Find minimum of unconstrained multivariable function
Finds the minimum of a problem specified by
where f(x) is a function that returns a scalar.
x is a vector or a matrix; see Matrix Arguments.
x = fminunc(fun,x0)
x = fminunc(fun,x0,options)
x = fminunc(problem)
[x,fval] = fminunc(...)
[x,fval,exitflag] = fminunc(...)
[x,fval,exitflag,output] = fminunc(...)
[x,fval,exitflag,output,grad] = fminunc(...)
[x,fval,exitflag,output,grad,hessian] = fminunc(...)
fminunc attempts to find a minimum of a scalar function of several variables, starting at an initial estimate. This is generally referred to as unconstrained nonlinear optimization.
Note: Passing Extra Parameters explains how to pass extra parameters to the objective function, if necessary.
x = fminunc(fun,x0,options) minimizes with the optimization options specified in options. Use optimoptions to set these options.
x = fminunc(problem) finds the minimum for problem, where problem is a structure described in Input Arguments.
Create the problem structure by exporting a problem from Optimization app, as described in Exporting Your Work.
[x,fval,exitflag,output,grad,hessian] = fminunc(...) returns in hessian the value of the Hessian of the objective function fun at the solution x. See Hessian.
Function Arguments contains general descriptions of arguments passed into fminunc. This section provides function-specific details for fun, options, and problem:
The function to be minimized. fun is a function that accepts a vector x and returns a scalar f, the objective function evaluated at x. The function fun can be specified as a function handle for a file
x = fminunc(@myfun,x0)
where myfun is a MATLAB® function such as
function f = myfun(x) f = ... % Compute function value at x
fun can also be a function handle for an anonymous function.
x = fminunc(@(x)norm(x)^2,x0);
If the gradient of fun can also be computed and the GradObj option is 'on', as set by
options = optimoptions(@fminunc,'GradObj','on')
then the function fun must return, in the second output argument, the gradient value g, a vector, at x. The gradient is the partial derivatives ∂f/∂xi of f at the point x. That is, the ith component of g is the partial derivative of f with respect to the ith component of x.
If the Hessian matrix can also be computed and the Hessian option is 'on', i.e., options = optimoptions(@fminunc,'GradObj','on','Hessian','on'), then the function fun must return the Hessian value H, a symmetric matrix, at x in a third output argument. The Hessian matrix is the second partial derivatives matrix of f at the point x. That is, the (i,j)th component of H is the second partial derivative of f with respect to xi and xj, ∂2f/∂xi∂xj. The Hessian is by definition a symmetric matrix.
Writing Scalar Objective Functions explains how to "conditionalize" the gradients and Hessians for use in solvers that do not accept them. Passing Extra Parameters explains how to parameterize fun, if necessary.
Options provides the function-specific details for the options values.
|Initial point for x|
|Options created with optimoptions|
Function Arguments contains general descriptions of arguments returned by fminunc. This section provides function-specific details for exitflag and output:
Integer identifying the reason the algorithm terminated. The following lists the values of exitflag and the corresponding reasons the algorithm terminated.
Magnitude of gradient smaller than the TolFun tolerance.
Change in x was smaller than the TolX tolerance.
Change in the objective function value was less than the TolFun tolerance.
Predicted decrease in the objective function was less than the TolFun tolerance.
Number of iterations exceeded MaxIter or number of function evaluations exceeded MaxFunEvals.
Algorithm was terminated by the output function.
Objective function at current iteration went below ObjectiveLimit.
Gradient at x
Hessian at x
Structure containing information about the optimization. The fields of the structure are
Number of iterations taken
Number of function evaluations
Measure of first-order optimality
Optimization algorithm used
Total number of PCG iterations (trust-region algorithm only)
Final displacement in x (medium-scale algorithm only)
fminunc computes the output argument hessian as follows:
When using the medium-scale algorithm, the function computes a finite-difference approximation to the Hessian at x using
The gradient grad if you supply it
The objective function fun if you do not supply the gradient
When using the trust-region algorithm, the function uses
options.Hessian, if you supply it, to compute the Hessian at x
A finite-difference approximation to the Hessian at x, if you supply only the gradient
fminunc uses these optimization options. Some options apply to all algorithms, some are only relevant when you are using the trust-region algorithm, and others are only relevant when you are using the quasi-newton algorithm. Use optimoptions to set or change options. See Optimization Options Reference for detailed information.
All fminunc algorithms use the following options:
If you use optimset (not recommended), use LargeScale instead of Algorithm.
Choose the fminunc algorithm. Choices are 'quasi-newton' and 'trust-region' (default).
The trust-region algorithm requires you to provide the gradient (see the preceding description of fun), or else fminunc uses the 'quasi-newton' algorithm. For information on choosing the algorithm, see Choosing the Algorithm.
Compare user-supplied derivatives (gradient of objective) to finite-differencing derivatives. The choices are 'on' or the default 'off'.
