Accelerating the pace of engineering and science

Documentation

Symbolic Math Toolbox Calculates Gradients and Hessians

If you have a Symbolic Math Toolbox™ license, you can easily calculate analytic gradients and Hessians for objective and constraint functions. There are two relevant Symbolic Math Toolbox functions:

• jacobian generates the gradient of a scalar function, and generates a matrix of the partial derivatives of a vector function. So, for example, you can obtain the Hessian matrix, the second derivatives of the objective function, by applying jacobian to the gradient. Part of this example shows how to use jacobian to generate symbolic gradients and Hessians of objective and constraint functions.

• matlabFunction generates either an anonymous function or a file that calculates the values of a symbolic expression. This example shows how to use matlabFunction to generate files that evaluate the objective and constraint function and their derivatives at arbitrary points.

Consider the electrostatics problem of placing 10 electrons in a conducting body. The electrons will arrange themselves so as to minimize their total potential energy, subject to the constraint of lying inside the body. It is well known that all the electrons will be on the boundary of the body at a minimum. The electrons are indistinguishable, so there is no unique minimum for this problem (permuting the electrons in one solution gives another valid solution). This example was inspired by Dolan, Moré, and Munson [58].

This example is a conducting body defined by the following inequalities:

 $z\le -|x|-|y|$ (6-60)
 ${x}^{2}+{y}^{2}+{\left(z+1\right)}^{2}\le 1.$ (6-61)

This body looks like a pyramid on a sphere.

There is a slight gap between the upper and lower surfaces of the figure. This is an artifact of the general plotting routine used to create the figure. This routine erases any rectangular patch on one surface that touches the other surface.

The syntax and structures of the two sets of toolbox functions differ. In particular, symbolic variables are real or complex scalars, but Optimization Toolbox™ functions pass vector arguments. So there are several steps to take to generate symbolically the objective function, constraints, and all their requisite derivatives, in a form suitable for the interior-point algorithm of fmincon:

To see the efficiency in using gradients and Hessians, see Compare to Optimization Without Gradients and Hessians.

Create the Variables

Generate a symbolic vector x as a 30-by-1 vector composed of real symbolic variables xij, i between 1 and 10, and j between 1 and 3. These variables represent the three coordinates of electron i: xi1 corresponds to the x coordinate, xi2 corresponds to the y coordinate, and xi3 corresponds to the z coordinate.

```x = cell(3, 10);
for i = 1:10
for j = 1:3
x{j,i} = sprintf('x%d%d',i,j);
end
end
x = x(:); % now x is a 30-by-1 vector
x = sym(x, 'real');```

The vector x is:

```x
x =
x11
x12
x13
x21
x22
x23
x31
x32
x33
x41
x42
x43
x51
x52
x53
x61
x62
x63
x71
x72
x73
x81
x82
x83
x91
x92
x93
x101
x102
x103```

Include the Linear Constraints

Write the linear constraints in Equation 6-60,

z ≤ –|x| – |y|,

as a set of four linear inequalities for each electron:

xi3xi1xi2 ≤ 0
xi3xi1 + xi2 ≤ 0
xi3 + xi1xi2 ≤ 0
xi3 + xi1 + xi2 ≤ 0

Therefore there are a total of 40 linear inequalities for this problem.

Write the inequalities in a structured way:

```B = [1,1,1;-1,1,1;1,-1,1;-1,-1,1];

A = zeros(40,30);
for i=1:10
A(4*i-3:4*i,3*i-2:3*i) = B;
end

b = zeros(40,1);```

You can see that A*x ≤ b represents the inequalities:

```A*x

ans =
x11 + x12 + x13
x12 - x11 + x13
x11 - x12 + x13
x13 - x12 - x11
x21 + x22 + x23
x22 - x21 + x23
x21 - x22 + x23
x23 - x22 - x21
x31 + x32 + x33
x32 - x31 + x33
x31 - x32 + x33
x33 - x32 - x31
x41 + x42 + x43
x42 - x41 + x43
x41 - x42 + x43
x43 - x42 - x41
x51 + x52 + x53
x52 - x51 + x53
x51 - x52 + x53
x53 - x52 - x51
x61 + x62 + x63
x62 - x61 + x63
x61 - x62 + x63
x63 - x62 - x61
x71 + x72 + x73
x72 - x71 + x73
x71 - x72 + x73
x73 - x72 - x71
x81 + x82 + x83
x82 - x81 + x83
x81 - x82 + x83
x83 - x82 - x81
x91 + x92 + x93
x92 - x91 + x93
x91 - x92 + x93
x93 - x92 - x91
x101 + x102 + x103
x102 - x101 + x103
x101 - x102 + x103
x103 - x102 - x101```

