Part III.

Calculus of variations
Lecture notes for MA342H

P. Karageorgis

[email protected]

1 / 43
Introduction

There are several applications that involve expressions of the form

Z b
J(y) = L(x, y(x), y ′ (x)) dx.
a

For instance, J(y) could represent area, length, energy and so on.
We are interested in finding the minimum/maximum values of J(y)
over all functions y(x) that satisfy some given constraints.
A function y(x) for which the minimum/maximum value is attained is
called an extremal of J(y). We shall assume that both y(x) and the
Lagrangian function L are twice continuously differentiable.
One may impose various constraints on y(x), but the most common
one amounts to specifying its values at the endpoints x = a, b. This
requires the graph of y(x) to pass through two given points.

2 / 43
Directional derivative

Consider a functional of the form

Z b
J(y) = L(x, y(x), y ′ (x)) dx.
a

Its directional derivative in the direction of a function φ is defined as

J(y + εφ) − J(y)

J ′ (y)φ = lim .
ε→0 ε
This is also known as the first variation of J(y). Explicitly, one has
Z b
′
Ly φ + Ly′ φ′ dx.


J (y)φ =
a

The function φ is called a test function, if it is twice continuously

differentiable and also vanishing at the endpoints x = a, b.

3 / 43
Directional derivative: Proof

The directional derivative J ′ (y)φ is defined in terms of

Z b 
J(y + εφ) − J(y) = L(x, y + εφ, y ′ + εφ′ ) − L(x, y, y ′ ) dx.
a

Denote the integrand by F (ε). Then a Taylor series expansion gives

F (ε) = Ly (x, y, y ′ ) · εφ + Ly′ (x, y, y ′ ) · εφ′ + . . . ,

where the dots indicate terms which contain higher powers of ε.

Once we now combine the last two equations, we find that

J(y + εφ) − J(y)

J ′ (y)φ = lim
ε→0 ε
Z b Z b
F (ε)
Ly φ + Ly′ φ′ dx.


= lim dx =
ε→0 a ε a

4 / 43
Euler-Lagrange equation

Theorem 1. Euler-Lagrange equation

Suppose that y(x) is an extremal of
Z b
J(y) = L(x, y(x), y ′ (x)) dx.
a

Then y(x) must satisfy the Euler-Lagrange equation

d
Ly′ = Ly .
dx

A solution of the Euler-Lagrange equation is also known as a critical

point of J(y). Critical points of J(y) are not necessarily extremals.
When the Lagrangian L depends on higher-order derivatives of y, the
Euler-Lagrange equation involves higher-order derivatives of L.

5 / 43
Euler-Lagrange equation: Proof

Since y(x) is an extremal of J(y), it must satisfy

Z b
′
Ly φ + Ly′ φ′ dx = 0


J (y)φ = 0 =⇒
a

for every test function φ. Integrating by parts, we now get

Z b 
d h ib
Ly − Ly′ φ dx + Ly′ φ = 0.
a dx a

Since the test function φ vanishes at the endpoints, this gives

Z b 
d
Ly − Ly φ dx = 0
′
a dx

for every test function φ. According to the fundamental lemma of

d
variational calculus, we must thus have Ly = dx Ly′ , as needed.
6 / 43
Fundamental lemma of variational calculus
Suppose that H(x) is continuously differentiable with
Z b
H(x)φ(x) dx = 0
a

for every test function φ. Then H(x) must be identically zero.

To prove this, consider an arbitrary subinterval [x1 , x2 ] and let

(x − x1 )3 (x2 − x)3
 
if x1 ≤ x ≤ x2
φ(x) = .
0 otherwise

Then φ is a test function which is positive on (x1 , x2 ) and we have

Z b Z x2
0= H(x)φ(x) dx = H(x)φ(x) dx.
a x1

If H(x) is positive/negative at a point, it must have the same sign on

some interval [x1 , x2 ] and this is not the case by the last equation.
7 / 43
Beltrami identity
Suppose that y(x) is a critical point of the functional
Z b
J(y) = L(y(x), y ′ (x)) dx
a

whose Lagrangian function L does not depend on x directly. Then

y ′ Ly′ − L = constant.

