Hard Problems from Advanced Partial Dierential
Equations (18.306)
Kenny Kamrin
June 27, 2004
1. We are given the PDE 2 = xx + yy = 0. We must nd solutions of the form
= x f (), where x/y. We also impose the condition that x y = 0 along
the line x = y. What are the possible values for and what are the corresponding
functions f () that solve this PDE?
Solution: Plugging into the PDE, we quickly get that f must fulll the following
ODE:
2 ( 2 + 1)f + 2( 2 + )f + ( 1)f = 0
(1)
We rst note that this isnt a standard Sturm-Liuville problem since our undetermined
constant appears in the f term. We can convert a second order ODE to a SturmLiouville problem only when the undetermined constant, i.e. the eigenvalue, occurs
only in the coecient of the f term. So we must use our ingenuity to tackle this
problem from a dierent perspective.
We notice that is the cotangent of the angle made by the point (x, y) with the
horizontal. This prompts us to try a change of variables so as to take advantage of the
angular correspondence to :
= cot
Similarly, we also get
d
d
d
=
= sin2
d
d(cot )
d
d
d2
d2
2
4
=
sin
sin
2
+
sin
d 2
d
d2
Lets go ahead and dene f () = g(cot ). We will instead solve for g and from there
we can get f . Applying our new variable and our new function, we have
0 = 2 ( 2 + 1)f + 2( 2 + )f + ( 1)f
= cot2 csc2 (sin2 sin 2 g + sin4 g ) + 2 cot (cot2 + )( sin2 ) g + ( 1) g
1
= cos2 g 2 sin cos g + ( 1) g = 0.
(2)
Observing the coecients of Equation 2, we are reminded of the identity
2 sin cos =
d
(cos2 )
d
(3)
which helps us notice the following:
2
(cos2 ) d
2 (cos g)
= cos2 [cos g 2 cos1 sin g (( 1) cos2 sin2 + cos ) g]
= cos2 g 2 cos sin g + [( 1) sin2 cos2 ] g
= cos2 g 2 cos sin g + ( 1) g 2 cos2 g.
This looks an awful lot like the left-hand side of Equation 2, so lets use this fact
to rewrite the ODE:
2
2
2
2
(cos2 ) d
2 (cos g) + cos g = cos g 2 sin cos g + ( 1) g = 0.
This yields
d2
(cos g) = 2 (cos g).
(4)
d2
Now, we rejoice, because letting = cos g, produces a second-order linear ODE
with constant coecients:
= 2 = () = A cos() + B sin()
= g() = (cos )[A cos() + B sin()]
is the general solution for g.
Now, let us consider the given conditions. 0 as y for constant x means
that (x, y ) = x f ()|0 = x g(arccot())|0 = 0. Thus we can deduce that
g(arccot(0)) = g(/2) = 0 whenever x =
0. Looking back to our general solution, we
see that when = /2, the prefactor (cos ) becomes 0 which will go to innity
and violate our condition unless < 0. Thus we require < 0. But under this re
quirement, g(/2) automatically equals 0, meaning that < 0 is the only constraint
2
needed in the x =
0 case. Now, we must check if we need to impose anymore re
strictions on so as to ensure that our condition holds when x = 0. In this case, we
need the leading term x to cancel in order to prevent
from going to along the
2
2
y-axis. Notice
that cos = cos (arccot(x/y)) = ( x + y /x) . Thus, we re-write
(x, y) = x2 + y 2 [A cos( arccot(x/y)) + B sin( arccot(x/y)], and see that when
x = 0, is a constant times y . goes to zero when y as long as is negative,
thus we need not add anymore restrictions to . So, we have that the only constraint
is < 0.
For the other condition, we start by writing it using = x g().
x = x1 g + x x = x1 (g sin cos g )
y = x1 g y = x1 g cos2
x y = x1 (g (sin cos + cos2 )g ).
Our condition says this needs to be zero whenever x = y, or, equivalently, =
/4, 5/4. Thus in the = /4 case, we have
g(/4) g (/4) = 0 = g(/4) = g (/4).
Plugging in our g:
g (/4) = ( cos1 sin (A cos +B sin )+ cos (A sin +B cos ))|=/4
= ( cos
)[A cos + B sin A sin + B cos ].
4
4
4
4
4
The condition g(/4) = g (/4) implies:
A cos(/4) + B sin(/4) = A cos(/4) + B sin(/4) A sin(/4) + B cos(/4)
= B cos(/4) = A sin(/4).
This yields B = A tan(/4) when is not even, and A = 0 with B arbitrary when
is even. In the case of = 5/4, we easily reach the analagous result
B cos(5/4) = A sin(5/4)
3
into which we plug in our recently determined expressions for A and B, yielding
tan(/4) = tan(5/4).
