Penalty Functions and Constrained Optimization

Kurt Bryan and Yosi Shibberu

Introduction
Weve had a pretty thorough introduction to unconstrained optimization. Now well
briefly consider constrained optimization. This is a rather more difficult subject. It makes use
of more heavy duty linear algebra and geometry, and the results and algorithms are generally
more complicated and technical. As such, well really just look at one basic technique, that
has the advantage of turning constrained problems into unconstrained problems! We can
then attack the problem using any of the unconstrained methods weve developed.
For motivation, consider minimizing the function f (x) = x4 subject to the constraint
x 1 (and ignore the fact that this is easy). The algorithms weve seen so far are strictly for
unconstrained optimization problems, and in this case that wont suffice; the constraint that
x 1 clearly influences the location of the minimum. The method of penalty functions provides a way to attack constrained optimization problems using algorithms for unconstrained
problems.
Heres how it works, via the example above. Define a function (t) by
(

(t) =

0
for t < 0
3
kt for t 0

(1)

where k is some positive constant. The function is a penalty function. It penalizes any
number t which is greater than zero (from the point of view of minimization). Heres a plot
of for k = 100:
800

600

400

200

1
t

The penalty function is also twice differentiable, even through zero.

The penalty function lets us attack the constrained problem by turning it into an unconstrained problem, as follows. The constraint x 1 is equivalent to 1 x 0. Define a
modified or penalized objective function
f(x) = f (x) + (1 x) = x4 + (1 x).
The function f(x) is identical to f if 1 x 0, i.e., if x 1, but rises sharply if x < 1. The
additional (1 x) term penalizes an optimization algorithm for choosing x < 1. Heres a
plot of just the penalty term, (1 x), with k = 100:
2500
2000
1500
1000
500

1
t

Heres a plot of f (x) and f(x), on a more compressed vertical scale:

20
18
16
14
12
y10
8
6
4
2
0

0.5

1.5
t

2.5

We can approximately minimize f (x) subject to the constraint x 1 by running an

unconstrained algorithm on the penalized objective function f(x); the penalty term will
strongly encourage the unconstrained algorithm to choose the best x which is greater than
or equal to one. The penalty term is also twice differentiable, so it should not cause any
trouble in an optimization algorithm which relies on first or second derivatives. The first
2

and second derivatives of (t) are just

(
0

(t) =

0
for t < 0
2
3kt for t 0

00

(t) =

0
for t < 0
6kt for t 0

If you run an unconstrained algorithm like golden section on f(x) in this case (with k = 100)
you find that the minimum is at x = 0.9012; the penalty approach didnt exactly solve the
problem, but it is reasonably close.
In fact, a reasonable procedure would be to increase the constant k, say by a factor of
10, and then re-run the unconstrained algorithm on f using 0.9012 as the initial guess. Increasing k enforces the constrained more rigorously, while using the previous final iterate as
an initial guess speeds up convergence (since we expect the minimum for the larger value of
k isnt that far from the minimum for the previous value of k). In this case increasing k to
104 moves the minimum to x = 0.989. We could then increase k and use x = 0.989 as an
initial guess, and continue this process until we obtain a reasonable estimate of the minimizer.
The General Case:
In general we want to minimize a function f (x) of n variables subject to both equality
and inequality constraints of the form
gi (x) 0, i = 1, . . . , m
hi (x) = 0, i = 1, . . . , n.

(2)
(3)

You should convince yourself that any equality or inequality constraints can be cast into the
above forms. The set of x in n dimensional space which satisfy the constraints is called the
feasible set, although it may be empty if the constraints are mutually contradictory.
We will call (, t) for 0, t lR, a penalty function if
1. is continuous.
2. (, t) 0 for all and t.
3. (, t) = 0 for t 0 and is strictly increasing for both > 0 and t > 0.
Its also desirable if has at least one continuous derivative in t, preferably two.
A typical example of a penalty function would be
(

(, t) =

0
for t < 0
tn for t 0

(4)

where n 1. This function has n 1 continuous derivatives in t, so taking n = 3 yields a

C 2 penalty function.
3

To minimize f (x) subject to constraints (2) and (3), define a modified objective function
by
f(x) = f (x) +

m
X
i=1

(i , gi (x)) +

n
X

((i , hi (x)) + (i , hi (x)))