To prove this, one uses the Euler-Lagrange equation to find that

d d
y ′ Ly′ = y ′′ Ly′ + y ′ Ly′ = y ′′ Ly′ + y ′ Ly .


dx dx
Since L(y, y ′ ) does not depend on x directly, the chain rule gives

d d
y ′ Ly′ = y ′′ Ly′ + y ′ Ly =


L.
dx dx
8 / 43
Example 1. Shortest path

Consider a function y(x) whose graph passes through two given

points. We wish to minimise the length of its graph
Z bp
J(y) = 1 + y ′ (x)2 dx.
a
p
Since L = 1 + y ′ (x)2 , the Euler-Lagrange equation gives

d y ′ (x)
Ly′ = Ly = 0 =⇒ Ly′ = p = c1 .
dx 1 + y ′ (x)2

We square both sides and then simplify to conclude that

y ′ (x)2 = c21 + c21 y ′ (x)2 =⇒ y ′ (x) = c2 .

This shows that y(x) must be a linear function. In other words, the
shortest path between any two points is given by a line.
9 / 43
Example 2. Minimal surface, page 1

Consider a function y(x) whose graph passes through two given

points. We wish to minimise the surface area of the solid which is
obtained by rotating the graph of y(x) around the x-axis, namely
Z b p
J(y) = 2πy(x) 1 + y ′ (x)2 dx.
a

Since the Lagrangian does not depend on x directly, one has

2πy(x)y ′ (x)2
c = y ′ Ly′ − L = p
p
− 2πy(x) 1 + y ′ (x)2
1 + y ′ (x)2

by the Beltrami identity. It now easily follows that

c y(x) y(x)2
= −p =⇒ 1 + y ′ (x)2 = .
2π 1 + y ′ (x)2 a2

10 / 43
Example 2. Minimal surface, page 2

The last equation is actually a separable equation which gives

r
y2
Z Z
dy dy dx
= 2
− 1 =⇒ p = .
dx a 2
y −a 2 a

Letting y = a cosh t, we now get y 2 − a2 = a2 sinh2 t, hence also

Z Z Z
dx a sinh t dt
= = dt.
a a sinh t
Since y = a cosh t by above, we may finally conclude that
 
x − x0 x − x0
= t =⇒ y = a cosh t = a cosh .
a a

This is an equation that describes the shape of a hanging chain.

11 / 43
Example 3. Brachistochrone, page 1

Consider a particle that slides along the graph of a function y(x) from
one point to another under the influence of gravity. We assume that
there is no friction and wish to minimise the overall travel time
Z bp
1 + y ′ (x)2 dx
J(y) = .
a v(y(x))

To find the speed v, we note that conservation of energy gives

1 1
mv 2 + mgy = mv02 + mgy0 .
2 2
Assuming that v0 = 0 for simplicity, the overall travel time is then
Z bp
1 + y ′ (x)2 dx
J(y) = p .
a 2g(y0 − y)

12 / 43
Example 3. Brachistochrone, page 2
Since the Lagrangian does not depend on x directly, one has
p
y ′ (x)2 1 + y ′ (x)2
c = y ′ Ly′ − L = p p −p
2g(y0 − y) 1 + y ′ (x)2 2g(y0 − y)
by the Beltrami identity. We clear denominators to get
p p
c 2g(y0 − y) 1 + y ′ (x)2 = −1

and then we square both sides to find that

a2 a2 − y0 + y(x)
1 + y ′ (x)2 = =⇒ y ′ (x)2 = .
y0 − y(x) y0 − y(x)
This is a separable equation that can also be written as
Z √
y0 − y dy
Z
p = dx.
a2 − y0 + y
13 / 43
Example 3. Brachistochrone, page 3