Create the Nonlinear Constraints, Their Gradients and Hessians

The nonlinear constraints in Equation 6-61 ,

${x}^{2}+{y}^{2}+{\left(z+1\right)}^{2}\le 1,$

are also structured. Generate the constraints, their gradients, and Hessians as follows:

```c = sym(zeros(1,10));
i = 1:10;
c = (x(3*i-2).^2 + x(3*i-1).^2 + (x(3*i)+1).^2 - 1).';

gradc = jacobian(c,x).'; % .' performs transpose

hessc = cell(1, 10);
for i = 1:10
end```

The constraint vector c is a row vector, and the gradient of c(i) is represented in the ith column of the matrix gradc. This is the correct form, as described in Nonlinear Constraints.

The Hessian matrices, hessc{1}...hessc{10}, are square and symmetric. It is better to store them in a cell array, as is done here, than in separate variables such as hessc1, ..., hesssc10.

Use the .' syntax to transpose. The ' syntax means conjugate transpose, which has different symbolic derivatives.

Create the Objective Function, Its Gradient and Hessian

The objective function, potential energy, is the sum of the inverses of the distances between each electron pair:

$\text{energy}=\sum _{i

The distance is the square root of the sum of the squares of the differences in the components of the vectors.

Calculate the energy, its gradient, and its Hessian as follows:

```energy = sym(0);
for i = 1:3:25
for j = i+3:3:28
dist = x(i:i+2) - x(j:j+2);
energy = energy + 1/sqrt(dist.'*dist);
end
end

Create the Objective Function File

The objective function should have two outputs, energy and gradenergy. Put both functions in one vector when calling matlabFunction to reduce the number of subexpressions that matlabFunction generates, and to return the gradient only when the calling function (fmincon in this case) requests both outputs. This example shows placing the resulting files in your current folder. Of course, you can place them anywhere you like, as long as the folder is on the MATLAB path.

```currdir = [pwd filesep]; % You may need to use currdir = pwd
filename = [currdir,'demoenergy.m'];

This syntax causes matlabFunction to return energy as the first output, and gradenergy as the second. It also takes a single input vector {x} instead of a list of inputs x11, ..., x103.

The resulting file demoenergy.m contains, in part, the following lines or similar ones:

```function [energy,gradenergy] = demoenergy(in1)
%DEMOENERGY
...
x101 = in1(28,:);
...
energy = 1./t140.^(1./2) + ...;
if nargout > 1
...
gradenergy = [(t174.*(t185 - 2.*x11))./2 - ...];
end```

This function has the correct form for an objective function with a gradient; see Writing Scalar Objective Functions.

Create the Constraint Function File

Generate the nonlinear constraint function, and put it in the correct format.

```filename = [currdir,'democonstr.m'];

The resulting file democonstr.m contains, in part, the following lines or similar ones:

```function [c,ceq,gradc,gradceq] = democonstr(in1)
%DEMOCONSTR
...
x101 = in1(28,:);
...
c = [t417.^2 + ...];
if nargout > 1
ceq = [];
end
if nargout > 2
end
if nargout > 3
end
```

This function has the correct form for a constraint function with a gradient; see Nonlinear Constraints.

Generate the Hessian Files

To generate the Hessian of the Lagrangian for the problem, first generate files for the energy Hessian and for the constraint Hessians.

The Hessian of the objective function, hessenergy, is a very large symbolic expression, containing over 150,000 symbols, as evaluating size(char(hessenergy)) shows. So it takes a substantial amount of time to run matlabFunction(hessenergy).

To generate a file hessenergy.m, run the following two lines:

```filename = [currdir,'hessenergy.m'];
matlabFunction(hessenergy,'file',filename,'vars',{x});```

In contrast, the Hessians of the constraint functions are small, and fast to compute:

```for i = 1:10
ii = num2str(i);
thename = ['hessc',ii];
filename = [currdir,thename,'.m'];
matlabFunction(hessc{i},'file',filename,'vars',{x});
end```

After generating all the files for the objective and constraints, put them together with the appropriate Lagrange multipliers in a file hessfinal.m as follows:

```function H = hessfinal(X,lambda)
%
% Call the function hessenergy to start
H = hessenergy(X);

