nleqslv: Solving systems of nonlinear equations with Broyden or Newton

Description

The function solves a system of nonlinear equations with either a Broyden or a full Newton method. It provides line search and trust region global strategies for difficult systems.

Usage

nleqslv(x, fn, jac=NULL, ...,
               method = c("Broyden", "Newton"),
               global = c("dbldog", "pwldog",
                          "cline", "qline", "gline", "hook", "none"),
               xscalm = c("fixed","auto"),
               jacobian=FALSE,
               control = list()
	)

Value

A list containing components

x

final values for x

fvec

function values

termcd

termination code as integer. The values returned are

1: Function criterion is near zero. Convergence of function values has been achieved.

2

x-values within tolerance. This means that the relative distance between two consecutive x-values is smaller than xtol but that the function value criterion is still larger than ftol. Function values may not be near zero; therefore the user must check if function values are acceptably small.

3

No better point found. This means that the algorithm has stalled and cannot find an acceptable new point. This may or may not indicate acceptably small function values.

4

Iteration limit maxit exceeded.

5

Jacobian is too ill-conditioned.

6

Jacobian is singular.

7

Jacobian is unusable.

-10

User supplied Jacobian is most likely incorrect.

message

a string describing the termination code

scalex

a vector containing the scaling factors, which will be the final values when automatic scaling was selected

njcnt

number of Jacobian evaluations

nfcnt

number of function evaluations, excluding those required for calculating a Jacobian and excluding the initial function evaluation (at iteration 0)

iter

number of outer iterations used by the algorithm. This excludes the initial iteration. The number of backtracks can be calculated as the difference between the nfcnt and iter components.

jac

the final Jacobian or the Broyden approximation if jacobian was set to TRUE. If no iterations were executed, as may happen when the initial guess are sufficiently close the solution, there is no Broyden approximation and the returned matrix will always be the actual Jacobian. If the matrix is singular or too ill-conditioned the returned matrix is of no value.

Arguments

x: A numeric vector with an initial guess of the root of the function.
fn: A function of x returning a vector of function values with the same length as the vector x.
jac: A function to return the Jacobian for the fn function. For a vector valued function fn the Jacobian must be a numeric matrix of the correct dimensions. For a scalar valued function fn the jac function may return a scalar. If not supplied numerical derivatives will be used.
...: Further arguments to be passed to fn and jac.
method: The method to use for finding a solution. See ‘Details’.
global: The global strategy to apply. See ‘Details’.
xscalm: The type of x scaling to use. See ‘Details’.
jacobian: A logical indicating if the estimated (approximate) jacobian in the solution should be returned. See ‘Details’.
control: A named list of control parameters. See ‘Details’.

Warning

You cannot use this function recursively. Thus function fn should not in its turn call nleqslv.

Details

The algorithms implemented in nleqslv are based on Dennis and Schnabel (1996).

Methods

Method Broyden starts with a computed Jacobian of the function and then updates this Jacobian after each successful iteration using the so-called Broyden update. This method often shows super linear convergence towards a solution. When nleqslv determines that it cannot continue with the current Broyden matrix it will compute a new Jacobian.

Method Newton calculates a Jacobian of the function fn at each iteration. Close to a solution this method will usually show quadratic convergence.

Both methods apply a so-called (backtracking) global strategy to find a better (more acceptable) iterate. The function criterion used by the algorithm is half of the sum of squares of the function values and “acceptable” means sufficient decrease of the current function criterion value compared to that of the previous iteration. A comprehensive discussion of these issues can be found in Dennis and Schnabel (1996). Both methods apply an unpivoted QR-decomposition to the Jacobian as implemented in Lapack. The Broyden method applies a rank-1 update to the Jacobian at the end of each iteration and is based on a simplified and modernized version of the algorithm described in Reichel and Gragg (1990).

Global strategies

When applying a full Newton or Broyden step does not yield a sufficiently smaller function criterion value nleqslv will attempt to decrease the steplength using one of several so-called global strategies.

The global argument indicates which global strategy to use or to use no global strategy

cline: a cubic line search
qline: a quadratic line search
gline: a geometric line search
dbldog: a trust region method using the double dogleg method as described in Dennis and Schnabel (1996)
pwldog: a trust region method using the Powell dogleg method as developed by Powell (1970).
hook: a trust region method described by Dennis and Schnabel (1996) as The locally constrained optimal (“hook”) step. It is equivalent to a Levenberg-Marquardt algorithm as described in Moré (1978) and Nocedal and Wright (2006).
none: Only a pure local Newton or Broyden iteration is used. The maximum stepsize (see below) is taken into account. The default maximum number of iterations (see below) is set to 20.

The double dogleg method is the default global strategy employed by this package.

Which global strategy to use in a particular situation is a matter of trial and error. When one of the trust region methods fails, one of the line search strategies should be tried. Sometimes a trust region will work and sometimes a line search method; neither has a clear advantage but in many cases the double dogleg method works quite well.