Letting y0 − y = a2 sin2 θ gives a2 − y0 + y = a2 cos2 θ, hence

Z √
y0 − y dy
Z
− dx = − p
a2 − y0 + y
a sin θ · 2a2 sin θ cos θ dθ
Z Z
2
= =a (1 − cos(2θ)) dθ.
a cos θ
Once we now simplify the integrals, we may finally conclude that

a2
 
2 1
x0 − x = a θ − sin(2θ) = (φ − sin φ) ,
2 2
a2 a2
y0 − y = a2 sin2 θ = (1 − cos(2θ)) = (1 − cos φ) ,
2 2
where φ = 2θ. These are the parametric equations of a cycloid.

14 / 43
Case 1. Fixed boundary conditions
Suppose that we wish to find the extremals of
Z b
J(y) = L(x, y(x), y ′ (x)) dx
a

subject to given boundary conditions y(a) = y0 and y(b) = y1 .

This is the standard variational problem which implies that
Z b 
d h ib
Ly − Ly′ φ dx + Ly′ φ = 0
a dx a

for every test function φ. Since the boundary terms vanish, one can
find the possible extremals by solving the Euler-Lagrange equation
d
Ly′ = Ly .
dx
Thus, we get a second-order ODE subject to two boundary conditions.
15 / 43
Case 1. Example

As a simple example, consider the functional

Z 1
J(y) = y ′ (x)2 dx
0

subject to the boundary conditions y(0) = 1 and y(1) = 3.

In this case, the Euler-Lagrange equation gives
d
Ly′ = Ly =⇒ 2y ′′ (x) = 0 =⇒ y(x) = ax + b.
dx
To ensure that the boundary conditions hold, we must have

1 = y(0) = b, 3 = y(1) = a + b.

Thus, the only possible extremal is given by y(x) = 2x + 1.

16 / 43
Case 2. Variable boundary conditions

Suppose that we wish to find the extremals of

Z b
J(y) = L(x, y(x), y ′ (x)) dx
a

when no boundary conditions are specified for y(a) and y(b).

Using the same approach as before, one finds that
Z b 
d h ib
Ly − Ly′ φ dx + Ly′ φ = 0
a dx a

for every function φ. This obviously includes all test functions, so the
Euler-Lagrange equation remains valid, but we must also have

Ly′ = 0 when x = a, b.

These conditions are known as the natural boundary conditions.

17 / 43
Case 2. Example

As a typical example, consider the functional

Z 1 
J(y) = y ′ (x)2 + y(x)y ′ (x) − 4y(x) dx.
0

Using the Euler-Lagrange equation, one finds that

d
Ly′ = Ly =⇒ 2y ′′ + y ′ = y ′ − 4
dx
=⇒ y(x) = −x2 + ax + b.

In view of the natural boundary conditions, we must also have

0 = Ly′ = 2y ′ + y = 2(a − 2x) − x2 + ax + b

when x = 0, 1. This gives 2a + b = 0 = 3a + b − 5, so we easily find

that a = 5 and b = −10. In other words, y(x) = −x2 + 5x − 10.
18 / 43
Case 3. Several unknown functions

Suppose that we wish to find the extremals of

Z b
J(y, z) = L(x, y(x), y ′ (x), z(x), z ′ (x)) dx
a

subject to given boundary conditions, say

y(a) = y0 , y(b) = y1 , z(a) = z0 , z(b) = z1 .

Viewing J(y, z) as a function of one variable, we must then have

d d
Ly′ = Ly , Lz ′ = Lz .
dx dx
In particular, one can find the extremals by solving a system of two
second-order equations subject to four boundary conditions.

19 / 43
Case 3. Example

As a typical example, consider the functional

Z 1 
J(y, z) = y ′ (x)z ′ (x) + y(x)2 dx
0

subject to the conditions y(0) = z(0) = 0 and y(1) = z(1) = 1.