% Add the Lagrange multipliers * the constraint Hessians
H = H + hessc1(X) * lambda.ineqnonlin(1);
H = H + hessc2(X) * lambda.ineqnonlin(2);
H = H + hessc3(X) * lambda.ineqnonlin(3);
H = H + hessc4(X) * lambda.ineqnonlin(4);
H = H + hessc5(X) * lambda.ineqnonlin(5);
H = H + hessc6(X) * lambda.ineqnonlin(6);
H = H + hessc7(X) * lambda.ineqnonlin(7);
H = H + hessc8(X) * lambda.ineqnonlin(8);
H = H + hessc9(X) * lambda.ineqnonlin(9);
H = H + hessc10(X) * lambda.ineqnonlin(10);```

Run the Optimization

Start the optimization with the electrons distributed randomly on a sphere of radius 1/2 centered at [0,0,–1]:

```rng default % for reproducibility
Xinitial = randn(3,10); % columns are normal 3-D vectors
for j=1:10
Xinitial(:,j) = Xinitial(:,j)/norm(Xinitial(:,j))/2;
% this normalizes to a 1/2-sphere
end
Xinitial(3,:) = Xinitial(3,:) - 1; % center at [0,0,-1]
Xinitial = Xinitial(:); % Convert to a column vector```

Set the options to use the interior-point algorithm, and to use gradients and the Hessian:

```options = optimoptions(@fmincon,'Algorithm','interior-point','GradObj','on',...
'HessFcn',@hessfinal,'Display','final');```

Call fmincon:

```[xfinal fval exitflag output] = fmincon(@demoenergy,Xinitial,...
A,b,[],[],[],[],@democonstr,options)```

The output is as follows:

```Local minimum found that satisfies the constraints.

Optimization completed because the objective function is non-decreasing in
feasible directions, to within the default value of the function tolerance,
and constraints are satisfied to within the default value of the constraint tolerance.

xfinal =
-0.0317
0.0317
-1.9990
0.6356
-0.6356
-1.4381
0.0000
-0.0000
-0.0000
0.0000
-1.0000
-1.0000
1.0000
-0.0000
-1.0000
-1.0000
-0.0000
-1.0000
0.6689
0.6644
-1.3333
-0.6667
0.6667
-1.3333
0.0000
1.0000
-1.0000
-0.6644
-0.6689
-1.3333

fval =
34.1365

exitflag =
1

output =
iterations: 19
funcCount: 28
constrviolation: 0
stepsize: 4.0372e-05
algorithm: 'interior-point'
firstorderopt: 4.0015e-07
cgiterations: 55
message: 'Local minimum found that satisfies the constraints.

Optimizati...'```

Even though the initial positions of the electrons were random, the final positions are nearly symmetric:

Compare to Optimization Without Gradients and Hessians

The use of gradients and Hessians makes the optimization run faster and more accurately. To compare with the same optimization using no gradient or Hessian information, set the options not to use gradients and Hessians:

```options = optimoptions(@fmincon,'Algorithm','interior-point',...
'Display','final');
[xfinal2 fval2 exitflag2 output2] = fmincon(@demoenergy,Xinitial,...
A,b,[],[],[],[],@democonstr,options)```

The output shows that fmincon found an equivalent minimum, but took more iterations and many more function evaluations to do so:

```Local minimum found that satisfies the constraints.

Optimization completed because the objective function is non-decreasing in
feasible directions, to within the default value of the function tolerance,
and constraints are satisfied to within the default value of the constraint tolerance.

xfinal2 =
0.0000
1.0000
-1.0000
0.6689
-0.6644
-1.3334
-0.6644
0.6689
-1.3334
0.0000
-1.0000
-1.0000
0.6357
0.6357
-1.4380
-0.0317
-0.0317
-1.9990
1.0000
0.0000
-1.0000
-1.0000
0.0000
-1.0000
0.0000
0.0000
-0.0000
-0.6667
-0.6667
-1.3334

fval2 =
34.1365

exitflag2 =
1

output2 =
iterations: 77
funcCount: 2431
constrviolation: 0
stepsize: 6.0588e-07
algorithm: 'interior-point'
firstorderopt: 2.9894e-06
cgiterations: 0
message: 'Local minimum found that satisfies the constraints.

Optimizati...'```

In this run the number of function evaluations (in output2.funcCount) is 2431, compared to 28 (in output.funcCount) when using gradients and Hessian.

Clear the Symbolic Variable Assumptions

The symbolic variables in this example have the assumption, in the symbolic engine workspace, that they are real. To clear this assumption from the symbolic engine workspace, it is not sufficient to delete the variables. You must clear the variables using the syntax

`syms x11 x12 x13 clear `

or reset the symbolic engine using the command

`reset(symengine)`

After resetting the symbolic engine you should clear all symbolic variables from the MATLAB workspace with the clear command, or clear variable_list.