Constraints and Lagrange Multipliers.

Physics 6010, Fall 2016

Constraints and Lagrange Multipliers.
Relevant Sections in Text: §1.3–1.6

Constraints
Often times we consider dynamical systems which are defined using some kind of
restrictions on the motion. For example, the spherical pendulum can be defined as a
particle moving in 3-d such that its distance from a given point is fixed. Thus the true
configuration space is defined by giving a simpler (usually bigger) configuration space along
with some constraints which restrict the motion to some subspace. Constraints provide a
phenomenological way to account for a variety of interactions between systems. We now
give a systematic treatment of this idea and show how to handle it using the Lagrangian
formalism.
For simplicity we will only consider holonomic constraints, which are restrictions which
can be expressed in the form of the vanishing of some set of functions – the constraints –
on the configuration space and time:

Cα (q, t) = 0, α = 1, 2, . . . m.

We assume these functions are smooth and independent so that if there are n coordinates
q i , then at each time t the constraints restrict the motion to a nice n − m dimensional
space. For example, the spherical pendulum has a single constraint on the three Cartesian
configruation variables (x, y, z):

C(x, y, z) = x2 + y 2 + z 2 − l2 = 0.

This constraint restricts the configuration to a two dimensional sphere of radius l centered
at the origin. To see another example of such constraints, see our previous discussion of
the double pendulum and pendulum with moving point of support.
We note that the constraints will restrict the velocities:
d ∂Cα i ∂Cα
Cα = q̇ + = 0.
dt ∂q i ∂t
For example in the spherical pendulum we have

xẋ + y ẏ + z ż = 0.

There are two ways to deal with such constraints. Firstly, one can simply solve the
constraints, i.e., find an independent set of generalized coordinates. We have been doing
1
 Constraints and Lagrange Multipliers.

this all along in our examples with constraints. For the spherical pendulum, we solve the
constraint by
x = l sin θ cos φ, y = l sin θ sin φ, z = l cos θ,

and express everything in terms of θ and φ, in particular the Lagrangian and EL equations.
In principle this can always be done, but in practice this might be difficult. There is another
method in which one can find the equations of motion without having to explicitly solve
the constraints. This is known as the method of Lagrange multipliers. This method is
not just popular in mechanics, but also features in “constrained optimization” problems,
e.g., in economics. As we shall see, the Lagrange multiplier method is more than just an
alternative approach to constraints – it provides additional physical information about the
forces which maintain the constraints.

Lagrange Multipliers

The method of Lagrange multipliers in the calculus of variations has an analog in

ordinary calculus. Suppose we are trying to find the critical points of a function f (x, y)
subject to a constraint C(x, y) = 0. That is to say, we want to find where on the curve
defined by the constraint the function has a maximum, minimum, saddle point. Again, we
could try to solve the constraint, getting a solution of the form y = g(x). Then we could
substitute this into the function f to get a (new) function h(x) = f (x, g(x)). Then we find
the critical points by solving h0 (x) = 0 for x = x0 whence the critical point is (x0 , g(x0 ))
This is analogous to our treatment of constraints in the variational calculus thus far (where
we solved the constraints via generalized coordinates before constructing the Lagrangian
and EL equations). There is another method, due to Lagrange, which does not require
explicit solution of the constraints and which gives useful physical information about the
constraints.
To begin with, when finding a critical point (x0 , y0 ) subject to the constraint C(x, y) =
0 we are looking for a point on the curve C(x, y) = 0 such that a displacement tangent
to the curve does not change the value of f to first order. Let the tangent vector to
C(x, y) = 0 at the point (x0 , y0 ) on the curve be denoted by ~t. We want

~t · ∇f (x0 , y0 ) = 0 where C(x0 , y0 ) = 0.

Evidently, at the critical point the gradient of f is orthogonal to the curve C(x, y) = 0.
Now, any vector orthogonal to the curve – orthogonal to ~t at (x0 , y0 ) – will be proportional
to the gradient of C at that point.* Thus the condition for a critical point (x0 , y0 ) of f

* This follows from the basic calculus result that the gradient of a function is orthogonal to
the locus of points where the function takes a constant value.
2
 Constraints and Lagrange Multipliers.