The corresponding Euler-Lagrange equations are then
d
Ly′ = Ly =⇒ z ′′ = 2y,
dx
d
Lz ′ = Lz =⇒ y ′′ = 0.
dx
Solving the latter gives y = ax + b, hence also y = x. Solving the
former, we now get z ′′ = 2x, so z = 31 x3 + cx + d = 31 (x3 + 2x).

20 / 43
Case 4. Isoperimetric constraints

Suppose that we wish to find the extremals of

Z b
J(y) = L(x, y(x), y ′ (x)) dx
a

subject to the boundary and integral constraints

Z b
y(a) = y0 , y(b) = y1 , M (x, y(x), y ′ (x)) dx = c.
a

Let us denote by I(y) the integral that appears in the last equation.
Then the extremals of J(y) are either critical points of I(y) or else
critical points of J(y) − λI(y) for some Lagrange multiplier λ ∈ R.
In particular, one has to solve the Euler-Lagrange equation for two
different Lagrangians, namely M and also L − λM .

21 / 43
Case 4. Example, page 1

We determine the possible extremals of

Z π
J(y) = y ′ (x)2 dx
0

subject to the boundary and integral constraints

Z π
y(0) = y(π) = 0, y(x) sin x dx = 1.
0

Let us denote by I(y) the integral that appears in the last equation.
Its critical points must satisfy the Euler-Lagrange equation
d
L = y(x) sin x =⇒ Ly′ = Ly =⇒ 0 = sin x
dx
and this means that I(y) has no critical points at all.

22 / 43
Case 4. Example, page 2
To find the critical points of J(y) − λI(y), we note that

d
L = y ′ (x)2 − λy(x) sin x =⇒ Ly′ = Ly
dx
=⇒ 2y ′′ = −λ sin x.

Integrating this equation twice, we conclude that

λ λ
y ′ (x) = cos x + a =⇒ y(x) = sin x + ax + b.
2 2
Since y(0) = y(π) = 0, it is easy to check that a = b = 0. Thus, it
remains to determine λ. Using the integral constraint, we get
Z π
λ π 2
Z
λπ
1= y(x) sin x dx = sin x dx = .
0 2 0 4

This gives λ = π4 , so the only possible extremal is y(x) = 2

π sin x.
23 / 43
Lagrange multipliers

Theorem 2. Lagrange multipliers

Suppose that y(x) is an extremal of
Z b
J(y) = L(x, y(x), y ′ (x)) dx
a

subject to the integral constraint I(y) = c, where

Z b
I(y) = M (x, y(x), y ′ (x)) dx.
a

Then the extremal y(x) must be either a critical point of I(y) or else
a critical point of J(y) − λI(y) for some Lagrange multiplier λ ∈ R.

This theorem is closely related to the corresponding theorem for the

extrema of a function f (x, y) subject to a constraint g(x, y) = c.
24 / 43
Lagrange multipliers: Proof, page 1

Let φ, ψ be some given test functions and define f, g : R2 → R by

f (ε, δ) = J(y + εφ + δψ), g(ε, δ) = I(y + εφ + δψ).

We note that g(0, 0) = I(y) = c by assumption, while

g(ε, 0) − g(0, 0) I(y + εφ) − I(y)

gε (0, 0) = lim = lim = I ′ (y)φ.
ε→0 ε ε→0 ε
Suppose y(x) is not a critical point of I(y). We can then find a test
function φ such that I ′ (y)φ ̸= 0. According to the implicit function
theorem, we can thus find a function ε = ε(δ) with ε(0) = 0 and

c = g(ε, δ) = I(y + εφ + δψ)

in a neighbourhood of δ = 0, namely for small enough δ.

25 / 43
Lagrange multipliers: Proof, page 2

Since y(x) is an extremal of J(y) subject to the given constraint,

f (ε, δ) = J(y + εφ + δψ)

attains an extremum at (0, 0) subject to the constraint g(ε, δ) = c.

