# A quick start¶

Once you have installed CVX (see Installation), you can start using it by entering a CVX specification into a Matlab script or function, or directly from the command prompt. To delineate CVX specifications from surrounding Matlab code, they are preceded with the statement cvx_begin and followed with the statement cvx_end. A specification can include any ordinary Matlab statements, as well as special CVX-specific commands for declaring primal and dual optimization variables and specifying constraints and objective functions.

Within a CVX specification, optimization variables have no numerical value; instead, they are special Matlab objects. This enables Matlab to distinguish between ordinary commands and CVX objective functions and constraints. As CVX reads a problem specification, it builds an internal representation of the optimization problem. If it encounters a violation of the rules of disciplined convex programming (such as an invalid use of a composition rule or an invalid constraint), an error message is generated. When Matlab reaches the cvx_end command, it completes the conversion of the CVX specification to a canonical form, and calls the underlying core solver to solve it.

If the optimization is successful, the optimization variables declared in the CVX specification are converted from objects to ordinary Matlab numerical values that can be used in any further Matlab calculations. In addition, CVX also assigns a few other related Matlab variables. One, for example, gives the status of the problem (i.e., whether an optimal solution was found, or the problem was determined to be infeasible or unbounded). Another gives the optimal value of the problem. Dual variables can also be assigned.

This processing flow will become clearer as we introduce a number of simple examples. We invite the reader to actually follow along with these examples in Matlab, by running the quickstart script found in the examples subdirectory of the CVX distribution. For example, if you are on Windows, and you have installed the CVX distribution in the directory D:\Matlab\cvx, then you would type

cd D:\Matlab\cvx\examples
quickstart


at the Matlab command prompt. The script will automatically print key excerpts of its code, and pause periodically so you can examine its output. (Pressing “Enter” or “Return” resumes progress.)

## Least squares¶

We first consider the most basic convex optimization problem, least-squares (also known as linear regression). In a least-squares problem, we seek $$x \in \mathbf{R}^n$$ that minimizes $$\|Ax-b\|_2$$, where $$A\in \mathbf{R}^{m \times n}$$ is skinny and full rank (i.e., $$m\geq n$$ and $$\operatorname*{\textbf{Rank}}(A)=n$$). Let us create the data for a small test problem in Matlab:

m = 16; n = 8;
A = randn(m,n);
b = randn(m,1);


Then the least-squares solution $$x=(A^TA)^{-1}A^Tb$$ is easily computed using the backslash operator:

x_ls = A \ b;


Using CVX, the same problem can be solved as follows:

cvx_begin
variable x(n)
minimize( norm(A*x-b) )
cvx_end


(The indentation is used for purely stylistic reasons and is optional.) Let us examine this specification line by line:

• cvx_begin creates a placeholder for the new CVX specification, and prepares Matlab to accept variable declarations, constraints, an objective function, and so forth.
• variable x(n) declares x to be an optimization variable of dimension $$n$$. CVX requires that all problem variables be declared before they are used in the objective function or constraints.
• minimize( norm(A*x-b) ) specifies the objective function to be minimized.
• cvx_end signals the end of the CVX specification, and causes the problem to be solved.

Clearly there is no reason to use CVX to solve a simple least-squares problem. But this example serves as sort of a “Hello world!” program in CVX; i.e., the simplest code segment that actually does something useful.

When Matlab reaches the cvx_end command, the least-squares problem is solved, and the Matlab variable x is overwritten with the solution of the least-squares problem, i.e., $$(A^TA)^{-1}A^Tb$$. Now x is an ordinary length-$$n$$ numerical vector, identical to what would be obtained in the traditional approach, at least to within the accuracy of the solver. In addition, several additional Matlab variables are created; for instance,

• cvx_optval contains the value of the objective function;
• cvx_status contains a string describing the status of the calculation (see Interpreting the results).