(where C(x0 , y0 ) = 0) is that the gradient of f and the gradient of C are proportional at
(x0 , y0 ). We write
∇f + λ∇C = 0, where C(x0 , y0 ) = 0
for some λ. This requirement is meant to hold only on the curve C = 0, so without loss of
generality we can take as the critical point condition

∇(f + λC) = 0, C = 0.

This constitutes three conditions on 3 unknowns; the unknowns being (x, y) and λ. The
function λ is known as a Lagrange multiplier. In fact, if we artificially enlarge our x-y plane
to a 3-d space parametrized by (x, y, λ) we can replace the above critical point condition
with
˜ + λC) = 0,
∇(f
˜ is the gradient in (x, y, λ) space. You should prove this as an exercise.
where ∇
To summarize: the critical points (x0 , y0 ) of a function f (x, y) constrained to a curve
C(x, y) = 0 can be obtained by finding unconstrained critical points (x0 , y0 , λ0 ) of a
function in the space of variables (x, y, λ):

f˜(x, y, λ) = f (x, y) + λC(x, y).

We can do the same thing with our variational principle. Suppose we have an action
for n degrees of freedom q i , i = 1, 2, . . . , n:
Z t2
S[q] = dt L(q(t), q̇(t), t)
t1

where the configuation space is subject to m constraints

Cα (q, t) = 0, α = 1, 2, . . . , m.

Let the solutions to the constraints be given in terms of generalized coordinates sA , A =

1, 2, . . . , n − m,
q i = F i (s, t),
i.e.,
Cα (F i (s, t), t) = 0.
The functions F i determine the graph of the solution set of Cα = 0 in the configuration
space. The correct equations of motion can be obtained by substituting the solutions
q i = F i (s, t) into the Lagrangian for q i , thus defining a Lagrangian for sA , and computing
the resulting EL equations for sA . Using the same technology you used in your homework
3
 Constraints and Lagrange Multipliers.

to study the effect of a point transformation on the EL equations (principally the chain
rule), it is not hard to see that the correct equations of motion are then
∂F i ∂L
 
d ∂L
− = 0.
∂sA ∂q i dt ∂ q̇ i q=F (s,t)
i
∂F have the geometric meaning of (a basis of) (n − m) tangent
We note that the functions ∂sA

vectors to the (n − m)-dimensional surface Cα = 0. Thus the equations of motion are the
statement that the projections of the EL equations along the surface must vanish.
Now we introduce the Lagrange multiplier method. We consider a modified action,
Z t2 Z t2
S̃[q, λ] = dt L̃ = S[q] + dt λα (t)Cα (q(t), t),
t1 t1

in which we have added m new configuration variables λα , α = 1, 2, . . . , m; these are the

Lagrange multipliers. The variation of the new action is
Z t2   Z t2  
∂L d ∂L i α α ∂Cα i
δ S̃ = dt − δq + dt δλ Cα + λ δq .
t1 ∂q i dt ∂ q̇ i t1 ∂q i

The EL equations of motion coming from L̃ are

∂L d ∂L ∂Cα
i
− i
+ λα i = 0,
∂q dt ∂ q̇ ∂q
which come from the variations in q i and also

Cα = 0,

which come from variations of λα . We have (n + m) equations for (n + m) unknowns. In

principle they can be solved to get the q i and the λα as functions of t.
What is the meaning of these equations? Well, the constraints are there, of course.
But what about the modified EL expressions? The EL equations you would have gotten
from L now have a “force term”, λα ∂C α
∂q i
. The force term is geometrically orthogonal to
the surface Cα = 0 in configuration space. This you can see from the identity (exercise)
∂ ∂Cα ∂F i
0= Cα (F (s), t) = .
∂sA ∂q i ∂sA
i
∂F represend n − m vectors tangent to the surface defined by C = 0.) Thus
(Recall that ∂sA α
the meaning of the EL equations coming from L̃ is that the EL expressions coming from
L no longer have to vanish, they simply have to be orthogonal to the constraint surface
since the equations of motion say that
∂L d ∂L ∂Cα
i
− i
= −λα i .
∂q dt ∂ q̇ ∂q
4
 Constraints and Lagrange Multipliers.