Using a standard calculus result, we conclude that either ∇g(0, 0) = 0
or else ∇f (0, 0) = λ∇g(0, 0) for some λ ∈ R. One may easily exclude
the former case since gε (0, 0) = I ′ (y)φ ̸= 0 by above. We thus have

∇f (0, 0) = λ∇g(0, 0) =⇒ fδ (0, 0) = λgδ (0, 0)

=⇒ J ′ (y)ψ = λI ′ (y)ψ

for all test functions ψ, so y(x) is a critical point of J(y) − λI(y).

26 / 43
Second variation
The first variation of J(y) is defined as the limit
J(y + εφ) − J(y)
J ′ (y)φ = lim
ε→0 ε
and one can use a Taylor series expansion to derive the formula
Z b
J ′ (y)φ = Ly φ + Ly′ φ′ dx.


a

The second variation of J(y) is defined as the limit

J(y + εφ) − J(y) − εJ ′ (y)φ
J ′′ (y)φ = lim 1 2
2ε
ε→0

and one can use a Taylor series expansion to derive the formula
Z b 
′′
J (y)φ = Lyy φ2 + 2Lyy′ φφ′ + Ly′ y′ (φ′ )2 dx.
a

27 / 43
Second variation: Sketch of proof

According to the definition of J(y), one has

Z b 
J(y + εφ) − J(y) = L(x, y + εφ, y ′ + εφ′ ) − L(x, y, y ′ ) dx.
a

Denote the integrand by F (ε). Then a Taylor series expansion gives

 
F (ε) = ε Ly φ + Ly′ φ′
ε2  
+ Lyy φ2 + 2Lyy′ φφ′ + Ly′ y′ (φ′ )2 + . . . ,
2
where the dots indicate terms which contain higher powers of ε.
Since the linear terms correspond to the first variation J ′ (y)φ, it
easily follows that the quadratic terms correspond to J ′′ (y)φ.

28 / 43
Necessary condition

Theorem 3. Necessary condition

If the functional J(y) attains a local minimum at the function y(x),
then we must actually have J ′′ (y)φ ≥ 0 for all functions φ.

This condition resembles the second derivative test from advanced

calculus. It is closely related to the expansion

ε2 ′′
J(y + εφ) = J(y) + εJ ′ (y)φ + J (y)φ + . . .
2
and the fact that J ′ (y)φ = 0 for all critical points of J(y).
It may happen that J ′′ (y)φ > 0 for all functions φ, even though J(y)
has no local minimum. In particular, the given condition is necessary
but not sufficient for the existence of a local minimum.

29 / 43
Necessary condition: Sketch of proof

Using the definition of J ′′ (y)φ, one can write

ε2 ′′ ε2
J(y + εφ) = J(y) + εJ ′ (y)φ + J (y)φ + R(y, ε, φ)
2 2
for some remainder term R such that R(y, ε, φ) → 0 as ε → 0.
Since y(x) is a point of a local minimum, this implies that

ε2 ′′ ε2
0 ≤ J(y + εφ) − J(y) = J (y)φ + R(y, ε, φ)
2 2
for all small enough ε. Letting ε → 0, we conclude that
 
J ′′ (y)φ = lim J ′′ (y)φ + R(y, ε, φ) ≥ 0.
ε→0

30 / 43
Legendre condition

Theorem 4. Legendre condition

If the functional J(y) attains a local minimum at the function y(x),
then one has Ly′ y′ (x, y, y ′ ) ≥ 0 throughout the interval [a, b].

As a simple example, consider the functional

Z 1 p
J(y) = x 1 + y ′ (x)2 dx.
−1

In this case, the Lagrangian L is easily seen to satisfy

xy ′ (x) x
Ly′ = p =⇒ Ly ′ y ′ = .
1 + y ′ (x)2 (1 + y ′ (x)2 )3/2

Since Ly′ y′ changes sign on the given interval, one finds that J(y)
has neither a local minimum nor a local maximum.
31 / 43
Legendre condition: Sketch of proof, page 1