Display diagnostic information about the function to be minimized or solved. The choices are 'on' or the default 'off'.
Maximum change in variables for finite-difference gradients (a positive scalar). The default is Inf.
Minimum change in variables for finite-difference gradients (a positive scalar). The default is 0.
Level of display:
Scalar or vector step size factor. When you set FinDiffRelStep to a vector v, forward finite differences delta are
delta = v.*sign(x).*max(abs(x),TypicalX);
and central finite differences are
delta = v.*max(abs(x),TypicalX);
Scalar FinDiffRelStep expands to a vector. The default is sqrt(eps) for forward finite differences, and eps^(1/3) for central finite differences.
The trust-region algorithm uses FinDiffRelStep only when DerivativeCheck is 'on'.
Finite differences, used to estimate gradients, are either 'forward' (the default), or 'central' (centered). 'central' takes twice as many function evaluations, but should be more accurate. The trust-region algorithm uses FinDiffType only when DerivativeCheck is 'on'.
Check whether objective function values are valid. 'on' displays an error when the objective function returns a value that is complex, Inf, or NaN. The default, 'off', displays no error.
Gradient for the objective function defined by the user. See the preceding description of fun to see how to define the gradient in fun. Set to 'on' to have fminunc use a user-defined gradient of the objective function. The default 'off' causes fminunc to estimate gradients using finite differences. You must provide the gradient, and set GradObj to 'on', to use the trust-region algorithm. This option is not required for the quasi-Newton algorithm.
If you use optimoptions (recommended), use Algorithm instead of LargeScale.
Use large-scale algorithm if possible when set to the default 'on'. Use medium-scale algorithm when set to 'off'.
The LargeScale algorithm requires you to provide the gradient (see the preceding description of fun), or else fminunc uses the medium-scale algorithm. For information on choosing the algorithm, see Choosing the Algorithm.
Maximum number of function evaluations allowed, a positive integer. The default value is 100*numberOfVariables.
Maximum number of iterations allowed, a positive integer. The default value is 400.
Specify one or more user-defined functions that an optimization function calls at each iteration, either as a function handle or as a cell array of function handles. The default is none (). See Output Function.
Plots various measures of progress while the algorithm executes. Select from predefined plots or write your own. Pass a function handle or a cell array of function handles. The default is none ().
For information on writing a custom plot function, see Plot Functions.
Termination tolerance on the function value, a positive scalar. The default is 1e-6.
Termination tolerance on x, a positive scalar. The default value is 1e-6.
Typical x values. The number of elements in TypicalX is equal to the number of elements in x0, the starting point. The default value is ones(numberofvariables,1). fminunc uses TypicalX for scaling finite differences for gradient estimation.
The trust-region algorithm uses TypicalX only for the DerivativeCheck option.
The trust-region algorithm uses the following options:
If 'on', fminunc uses a user-defined Hessian for the objective function. The Hessian is either defined in the objective function (see fun), or is indirectly defined when using HessMult.
If 'off' (default), fminunc approximates the Hessian using finite differences.
Function handle for Hessian multiply function. For large-scale structured problems, this function computes the Hessian matrix product H*Y without actually forming H. The function is of the form
W = hmfun(Hinfo,Y)
where Hinfo contains the matrix used to compute H*Y.
The first argument must be the same as the third argument returned by the objective function fun, for example by
[f,g,Hinfo] = fun(x)
Y is a matrix that has the same number of rows as there are dimensions in the problem. W = H*Y although H is not formed explicitly. fminunc uses Hinfo to compute the preconditioner. See Passing Extra Parameters for information on how to supply values for any additional parameters hmfun needs.
See Minimization with Dense Structured Hessian, Linear Equalities for an example.
Sparsity pattern of the Hessian for finite differencing. If it is not convenient to compute the sparse Hessian matrix H in fun, the trust-region method in fminunc can approximate H via sparse finite differences (of the gradient) provided the sparsity structure of H —i.e., locations of the nonzeros—is supplied as the value for HessPattern. In the worst case, if the structure is unknown, you can set HessPattern to be a dense matrix and a full finite-difference approximation is computed at each iteration (this is the default). This can be very expensive for large problems, so it is usually worth the effort to determine the sparsity structure.
Maximum number of PCG (preconditioned conjugate gradient) iterations, a positive scalar. The default is max(1,floor(numberOfVariables/2)). For more information, see Algorithms.
Upper bandwidth of preconditioner for PCG, a nonnegative integer. By default, fminunc uses diagonal preconditioning (upper bandwidth of 0). For some problems, increasing the bandwidth reduces the number of PCG iterations. Setting PrecondBandWidth to Inf uses a direct factorization (Cholesky) rather than the conjugate gradients (CG). The direct factorization is computationally more expensive than CG, but produces a better quality step towards the solution.
Termination tolerance on the PCG iteration, a positive scalar. The default is 0.1.