When the function to be solved returns non-finite function values for a parameter vector x and the algorithm is not evaluating a numerical Jacobian, then any non-finite values will be replaced by a large number forcing the algorithm to backtrack, i.e. decrease the line search factor or decrease the trust region radius.

Scaling

The elements of vector x may be scaled during the search for a zero of fn. The xscalm argument provides two possibilities for scaling

fixed: the scaling factors are set to the values supplied in the control argument and remain unchanged during the iterations. The scaling factor of any element of x should be set to the inverse of the typical value of that element of x, ensuring that all elements of x are approximately equal in size.
auto: the scaling factors are calculated from the euclidean norms of the columns of the Jacobian matrix. When a new Jacobian is computed, the scaling values will be set to the euclidean norm of the corresponding column if that is larger than the current scaling value. Thus the scaling values will not decrease during the iteration. This is the method described in Moré (1978). Usually manual scaling is preferable.

Jacobian

When evaluating a numerical Jacobian, an error message will be issued on detecting non-finite function values. An error message will also be issued when a user supplied jacobian contains non-finite entries.

When the jacobian argument is set to TRUE the final Jacobian or Broyden matrix will be returned in the return list. The default value is FALSE; i.e. to not return the final matrix. There is no guarantee that the final Broyden matrix resembles the actual Jacobian.

The package can cope with a singular or ill-conditioned Jacobian if needed by setting the allowSingular component of the control argument. The method used is described in Dennis and Schnabel (1996); it is equivalent to a Levenberg-Marquardt type adjustment with a small damping factor. There is no guarantee that this method will be successful. Warning: nleqslv may report spurious convergence in this case.

By default nleqslv returns an error if a Jacobian becomes singular or very ill-conditioned. A Jacobian is considered to be very ill-conditioned when the estimated inverse condition is less than or equal to a specified tolerance with a default value equal to $10^{-12}$; this can be changed and made smaller with the cndtol item of the control argument. There is no guarantee that any change will be effective.

Control options

The control argument is a named list that can supply any of the following components:

xtol: The relative steplength tolerance. When the relative steplength of all scaled x values is smaller than this value convergence is declared. The default value is $10^{-8}$.
ftol: The function value tolerance. Convergence is declared when the largest absolute function value is smaller than ftol. The default value is $10^{-8}$.
btol: The backtracking tolerance. When the relative steplength in a backtracking step to find an acceptable point is smaller than the backtracking tolerance, the backtracking is terminated. In the Broyden method a new Jacobian will be calculated if the Jacobian is outdated. The default value is $10^{-3}$.
cndtol: The tolerance of the test for ill conditioning of the Jacobian or Broyden approximation. If less than the machine precision it will be silently set to the machine precision. When the estimated inverse condition of the (approximated) Jacobian matrix is less than or equal to the value of cndtol the matrix is deemed to be ill-conditioned, in which case an error will be reported if the allowSingular component is set to FALSE. The default value is $10^{-12}$.
sigma: Reduction factor for the geometric line search. The default value is 0.5.
scalex: a vector of scaling values for the parameters. The inverse of a scale value is an indication of the size of a parameter. The default value is 1.0 for all scale values.
maxit: The maximum number of major iterations. The default value is 150 if a global strategy has been specified. If no global strategy has been specified the default is 20.
trace: Non-negative integer. A value of 1 will give a detailed report of the progress of the iteration. For a description see Iteration report.
chkjac: A logical value indicating whether to check a user supplied Jacobian, if supplied. The default value is FALSE. The first 10 errors are printed. The code for this check is derived from the code in Bouaricha and Schnabel (1997).
delta: Initial (scaled) trust region radius. A value of $-1.0$ or "cauchy" is replaced by the length of the Cauchy step in the initial point. A value of $-2.0$ or "newton" is replaced by the length of the Newton step in the initial point. Any numeric value less than or equal to 0 and not equal to $-2.0$, will be replaced by $-1.0$; the algorithm will then start with the length of the Cauchy step in the initial point. If it is numeric and positive it will be set to the smaller of the value supplied or the maximum stepsize. If it is not numeric and not one of the permitted character strings then an error message will be issued. The default is $-2.0$.
stepmax: Maximum scaled stepsize. If this is negative then the maximum stepsize is set to the largest positive representable number. The default is $-1.0$, so there is no default maximum stepsize.
dsub: Number of non zero subdiagonals of a banded Jacobian. The default is to assume that the Jacobian is not banded. Must be specified if dsuper has been specified and must be larger than zero when dsuper is zero.
dsuper: Number of non zero super diagonals of a banded Jacobian. The default is to assume that the Jacobian is not banded. Must be specified if dsub has been specified and must be larger than zero when dsub is zero.
allowSingular: A logical value indicating if a small correction to the Jacobian when it is singular or too ill-conditioned is allowed. If the correction is less than 100*.Machine$double.eps the correction cannot be applied and an unusable Jacobian will be reported. The method used is similar to a Levenberg-Marquardt correction and is explained in Dennis and Schnabel (1996) on page 151. It may be necessary to choose a higher value for cndtol to enforce the correction. The default value is FALSE.