If it happens that Ly′ y′ < 0 at some point, then Ly′ y′ < 0 on some
interval [x0 − ε, x0 + ε] by continuity. Consider the test function
(   )
sin3 π(x−xε
0)
if |x − x 0 | ≤ ε
φ(x) = .
0 otherwise

This function is bounded for all ε > 0, but its derivative becomes
unbounded as ε → 0. One thus expects the second variation
Z b 
′′
J (y)φ = Lyy φ2 + 2Lyy′ φφ′ + Ly′ y′ (φ′ )2 dx
a

to become negative as ε → 0. This contradicts our previous theorem

which asserts that J ′′ (y)φ ≥ 0 at a point of a local minimum.
It remains to show that J ′′ (y)φ < 0 for all small enough ε > 0.

32 / 43
Legendre condition: Sketch of proof, page 2

Since the Lagrangian is twice continuously differentiable, one has

Z x0 +ε  
J ′′ (y)φ = Lyy φ2 + 2Lyy′ φφ′ + Ly′ y′ (φ′ )2 dx
x0 −ε
Zx0 +ε  
≤ C1 φ2 + C2 |φφ′ | − C3 (φ′ )2 dx
x0 −ε

for some constants C1 , C2 , C3 > 0. Letting u = πε (x − x0 ), we get

π
3πC2 9π 2 C3
Z  
′′ ε 4 2
J (y)φ ≤ C1 + − sin u cos u du
π −π ε ε2
C4
= 2εC1 + 6πC2 −
ε
for some constants C1 , C2 , C4 > 0 and the result now follows.

33 / 43
Poincaré inequality

Suppose that φ is a test function on the interval [a, b]. Then

Z x
|φ(x)| = |φ(x) − φ(a)| ≤ |φ′ (y)| dy,
a

so one may use the Cauchy-Schwarz inequality to find that

Z x Z b Z b
′
2
φ(x) ≤ dy 2
φ (y) dy = (x − a) φ′ (x)2 dx.
a a a

Integrating over [a, b], we thus obtain the Poincaré inequality

b b
(b − a)2
Z Z
φ(x)2 dx ≤ φ′ (x)2 dx.
a 2 a

34 / 43
Sufficient condition

Theorem 5. Sufficient condition

Suppose that y(x) is a critical point of
Z b
J(y) = L(x, y(x), y ′ (x)) dx
a

subject to the boundary conditions y(a) = y0 and y(b) = y1 . Suppose

also that there exists some constant δ > 0 such that
Z b
′′
J (y)φ ≥ δ φ′ (x)2 dx
a

for all test functions φ. Then J(y) attains a local minimum at y(x).

35 / 43
Sufficient condition: Sketch of proof, page 1
We use Taylor’s theorem with remainder to write

ε2 b 
Z 
J(y + εφ) = J(y) + Lyy φ2 + 2Lyy′ φφ′ + Ly′ y′ (φ′ )2 dx
2 a
with the second derivatives of L evaluated at a point of the form

(x, y + tεφ, y ′ + tεφ′ ), 0 ≤ t ≤ 1.

Since L is twice continuously differentiable, this implies that

ε2 ′′
J(y + εφ) = J(y) + J (y)φ
2
ε2 b 
Z 
+ R1 φ2 + 2R2 φφ′ + R3 (φ′ )2 dx
2 a
for some functions R1 , R2 , R3 which approach zero as ε → 0. We
now estimate the integral that appears on the right hand side.
36 / 43
Sufficient condition: Sketch of proof, page 2

Let us denote the last integral by I. We then have

Z b  Z b 
|I| ≤ 2
|R1 | + |R2 | φ(x) dx + |R2 | + |R3 | φ′ (x)2 dx
a a

and we can use the Poincaré inequality to conclude that

Z b
|I| ≤ R(ε, x) · φ′ (x)2 dx
a

for some positive function R which approaches zero as ε → 0.

In view of our assumption on J ′′ (y)φ, this implies that
b
ε2
Z
J(y + εφ) − J(y) ≥ (δ − R(ε, x)) · φ′ (x)2 dx ≥ 0
2 a

for all small enough ε, so J(y) attains a local minimum at y(x).