All of these quantities—x, cvx_optval, and cvx_status, etc.—may now be freely used in other Matlab statements, just like any other numeric or string values. [1]

There is not much room for error in specifying a simple least-squares problem, but if you make one, you will get an error or warning message. For example, if you replace the objective function with

maximize( norm(A*x-b) );


which asks for the norm to be maximized, you will get an error message stating that a convex function cannot be maximized (at least in disciplined convex programming):

??? Error using ==> maximize
Disciplined convex programming error:
Objective function in a maximization must be concave.


## Bound-constrained least squares¶

Suppose we wish to add some simple upper and lower bounds to the least-squares problem above: i.e.,

$\begin{split}\begin{array}{ll} \mbox{minimize} & \|Ax-b\|_2\\ \mbox{subject to} & l \preceq x \preceq u \end{array}\end{split}$

where $$l$$ and $$u$$ are given data vectors with the same dimension as $$x$$. The vector inequality $$u \preceq v$$ means componentwise, i.e., $$u_i \leq v_i$$ for all $$i$$. We can no longer use the simple backslash notation to solve this problem, but it can be transformed into a quadratic program (QP) which can be solved without difficulty with a standard QP solver. [2]

Let us provide some numeric values for l and u:

bnds = randn(n,2);
l = min( bnds, [], 2 );
u = max( bnds, [], 2 );


If you have the Matlab Optimization Toolbox, you can use quadprog to solve the problem as follows:

x_qp = quadprog( 2*A'*A, -2*A'*b, [], [], [], [], l, u );


This actually minimizes the square of the norm, which is the same as minimizing the norm itself. In contrast, the CVX specification is given by

cvx_begin
variable x(n)
minimize( norm(A*x-b) )
subject to
l <= x <= u
cvx_end


Two new lines of CVX code have been added to the CVX specification:

• The subject to statement does nothing—CVX provides this statement simply to make specifications more readable. As with indentation, it is optional.
• The line l <= x <= u represents the $$2n$$ inequality constraints.

As before, when the cvx_end command is reached, the problem is solved, and the numerical solution is assigned to the variable x. Incidentally, CVX will not transform this problem into a QP by squaring the objective; instead, it will transform it into an SOCP. The result is the same, and the transformation is done automatically.

In this example, as in our first, the CVX specification is longer than the Matlab alternative. On the other hand, it is easier to read the CVX version and relate it to the original problem. In contrast, the quadprog version requires us to know in advance the transformation to QP form, including the calculations such as 2*A'*A and -2*A'*b. For all but the simplest cases, a CVX specification is simpler, more readable, and more compact than equivalent Matlab code to solve the same problem.

## Other norms and functions¶

Now let us consider some alternatives to the least-squares problem. Norm minimization problems involving the $$\ell_\infty$$ or $$\ell_1$$ norms can be reformulated as LPs, and solved using a linear programming solver such as linprog in the Matlab Optimization Toolbox; see, e.g., Section 6.1 of Convex Optimization. However, because these norms are part of CVX’s base library of functions, CVX can handle these problems directly.

For example, to find the value of $$x$$ that minimizes the Chebyshev norm $$\|Ax-b\|_\infty$$, we can employ the linprog command from the Matlab Optimization Toolbox:

f    = [ zeros(n,1); 1          ];
Ane  = [ +A,         -ones(m,1)  ; ...
-A,         -ones(m,1) ];
bne  = [ +b;         -b         ];
xt   = linprog(f,Ane,bne);
x_cheb = xt(1:n,:);


With CVX, the same problem is specified as follows:

cvx_begin
variable x(n)
minimize( norm(A*x-b,Inf) )
cvx_end


The code based on linprog, and the CVX specification above will both solve the Chebyshev norm minimization problem, i.e., each will produce an $$x$$ that minimizes $$\|Ax-b\|_\infty$$. Chebyshev norm minimization problems can have multiple optimal points, however, so the particular $$x$$‘s produced by the two methods can be different. The two points, however, must have the same value of $$\|Ax-b\|_\infty$$.