i=1

where the i and i are positive constants that control how strongly the constraints will be
enforced. The penalty functions in the first sum modify the original objective function so
that if any inequality constraint is violated, a large penalty is invoked; if all constraints are
satisfied, no penalty. Similarly the second summation penalizes equality constraints which
are not satisfied, by penalizing both hi (x) < 0 and hi (x) > 0. We minimize the function f
with no constraints, and count on the penalty terms to keep the solution near the feasible set,
although no finite choice for the penalty parameters typically keeps the solution in the feasible set. After having minimized f with an unconstrained method, for a given set of i and
i , we may then increase the i and i and use the terminal iterate as the initial guess for a
new minimization, and continue this process until we obtain a sufficiently accurate minimum.
Example: Let us minimize the function f (x, y) = x2 + y 2 subject to the inequality constraint x + 2y 6 and the equality constraint x y = 3. In this case the constraints can by
written as
g1 (x, y) 0,
h1 (x, y) = 0,
where g1 (x, y) = 6 x 2y and h1 (x, y) = 3 x + y. We use the penalty function defined
in equation (4) with 1 = 5 and 1 = 5 to start. The modified objective function is
f(x, y) = f (x, y) + (5, g1 (x, y)) + (5, h1 (x, y)) + (5, h1 (x, y)).
Run any standard unconstrained algorithm on this, e.g., a BFGS quasi-Newton method;
the minimum occurs at x = 3.506 and y = 1.001. The equality constraint is violated
(3 x + y = 0.494), as is the inequality constraint (6 x 2y = 0.449 > 0). To increase
the accuracy with which the constraints are enforced, increase the penalty parameters. It is
very helpful to use the final estimate from the more modest penalty parameters as the initial
guess for the larger parameters. With 1 = 1 = 50 we obtain x = 3.836 and y = 1.008.
Increasing 1 = 1 = 500 we obtain x = 3.947, y = 1.003. The actual answer is x = 4, y = 1.
Increasing the penalty parameters does improve the accuracy of the final answer, but it will
also slow down the unconstrained algorithms convergence, for f(x) will then have a very
large gradient and the algorithm will spend a lot of time hunting for an accurate minimum.
Under appropriate assumptions one can prove that as the penalty parameters are increased without bound, any convergent subsequence of solutions to the unconstrained penalized problems must converge to a solution of the original constrained problem.

Pros and Cons of Penalty Functions

The obvious advantage to the penalty function approach is that we obtain a hands-off
method for converting constrained problems of any type into unconstrained problems. Also,
we dont have to worry about finding an initial feasible point (sometimes a problem).
Another advantage to the penalty function approach is that (in my humble experience)
many constraints in the real world are soft, in the sense that they need not be satisfied
precisely. The penalty function approach is well-suited to this type of problem.
The drawback to penalty function methods is that the solution to the unconstrained
penalized problem will not be an exact solution to the original problem (except in the
limit as described above). In some cases penalty methods cant be applied because the
objective function is actually undefined outside the feasible set. Ive worked on problems
like this, where computing the objective function involves solving a PDE on some region
and the independent variables control the geometry of the region. Infeasible values of the
independent variables correspond to geometrically impossible shapes!
Another drawback to penalty methods is that as we increase the penalty parameters
to more strictly enforce the constraints, the unconstrained formulation becomes very illconditioned, with large gradients and abrupt function changes. It also turns out that there
are more efficient (but more elaborate and difficult) methods for approaching constrained
optimization problems, but they are beyond what well cover in this course.
Barrier Function Methods
These are closely related to penalty function methods, and in fact might as well be
considered a type of penalty function method. These methods are generally applicable only
to inequality constrained optimization problems. Barrier methods have the advantage that
they always maintain feasible iterates, unlike the penalty methods above.
The most common is the log barrier method. Suppose we have an objective function f (x)
with inequality constraints gi (x) 0 for 1 i m. Form a modified or penalized objective
function
m
X
f(x) = f (x)
ri ln(gi (x))
i=1

where the ri > 0. Notice that f(x) is undefined if any gi (x) 0, so we can only evaluate
f in the interior of the feasible region. However, even inside the feasible region the penalty
term is non-zero (but it becomes an anti-penalty if gi 1).
Suppose we start some choice for the ri and with initial feasible point x0 , and minimize
f. The terminal point xk , must be a feasible point, because the log terms in the definition
of f form a barrier of infinite height which prevents the optimization routine from leaving
the interior of the feasible region.
Example: Heres a 1D example with objective function f (t) = t4 and constraint t 1.
5

The penalized objective function is f(t) = t4 2 ln(t 1) (so I took r1 = 2):

80

60

40

20

0 1 1.2 1.4 1.6 1.8

2
t

2.2 2.4 2.6 2.8

In general a barrier method works in a similar way to the penalty methods above. We start
with some positive ri and feasible point x0 . Minimize f using an unconstrained algorithm.
Now decrease the value of the ri and re-optimize, using the final iterate as an initial guess
for the newly decreased ri . Continue until an acceptable minimum is found.
One point on which we need to be careful is the line searchyou dont want to evaluate
f at any point outside the feasible set (or at least you need to deal with this gracefully).
Example: Let f (x, y) = x2 + y 2 . We want to minimize f subject to 6 x 2y 0.
If we take r1 = 5 in the definition of f (so f(x, y) = x2 + y 2 5 ln(x + 2y 6)) and start
with feasible point (5, 5) we obtain a minimum at (1.53, 3.05). Decreasing r1 to 0.5 gives a
minimum at (1.24, 2.48), and decreasing r1 to 0.05 gives a minimum at (1.204, 2.408) (the
true minimum is at (1.2, 2.4)).

One issue in using a barrier method is that of finding an initial feasible point which
is in the interior of the feasible region. In many cases such a point will be obvious from
considerations specific to the problem. If not, it can be rather difficult to find such a point
(or perhaps prove that the feasible region is in fact empty if the constraints are mutually
exclusive). One idea would be to use penalty functions, but on constraints gi (x) < 0
with f 0. If a solution a to this can be found with f(a) = 0 then a is a feasible point
which is in the interior of the feasible region defined by gi (x) 0.