which can
be expensive or hard to do (differentiation by hand?)
Newtons method converges fast if the Jacobian J (x? ) is
well-conditioned, otherwise it can blow up.
For large systems one can use so called quasi-Newton methods:
Approximate the Jacobian with another matrix e
J and solve
Jx = f(xk ).
e
Damp the step by a step length k . 1,

xk+1 = xk + k x.

Update e
J by a simple update, e.g., a rank-1 update (recall homework
2).

A. Donev (Courant Institute) Lecture VI 10/14/2010 23 / 31

 Systems of Non-Linear Equations

In practice

It is much harder to construct general robust solvers in higher

dimensions and some problem-specific knowledge is required.
There is no built-in function for solving nonlinear systems in
MATLAB, but the Optimization Toolbox has fsolve.
In many practical situations there is some continuity of the problem
so that a previous solution can be used as an initial guess.
For example, implicit methods for differential equations have a
time-dependent Jacobian J(t) and in many cases the solution x(t)
evolves smootly in time.
For large problems specialized sparse-matrix solvers need to be used.
In many cases derivatives are not provided but there are some
techniques for automatic differentiation.

A. Donev (Courant Institute) Lecture VI 10/14/2010 24 / 31

 Intro to Unconstrained Optimization

Formulation
Optimization problems are among the most important in engineering
and finance, e.g., minimizing production cost, maximizing profits,
etc.
minn f (x)
xR
where x are some variable parameters and f : Rn R is a scalar
objective function.
Observe that one only need to consider minimization as
max f (x) = minn [f (x)]
xRn xR
A local minimum x? is optimal in some neighborhood,
?
f (x ) f (x) x s.t. kx x? k R > 0.
(think of finding the bottom of a valley)
Finding the global minimum is generally not possible for arbitrary
functions
(think of finding Mt. Everest without a satelite)
A. Donev (Courant Institute) Lecture VI 10/14/2010 25 / 31
 Intro to Unconstrained Optimization

Connection to nonlinear systems

Assume that the objective function is differentiable (i.e., first-order

Taylor series converges or gradient exists).
Then a necessary condition for a local minimizer is that x? be a
critical point
 
f
g (x? ) = x f (x? ) = (x? ) = 0
xi i

which is a system of non-linear equations!

In fact similar methods, such as Newton or quasi-Newton, apply to
both problems.
Vice versa, observe that solving f (x) = 0 is equivalent to an
optimization problem h i
min f (x)T f (x)
x
although this is only recommended under special circumstances.
A. Donev (Courant Institute) Lecture VI 10/14/2010 26 / 31
 Intro to Unconstrained Optimization

Sufficient Conditions

Assume now that the objective function is twice-differentiable (i.e.,

Hessian exists).
A critical point x? is a local minimum if the Hessian is positive
definite
H (x? ) = 2x f (x? )  0
which means that the minimum really looks like a valley or a convex
bowl.
At any local minimum the Hessian is positive semi-definite,
2x f (x? )  0.
Methods that require Hessian information converge fast but are
expensive (next class).

A. Donev (Courant Institute) Lecture VI 10/14/2010 27 / 31

 Intro to Unconstrained Optimization

Direct-Search Methods

A direct search method only requires f (x) to be continuous but

not necessarily differentiable, and requires only function evaluations.
Methods that do a search similar to that in bisection can be devised
in higher dimensions also, but they may fail to converge and are
usually slow.
The MATLAB function fminsearch uses the Nelder-Mead or
simplex-search method, which can be thought of as rolling a simplex
downhill to find the bottom of a valley. But there are many others
and this is an active research area.
Curse of dimensionality: As the number of variables
(dimensionality) n becomes larger, direct search becomes hopeless
since the number of samples needed grows as 2n !

A. Donev (Courant Institute) Lecture VI 10/14/2010 28 / 31

 Intro to Unconstrained Optimization

Minimum of 100(x2 x12 )2 + (a x1 )2 in MATLAB

% R o s e n b r o c k o r banana f u n c t i o n :
a = 1;
banana = @( x ) 1 0 0 ( x (2) x ( 1 ) 2 ) 2 + ( ax ( 1 ) ) 2 ;

% T h i s f u n c t i o n must a c c e p t a r r a y a r g u m e n t s !
b a n a n a x y = @( x1 , x2 ) 1 0 0 ( x2x1 . 2 ) . 2 + ( ax1 ) . 2 ;

f i g u r e ( 1 ) ; e z s u r f ( banana xy , [ 0 , 2 , 0 , 2 ] )

[ x , y ] = meshgrid ( l i n s p a c e ( 0 , 2 , 1 0 0 ) ) ;
f i g u r e ( 2 ) ; c o n t o u r f ( x , y , banana xy ( x , y ) ,100)

% C o r r e c t a n s w e r s a r e x = [ 1 , 1 ] and f ( x )=0
[ x , f v a l ] = f m i n s e a r c h ( banana , [ 1 . 2 , 1 ] , o p t i m s e t ( TolX , 1 e 8))
x = 0.999999999187814 0.999999998441919
fval = 1 . 0 9 9 0 8 8 9 5 1 9 1 9 5 7 3 e 18

A. Donev (Courant Institute) Lecture VI 10/14/2010 29 / 31

 Intro to Unconstrained Optimization

Figure of Rosenbrock f (x)

100 (x2x12)2+(ax1)2
2

1.8

1.6
800
1.4

600
1.2

400 1

200 0.8

0.6
0
2
0.4
1 1
0 0.5 0.2
0
1
0.5 0
x2 2 1 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x1

A. Donev (Courant Institute) Lecture VI 10/14/2010 30 / 31

 Conclusions

Conclusions/Summary

Root finding is well-conditioned for simple roots (unit multiplicity),

ill-conditioned otherwise.
Methods for solving nonlinear equations are always iterative and the
order of convergence matters: second order is usually good enough.
A good method uses a higher-order unsafe method such as Newton
method near the root, but safeguards it with something like the
bisection method.
Newtons method is second-order but requires derivative/Jacobian
evaluation. In higher dimensions having a good initial guess for
Newtons method becomes very important.
Quasi-Newton methods can aleviate the complexity of solving the
Jacobian linear system.

A. Donev (Courant Institute) Lecture VI 10/14/2010 31 / 31