Similarly, to minimize the $$\ell_1$$ norm $$\|\cdot\|_1$$, we can use linprog as follows:

f    = [ zeros(n,1); ones(m,1);  ones(m,1)  ];
Aeq  = [ A,          -eye(m),    +eye(m)    ];
lb   = [ -Inf(n,1);  zeros(m,1); zeros(m,1) ];
xzz  = linprog(f,[],[],Aeq,b,lb,[]);
x_l1 = xzz(1:n,:) - xzz(n+1:end,:);


The CVX version is, not surprisingly,

cvx_begin
variable x(n)
minimize( norm(A*x-b,1) )
cvx_end


CVX automatically transforms both of these problems into LPs, not unlike those generated manually for linprog.

The advantage that automatic transformation provides is magnified if we consider functions (and their resulting transformations) that are less well-known than the $$\ell_\infty$$ and $$\ell_1$$ norms. For example, consider the norm

$\| Ax-b\|_{\mathrm{lgst},k} = |Ax-b|_{[1]}+ \cdots + |Ax-b|_{[k]},$

where $$|Ax-b|_{[i]}$$ denotes the $$i$$th largest element of the absolute values of the entries of $$Ax-b$$. This is indeed a norm, albeit a fairly esoteric one. (When $$k=1$$, it reduces to the $$\ell_\infty$$ norm; when $$k=m$$, the dimension of $$Ax-b$$, it reduces to the $$\ell_1$$ norm.) The problem of minimizing $$\| Ax-b\|_{\mathrm{lgst},k}$$ over $$x$$ can be cast as an LP, but the transformation is by no means obvious so we will omit it here. But this norm is provided in the base CVX library, and has the name norm_largest, so to specify and solve the problem using CVX is easy:

k = 5;
cvx_begin
variable x(n);
minimize( norm_largest(A*x-b,k) );
cvx_end


Unlike the $$\ell_1$$, $$\ell_2$$, or $$\ell_\infty$$ norms, this norm is not part of the standard Matlab distribution. Once you have installed CVX, though, the norm is available as an ordinary Matlab function outside a CVX specification. For example, once the code above is processed, x is a numerical vector, so we can type

cvx_optval
norm_largest(A*x-b,k)


The first line displays the optimal value as determined by CVX; the second recomputes the same value from the optimal vector x as determined by CVX.

The list of supported nonlinear functions in CVX goes well beyond norm and norm_largest. For example, consider the Huber penalty minimization problem

$\begin{split}\begin{array}{ll} \text{minimize} & \sum_{i=1}^m \phi( (Ax-b)_i ) \end{array},\end{split}$

with variable $$x \in \mathbf{R}^n$$, where $$\phi$$ is the Huber penalty function

$\begin{split}\phi(z) = \begin{cases} |z|^2 & |z|\leq 1 \\ 2|z|-1 & |z|\geq 1\end{cases}.\end{split}$

The Huber penalty function is convex, and has been provided in the CVX function library. So solving the Huber penalty minimization problem in CVX is simple:

cvx_begin
variable x(n);
minimize( sum(huber(A*x-b)) );
cvx_end


CVX automatically transforms this problem into an SOCP, which the core solver then solves. (The CVX user, however, does not need to know how the transformation is carried out.)

## Other constraints¶

We hope that, by now, it is not surprising that adding the simple bounds $$l\preceq x\preceq u$$ to the problems above is as simple as inserting the line l <= x <= u before the cvx_end statement in each CVX specification. In fact, CVX supports more complex constraints as well. For example, let us define new matrices C and d in Matlab as follows,

p = 4;
C = randn(p,n);
d = randn(p,1);


Now let us add an equality constraint and a nonlinear inequality constraint to the original least-squares problem:

cvx_begin
variable x(n);
minimize( norm(A*x-b) );
subject to
C*x == d;
norm(x,Inf) <= 1;
cvx_end


Both of the added constraints conform to the DCP rules, and so are accepted by CVX. After the cvx_end command, CVX converts this problem to an SOCP, and solves it.