The quasi-newton algorithm uses the following options:
Method for choosing the search direction in the Quasi-Newton algorithm. The choices are:
Initial quasi-Newton matrix. This option is only available if you set InitialHessType to 'user-supplied'. In that case, you can set InitialHessMatrix to one of the following:
Initial quasi-Newton matrix type. The options are:
A tolerance (stopping criterion) that is a scalar. If the objective function value at an iteration is less than or equal to ObjectiveLimit, the iterations halt, since the problem is presumably unbounded. The default value is -1e20.
Minimize the function .
Create a file myfun.m:
function f = myfun(x) f = 3*x(1)^2 + 2*x(1)*x(2) + x(2)^2; % Cost function
Then call fminunc to find a minimum of myfun near [1,1]:
x0 = [1,1]; [x,fval] = fminunc(@myfun,x0);
After a few iterations, fminunc returns the solution, x, and the value of the function at x, fval:
x,fval x = 1.0e-006 * 0.2541 -0.2029 fval = 1.3173e-013
To minimize this function with the gradient provided, modify myfun.m so the gradient is the second output argument:
function [f,g] = myfun(x) f = 3*x(1)^2 + 2*x(1)*x(2) + x(2)^2; % Cost function if nargout > 1 g(1) = 6*x(1)+2*x(2); g(2) = 2*x(1)+2*x(2); end
Indicate that the gradient value is available by creating optimization options with the GradObj option set to 'on' using optimoptions.
options = optimoptions('fminunc','GradObj','on'); x0 = [1,1]; [x,fval] = fminunc(@myfun,x0,options);
After several iterations fminunc returns the solution, x, and the value of the function at x, fval:
x,fval x = 1.0e-015 * 0.1110 -0.8882 fval = 6.2862e-031
To minimize the function f(x) = sin(x) + 3 using an anonymous function
f = @(x)sin(x)+3; x = fminunc(f,4)
fminunc returns a solution
x x = 4.7124
Instead use the lsqnonlin function, which has been optimized for problems of this form.
To use the trust-region method, you must provide the gradient in fun (and set the GradObj option to 'on' using optimoptions). A warning is given if no gradient is provided and the Algorithm option is 'trust-region'.
fminunc only minimizes over the real numbers, that is, x must only consist of real numbers and f(x) must only return real numbers. When x has complex variables, they must be split into real and imaginary parts.
To use the trust-region algorithm, you must supply the gradient in fun (and GradObj must be set 'on' in options).
Trust Region Algorithm Coverage and Requirements
|Additional Information Needed||For Large Problems|
Must provide gradient for f(x) in fun.
By default fminunc chooses the trust-region algorithm if you supply the gradient in fun and set GradObj to 'on' using optimoptions. This algorithm is a subspace trust-region method and is based on the interior-reflective Newton method described in  and . Each iteration involves the approximate solution of a large linear system using the method of preconditioned conjugate gradients (PCG). See fminunc trust-region Algorithm, Trust-Region Methods for Nonlinear Minimization and Preconditioned Conjugate Gradient Methodpt.
The quasi-newton algorithm uses the BFGS Quasi-Newton method with a cubic line search procedure. This quasi-Newton method uses the BFGS (,,, and ) formula for updating the approximation of the Hessian matrix. You can select the DFP (,, and ) formula, which approximates the inverse Hessian matrix, by setting the HessUpdate option to 'dfp' (and the Algorithm option to 'quasi-newton'). You can select a steepest descent method by setting HessUpdate to 'steepdesc' (and Algorithm to 'quasi-newton'), although this is not recommended.
 Broyden, C.G., "The Convergence of a Class of Double-Rank Minimization Algorithms," Journal Inst. Math. Applic., Vol. 6, pp. 76-90, 1970.
 Coleman, T.F. and Y. Li, "An Interior, Trust Region Approach for Nonlinear Minimization Subject to Bounds," SIAM Journal on Optimization, Vol. 6, pp. 418-445, 1996.
 Coleman, T.F. and Y. Li, "On the Convergence of Reflective Newton Methods for Large-Scale Nonlinear Minimization Subject to Bounds," Mathematical Programming, Vol. 67, Number 2, pp. 189-224, 1994.
 Davidon, W.C., "Variable Metric Method for Minimization," A.E.C. Research and Development Report, ANL-5990, 1959.
 Fletcher, R., "A New Approach to Variable Metric Algorithms," Computer Journal, Vol. 13, pp. 317-322, 1970.
 Fletcher, R., "Practical Methods of Optimization," Vol. 1, Unconstrained Optimization, John Wiley and Sons, 1980.
 Fletcher, R. and M.J.D. Powell, "A Rapidly Convergent Descent Method for Minimization," Computer Journal, Vol. 6, pp. 163-168, 1963.
 Goldfarb, D., "A Family of Variable Metric Updates Derived by Variational Means," Mathematics of Computing, Vol. 24, pp. 23-26, 1970.
 Shanno, D.F., "Conditioning of Quasi-Newton Methods for Function Minimization," Mathematics of Computing, Vol. 24, pp. 647-656, 1970.