One physically interprets this “force term” as the force required to keep the motion on
this surface.
It is easy to verify that these modified equations, (n + m) in number, are equivalent
the correct (n − m) equations obtained for sA earlier. Indeed, we have the m equations
of constraint. And, given this constraint, to say the EL expression coming from L is
orthogonal to Cα = 0 is the same as saying its projection tangent to the surface vanishes,
i.e.,
∂F i ∂L ∂F i
   
d ∂L α ∂Cα
− = −λ = 0,
∂sA ∂q i dt ∂ q̇ i q=F (s,t) ∂sA ∂q i
which is precisely the content of the equations for the sA we obtained above.

Example: Plane pendulum revisited

Let us study the plane pendulum using Lagrange multipliers. We model the system as
moving in a plane with coordinates (x, y) subject the constraint

C = x2 + y 2 − l2 = 0.

Without the constraint the Lagrangian would be simply

1
L = m(ẋ2 + ẏ 2 ) − mgy.
2
According to our general prescription for incorporating the constraint, we construct the
modified Lagrangian
1
L̃ = m(ẋ2 + ẏ 2 ) − mgy + λ(x2 + y 2 − l2 ).
2
The critical points for the action built from L̃, with the configuration space parametrized
by (x, y, λ), should give us the critical points along the surface C = 0. To find the critical
points we construct the EL equations as usual. We get

x2 + y 2 − l2 = 0,

coming from the variation of λ, and

2λx − mẍ = 0, 2λy − mg − mÿ = 0,

coming from the variations of x and y, respectively.

Here we can see more explicitly how the Lagrange multiplier defines a force term
beyond the gravitational force. This “force of constraint” represents the force of the rigid
pendulum arm upon the particle and is given by

F~constraint = 2λxx̂ + 2λy ŷ.

5
 Constraints and Lagrange Multipliers.

The typical analysis of EL equations involving Lagrange multipliers can now be nicely
demonstrated. First, the three EL equations can be solved for λ (exercise)
m
λ= (xẍ + y ÿ + gy) .
2l2
Next, differentiation of the constraint twice reveals:

C̈ = 0 =⇒ xẍ + y ÿ = −(ẋ2 + ẏ 2 ),

so that the multiplier λ can be solved for in terms of the original velocity phase space
variables:
m
λ = − 2 (ẋ2 + ẏ 2 − gy).
2l
Substituting this result back into the EL equations for x and y we get the equations of
motion for x and y with the effect of the constraint — physically, the tension in the rod
— taken into account:
m m
mẍ = − 2 (ẋ2 + ẏ 2 − gy)x, mÿ = − 2 (ẋ2 + ẏ 2 − gy)y − mg.
l l

Note we never had to solve the constraint! Still, as a nice exercise you can check that,
after solving the constraint with x = l cos φ, y = −l sin φ, these remaining 2 equations are
equivalent the familiar equation of motion for a plane pendulum, namely,
g
φ̈ = − sin φ,
l
where φ is the angular displacement from equilibrium.
Using Lagrangian multipliers, the equations of motion for x and y tell us that the
pendulum moves according to a superposition of forces consisting of (i) gravity, (ii) the
force of constraint F~constraint needed to keep the mass moving in a circle of radius l.
This latter force is supplied by the Lagrange multiplier terms in the equation of motion.
Indeed, thanks to these Lagrange multiplier terms, the radial component of the net force
is (exercise)
~r ~ m
· F = − (ẋ2 + ẏ 2 ),
l l
which is the centripetal force, as it should be.
To summarize: Given a dynamical system with coordinates q i and Lagrangian L, we
can impose constraints Cα (q, t) = 0 by the following recipe.
(i) Add variables λα – the Lagrange multipliers – to the configuration space,
(ii) Define a Lagrangian on the augmented velocity phase space L̃ = L + λα Cα ,
(iii) Compute the usual EL equations from L̃ for the q i and λα degrees of freedom.
6
 Constraints and Lagrange Multipliers.

The resulting equations will include the constraints themselves as equations of motion
coming from variations of λα . The equations coming from the variations of the q i will
have extra terms involving the multipliers. For Newtonian systems these terms represent
the forces in the system which are necessary to enforce the constraints. If desired, one can
use the equations of motion, the constraints, and the time derivatives of the constraints
to solve for the multipliers in terms of the velocity phase space. One can then reduce the
original equations to only be built from the original degrees of freedom.
Thus the Lagrange multiplier method has distinct advantages over our previous ap-
proach in which we just solve the constraints at the beginning.. Namely, you do not have
to explcitly solve the constraints in order to compute the equations of motion, and the
equations of motion have additional physical information: the forces of constraint.