Expressions using comparison operators (==, >=, etc.) behave quite differently when they involve CVX optimization variables, or expressions constructed from CVX optimization variables, than when they involve simple numeric values. For example, because x is a declared variable, the expression C*x==d causes a constraint to be included in the CVX specification, and returns no value at all. On the other hand, outside of a CVX specification, if x has an appropriate numeric value—for example immediately after the cvx_end command—that same expression would return a vector of 1s and 0s, corresponding to the truth or falsity of each equality. [3] Likewise, within a CVX specification, the statement norm(x,Inf)<=1 adds a nonlinear constraint to the specification; outside of it, it returns a 1 or a 0 depending on the numeric value of x (specifically, whether its $$\ell_\infty$$-norm is less than or equal to, or more than, $$1$$).

Because CVX is designed to support convex optimization, it must be able to verify that problems are convex. To that end, CVX adopts certain rules that govern how constraint and objective expressions are constructed. For example, CVX requires that the left- and right-hand sides of an equality constraint be affine. So a constraint such as

norm(x,Inf) == 1;


results in the following error:

??? Error using ==> cvx.eq
Disciplined convex programming error:
Both sides of an equality constraint must be affine.


Inequality constraints of the form $$f(x) \leq g(x)$$ or $$g(x) \geq f(x)$$ are accepted only if $$f$$ can be verified as convex and $$g$$ verified as concave. So a constraint such as

norm(x,Inf) >= 1;


results in the following error:

??? Error using ==> cvx.ge
Disciplined convex programming error:
The left-hand side of a ">=" inequality must be concave.


The specifics of the construction rules are discussed in more detail in The DCP ruleset. These rules are relatively intuitive if you know the basics of convex analysis and convex optimization.

For our final example in this section, let us show how traditional Matlab code and CVX specifications can be mixed to form and solve multiple optimization problems. The following code solves the problem of minimizing $$\|Ax-b\|_2 +\gamma \|x\|_1$$, for a logarithmically spaced vector of (positive) values of $$\gamma$$. This gives us points on the optimal trade-off curve between $$\|Ax-b\|_2$$ and $$\|x\|_1$$. An example of this curve is given in the figure below.

gamma = logspace( -2, 2, 20 );
l2norm = zeros(size(gamma));
l1norm = zeros(size(gamma));
fprintf( 1, '   gamma       norm(x,1)    norm(A*x-b)\n' );
fprintf( 1, '---------------------------------------\n' );
for k = 1:length(gamma),
fprintf( 1, '%8.4e', gamma(k) );
cvx_begin
variable x(n);
minimize( norm(A*x-b)+gamma(k)*norm(x,1) );
cvx_end
l1norm(k) = norm(x,1);
l2norm(k) = norm(A*x-b);
fprintf( 1, '   %8.4e   %8.4e\n', l1norm(k), l2norm(k) );
end
plot( l1norm, l2norm );
xlabel( 'norm(x,1)' );
ylabel( 'norm(A*x-b)' );
grid on


The minimize statement above illustrates one of the construction rules to be discussed in The DCP ruleset. A basic principle of convex analysis is that a convex function can be multiplied by a nonnegative scalar, or added to another convex function, and the result is then convex. CVX recognizes such combinations and allows them to be used anywhere a simple convex function can be—such as an objective function to be minimized, or on the appropriate side of an inequality constraint. So in our example, the expression

norm(A*x-b)+gamma(k)*norm(x,1)


is recognized as convex by CVX, as long as gamma(k) is positive or zero. If gamma(k) were negative, then this expression becomes the sum of a convex term and a concave term, which causes CVX to generate the following error:

??? Error using ==> cvx.plus
Disciplined convex programming error:
Addition of convex and concave terms is forbidden.

 [1] If you type who or whos at the command prompt, you may see other, unfamiliar variables as well. Any variable that begins with the prefix cvx_ is reserved for internal use by CVX itself, and should not be changed.
 [2] There are also a number of solvers specifically designed to solve bound-constrained least-squares problems, such as BCLS by Michael Friedlander.
 [3] In fact, immediately after the cvx_end command above, you would likely find that most if not all of the values returned would be 0. This is because, as is the case with many numerical algorithms, solutions are determined only to within some nonzero numeric tolerance. So the equality constraints will be satisfied closely, but often not exactly.

← Installation

→ The Basics