References

Bouaricha, A. and Schnabel, R.B. (1997), Algorithm 768: TENSOLVE: A Software Package for Solving Systems of Nonlinear Equations and Nonlinear Least-squares Problems Using Tensor Methods, Transactions on Mathematical Software, 23, 2, pp. 174--195.

Dennis, J.E. Jr and Schnabel, R.B. (1996), Numerical Methods for Unconstrained Optimization and Nonlinear Equations, Siam.

Moré, J.J. (1978), The Levenberg-Marquardt Algorithm, Implementation and Theory, In Numerical Analysis, G.A. Watson (Ed.), Lecture Notes in Mathematics 630, Springer-Verlag, pp. 105--116.

Golub, G.H and C.F. Van Loan (1996), Matrix Computations (3rd edition), The John Hopkins University Press.

Higham, N.J. (2002), Accuracy and Stability of Numerical Algorithms, 2nd ed., SIAM, pp. 10--11.

Nocedal, J. and Wright, S.J. (2006), Numerical Optimization, Springer.

Powell, M.J.D. (1970), A hybrid method for nonlinear algebraic equations, In Numerical Methods for Nonlinear Algebraic Equations, P. Rabinowitz (Ed.), Gordon & Breach.

Powell, M.J.D. (1970), A Fortran subroutine for solving systems nonlinear equations, In Numerical Methods for Nonlinear Algebraic Equations, P. Rabinowitz (Ed.), Gordon & Breach.

Reichel, L. and W.B. Gragg (1990), Algorithm 686: FORTRAN subroutines for updating the QR decomposition, ACM Trans. Math. Softw., 16, 4, pp. 369--377.

Examples

Run this code

# Dennis Schnabel example 6.5.1 page 149
dslnex <- function(x) {
    y <- numeric(2)
    y[1] <- x[1]^2 + x[2]^2 - 2
    y[2] <- exp(x[1]-1) + x[2]^3 - 2
    y
}

jacdsln <- function(x) {
    n <- length(x)
    Df <- matrix(numeric(n*n),n,n)
    Df[1,1] <- 2*x[1]
    Df[1,2] <- 2*x[2]
    Df[2,1] <- exp(x[1]-1)
    Df[2,2] <- 3*x[2]^2

    Df
}

BADjacdsln <- function(x) {
    n <- length(x)
    Df <- matrix(numeric(n*n),n,n)
    Df[1,1] <- 4*x[1]
    Df[1,2] <- 2*x[2]
    Df[2,1] <- exp(x[1]-1)
    Df[2,2] <- 5*x[2]^2

    Df
}

xstart <- c(2,0.5)
fstart <- dslnex(xstart)
xstart
fstart

# a solution is c(1,1)

nleqslv(xstart, dslnex, control=list(btol=.01))

# Cauchy start
nleqslv(xstart, dslnex, control=list(trace=1,btol=.01,delta="cauchy"))

# Newton start
nleqslv(xstart, dslnex, control=list(trace=1,btol=.01,delta="newton"))

# final Broyden approximation of Jacobian (quite good)
z <- nleqslv(xstart, dslnex, jacobian=TRUE,control=list(btol=.01))
z$x
z$jac
jacdsln(z$x)

# different initial start; not a very good final approximation
xstart <- c(0.5,2)
z <- nleqslv(xstart, dslnex, jacobian=TRUE,control=list(btol=.01))
z$x
z$jac
jacdsln(z$x)

if (FALSE) {
# no global strategy but limit stepsize
# but look carefully: a different solution is found
nleqslv(xstart, dslnex, method="Newton", global="none", control=list(trace=1,stepmax=5))

# but if the stepsize is limited even more the c(1,1) solution is found
nleqslv(xstart, dslnex, method="Newton", global="none", control=list(trace=1,stepmax=2))

# Broyden also finds the c(1,1) solution when the stepsize is limited
nleqslv(xstart, dslnex, jacdsln, method="Broyden", global="none", control=list(trace=1,stepmax=2))
}

# example with a singular jacobian in the initial guess
f <- function(x) {
    y <- numeric(3)
    y[1] <- x[1] + x[2] - x[1]*x[2] - 2
    y[2] <- x[1] + x[3] - x[1]*x[3] - 3
    y[3] <- x[2] + x[3] - 4
    return(y)
}

Jac <- function(x) {
    J <- matrix(0,nrow=3,ncol=3)
    J[,1] <- c(1-x[2],1-x[3],0)
    J[,2] <- c(1-x[1],0,1)
    J[,3] <- c(0,1-x[1],1)
    J
}

# exact solution
xsol <- c(-.5, 5/3 , 7/3)
xsol

xstart <- c(1,2,3)
J <- Jac(xstart)
J
rcond(J)

z <- nleqslv(xstart,f,Jac, method="Newton",control=list(trace=1,allowSingular=TRUE))
all.equal(z$x,xsol)

Run the code above in your browser using DataLab