This combined with the -periodicity of the tangent function means that /4 =
5/4 + n for some integer n. Therefore, we deduce the additional constraint that
must be an integer. So, altogether, we have:
For Z and < 0,
A x cos [cos() + tan(/4) sin()]
B x cos [sin()]
if is not even,
if is even.
where = arccot() as previously dened, and the constants A and B are arbitrary.
2. The nonlinear KdV equation ut 6uux + uxxx = 0 can be shown to have a similar
ity solution of the form u = (3t)2/3 g() for = x(3t)1/3 . Applying this solution
reduces the PDE to the ODE g () + (6g() )g () 2g() = 0. Show that this
ODE reduces to the second-order ODE V V 2V 3 = 0 upon a proper choice of
the constant in the variable change g() = (dV ()/d) V ()2 (and given that V
decays exponentially for large ). How does the solution relate to Airy functions for
large ?
Solution: We rst plug the given variable change directly into the ODE for g, yielding
the following ODE for V :
V 2V V + 62 V V 6V V 6V 2 V V
12V (V )2 + 12V 3 V + 2V V 2V + 2V 2 = 0.
We try = 1. A keen observation allows us to notice that with this selection for ,
the ODE can be written as
d2
d
(V V 2V 3 ) 2V (V V 2V 3 ) = 0.
2
d
d
Letting F () =
d
(V
d
V 2V 3 ), we can rewrite the ODE as
dF
= 2V F
d
which quickly reduces to
d
(log F ) = 2V
d
4
leaving us with the general solution
F () = Ae2
V ()d
for some constant A. We are given
that V is exponentially decaying for large , so in
this regime, the antiderivative V d is also an exponentially decaying function. And
thus as gets large, the right side approaches Ae20 = A. So as ,
F () =
d
(V V 2V 3 ) A
d
V V 2V 3 A + B
(5)
for some constant B. As , all three terms on the left side of (5) approach 0
since V is exponentially decaying. Therefore, A and B must both be 0, and our ODE
has succesfully reduced to the second-order ODE
V V 2V 3 = 0
(6)
This marks the completion of the rst part of this problem.
V decaying exponentially means that for large , the V 3 term in Equation 6 is expo
nentially smaller than the other two terms. Likewise we will neglect the V 3 term. We
are left to analyze the ODE
V V = 0
(7)
for 1. We note that if we could somehow convert every (d/d) into and vice
versa, then the ODE would be rst-order and easily solvable. Recalling that Fourier
Transforms have this feature, we express V in terms of its transform:
V () = (1/2)
V (k) eik dk.
Now, we can write
V () = (1/2)
k 2 V (k) eik dk
and
k=
ik
eik
ik
(k) e dk.
V
V () =
V (k) e dk = V (k)
i k=
i
k+
As long as the integration contour is chosen to ensure V (k)eik
= 0, Equation 7
k
can be written in terms of V as:
(k)
V
(1/2)
eik dk = 0
k 2 V (k) +
i
V (k) eik dk =
By uniqueness of the Fourier Transform, this means
V (k)
k 2 V (k) = 0.
i
(8)
This rst order ODE has the solution
3
V (k) = Ceik /3 .
(9)
Now that we have a candidate for V (k), we can only accept it once we verify that there is
k+
=
an appropriate contour whose endpoints are at such that f (k) = V (k)eik
k
0. The function f (k) = Cei(k+k /3) is analytic k C, so we can deform the integration
contour however we wish within C as long as the endpoints remain at . Let k = rei .
The exponent in f (k) is dominated by its cubic term for |k | large, thus for the sake of
detecting decay behavior, we may neglect the k term. The real part of this reduced
exponent is Re(ik 3 /3) = Re(ir3 e3i /3) = (r3 /3) sin(3). Decay behavior requires this
to be negative in the region through which the ends of the contour pass. Consequently,
the ends must approach through regionsfullling 2n < 3 < + 2n. This
0 < < /3
good region can be more simply written as
2/3 < <
4/3 < < 5/3.
So, our goal is met for any contour that approaches + through the 0 < < /3
region, through the 2/3 < < region, and does not attain a large magnitude
i
in a region outside the three listed above. For example, the contour k(t) = t + 1+t
2
would suce. Therefore, we can nally say that the V we received is indeed legitimate.
Transforming out, our solution for V is
3
V () = (1/2)
Cei(k+k /3) dk
= (1/)
C cos(k + k 3 /3) dk
(10)
(11)
= C Ai()
(12)
This solution, though only one of two independent solutions, is the one we desire since
it decays exponentially as +.