37 / 43
Sufficient condition: Example

Consider the shortest path example that corresponds to the case

Z bp
J(y) = 1 + y ′ (x)2 dx.
a

In this case, the critical points are lines, namely functions y(x) whose
derivative is constant, say y ′ (x) = c. One can easily check that
1 1
Ly′ y′ = = =δ
(1 + y ′ (x)2 )3/2 (1 + c2 )3/2

for some constant δ > 0, while Lyy = Lyy′ = 0. This implies that
Z b
′′
J (y)φ = δφ′ (x)2 dx,
a

so the sufficient condition is satisfied and J(y) has a local minimum.

38 / 43
Invariance

If there is a transformation (x, y) → (x∗ , y∗ ) such that

Z b Z b∗
′
L(x, y, y ) dx = L(x∗ , y∗ , y∗′ ) dx∗ for all a < b,
a a∗

we say that J(y) is invariant under the given transformation.

A very common example is time invariance

x∗ = x + ε, y∗ = y.

This case arises whenever L is independent of x, for instance.

Another common example is translation invariance

x∗ = x, y∗ = y + ε.

This case arises whenever L is independent of y, for instance.

39 / 43
Noether’s theorem

Theorem 6. Noether’s theorem

Suppose J(y) is invariant under a family of transformations

(x, y) −→ (x∗ , y∗ ) = (f (x, y, ε), g(x, y, ε))

such that x∗ = x and y∗ = y when ε = 0. Then the quantity

Q = α L − y ′ Ly′ + βLy′

is independent of x whenever y(x) is a critical point of J(y) and

∂x∗ ∂y∗
α= , β= .
∂ε ε=0 ∂ε ε=0

The Beltrami identity is a very special case of this theorem.

40 / 43
Noether’s theorem: Sketch of proof, page 1

We simplify the invariance condition using a Taylor series expansion

and keeping the linear terms only. This gives the identity

L(x∗ , y∗ , y∗′ ) = L(x, y, y ′ ) + (x∗ − x)Lx (x, y, y ′ )

+ (y∗ − y)Ly (x, y, y ′ ) + (y∗′ − y ′ )Ly′ (x, y, y ′ ).

Let us express this identity in the more compact form

L∗ = L + ∆x · Lx + ∆y · Ly + ∆y ′ · Ly′ .

Keeping linear terms as before, we get ∆x = x∗ − x = αε and

∆y = y∗ (x) − y(x) = y∗ (x) − y∗ (x∗ ) + y∗ (x∗ ) − y(x)

= −y ′ (x)∆x + βε = (β − αy ′ )ε.

41 / 43
Noether’s theorem: Sketch of proof, page 2

We now integrate the identity above. Since x∗ = x + αε, we have

Z b Z b∗
L dx = L∗ dx∗
a a∗
Z b∗
L + ∆x · Lx + ∆y · Ly + ∆y ′ · Ly′ dx∗


=
a∗
Z b
L + αε · Lx + ∆y · Ly + ∆y ′ · Ly′ + α′ ε · L dx.


=
a

Rearranging terms and integrating by parts, we conclude that

Z b ib Z b  
h d
0= (αε · L)′ dx + ∆y · Ly′ + ∆y Ly − Ly′ dx.
a a a dx

Here, the rightmost integral is zero by the Euler-Lagrange equation.

42 / 43
Noether’s theorem: Sketch of proof, page 3

In view of our computation above, we must thus have

h ib
0 = αε · L + ∆y · Ly′
a
h ib
= αε · L + (β − αy ′ )ε · Ly′ .
a

Since the endpoints a, b are arbitrary, this actually means that

Q = αL + (β − αy ′ )Ly′

is independent of x. Rearranging terms, we conclude that

Q = α(L − y ′ Ly′ ) + βLy′

is independent of x. This finally completes the proof.

43 / 43