Proposition (19981018).
In the following cases, the conclusion of Problem 1 (the uniqueness
of critical point of w in D) holds without additional conditions
under the assumption of Problem 1.
-
If W is a homogeneous polynomial.
-
If W is a polynomial of total degree 4
or less (not necessarily homogeneous).
Remarks.
-
For both cases in the Proposition,
no term of y-degree 1 is allowed under the assumptions of
the Problem 1.
-
For W of homogeneous degree, the y dependence of W is in a
single term under the assumptions of Problem 1.
Hence this case is actually a special case of Proposition (19981226).
-
W with terms of degree 4 or less are the examples
(the only known class of examples with non-trivial parameters in the
coefficients) of
W with more than one terms of non-trivial dependence on y.
Put F(x,y)=(y/x2) (Wx2/Wy)(x,y).
Note that if F(x,y)> 1
holds on D-{(x,0)|x>0}, then y/x2 is
a Lyapnov function (on the same set)
of a dynamical system (Xn,Yn)
determined by grad W; namely,
if (Xn+1,Yn+1)= (grad W)(Xn,Yn)
and if
(Xn,Yn) is in D-{(x,0)|x>0}, then
Yn+1/Xn+12
<Yn/Xn2 holds.
Therefore the condition F(x,y)> 1
serves as a sufficient condition for the absence of
critical point of w in D - {(x,0)|x>0}.
In this respect, we have the following result.
Proposition (19981226).
Consider a W which satisfies the conditions in Problem 1.
Assume furthermore that W has just one term with y degree greater than 1,
and the remaining terms do not depend on y; namely,
W(x,y)=f(x) + b xk yk'
for a polynomial f of one variable with positive coefficients and
constants b>0, k>=0, k'>=2.
Denote by L (resp., H) the lowest (resp., highest) degree among the terms in f.
Then, if at least one of the following holds then the infimum of F(x,y)
(defined above) in D is attained at the boundary (possibly at infinity) of D.
- k=0 or k'=2,
- L >= 4 beta/(k'+2) or H <= 4 beta/(k'+2)
(where beta=k'+(k/2)),
- otherwise, f'(x0) x01-2 beta
>= b k (k'+2)/(k'-2), where
x=x0 is the unique positive root of
x f''(x)/f'(x) = 4 beta/(k'+2)-1.
Remarks.
-
Note that F(x,x2) > 1 under
the assumptions of Problem 1.
Hence, to have F(x,y) > 1 in D, one only has to check
boundaries of D other than the parabolla y=x2.
-
The sufficient conditions in the Proposition are irrelevant of
the detailed properties of W in the interior of D.
All the conditions are reduced to the properties of functions of
(at most) one variable.
If F(x,y) > 1 holds at the boundaries of D, then
the remark above the Proposition implies that there are no
critical point of w in the interior of D.
A proof of Proposition (19981226).
Note first that the assumptions for
W(x,y)=f(x) + b xk yk'
(b>0, k>=0, k'>=2)
in Problem 1 implies that L>=3 (i.e., f is a polynomial with terms of
degree no less than 3) and k+k'>=3.
Put R(x,z):=F(x,x2z)1/2
=z1/2
(Wx/Wy1/2)(x,x2z)
=(bk')-1/2(f'(x) x-beta+1z1-k'/2
+bk xbetaz1+k'/2),
where we put beta=k'+(k/2).
If k=0 or k'=2 then R(x,z) is obviously monotone in z, so that
F(x,y)=R(x,y/x2)2 takes its extremum at the boundary
z=0 or z=1 of D. Henceforth we may assume k>1 and k'>2.
Put g(x) := x f''(x)/f'(x).
If R(x,z) attains its extremum at (x,z), then it must hold that
R,x(x,z)=R,z(x,z)=0,
where R,x denotes the partial derivative of R in x.
By explicit calculation, we have, as extremum condition,
g(x) = 4 beta/(k'+2) - 1,
b k zk' = (k'-2) f'(x) x1-2 beta/(k'+2).
If f is a monomial, f(x)=a xm (a>0, m>2), then
g(x) = 4 beta/(k'+2) - 1 implies L=H=m=4 beta/(k'+2),
in which case the extremum of R
is taken on a curve extending to x to infinity, hence the infimum of F is
attained at the boundary of D.
Henceforth we may assume that f has more
than 2 terms, and put
f(x)=aLxL + ... + aHxH
(aL aH >0, H>L>=3).
By explicit calculation we find g'(x) >0, x>0
(A note for correlation inequality enthusiasts: Use Griffiths' inequality
for ferromagnetic spin system of 1 degree of freedom!).
Therefore, L-1=g(0+)< g(x) < g(infinity)=H-1, x>0.
In particular, if
L >= 4 beta/(k'+2) or H <= 4 beta/(k'+2), then
g(x) = 4 beta/(k'+2) - 1 has no solution for positive x, hence
infimum of F is not attained in the interior of D.
Henceforth we may assume H > 4 beta/(k'+2) > L.
In this case, let x0 be the unique positive root of
x f''(x)/f'(x) = g(x)= 4 beta/(k'+2)-1.
Then the conditions for an extremum of R imply that an extremum
must be of a form (x0,z0), where
b k z0k' =
(k'-2) f'(x0) x01-2 beta/(k'+2).
Hence if the right hand side of this equality is no smaller than b k,
then the extremum of R cannot be attained in the interior of D.
All the cases listed in the Proposition is thus exhausted, and the proof
is complete.
Remarks.
-
One does have to check F(x,y) > 1 at the boundaries,
especially for large x.
For example,
W(x,y)=x3/3 + 4 x2y3 /3
satisfies the assumptions of the Proposition (19981226)
because H < =4 beta/(k'+2), but
the corresponding w=W(x,y)-(x2+y2)/2 has
3 critical points in D:
(1,0), (0.939352,0.283324), (0.734382,0.46355)
(the large blobs in the right figure;
one of them is the minimum of w, and is supposed to be in the dark area).
In fact,
F(x,y) = (1+8y3/(3x))2/(4y) -> 1/(4y),
x->infinity,
so that if y > 1/4, then F(x,y) eventually is less than 1 when x is large,
hence y/x2 is not a Lyapnov function in D.
-
This page (and the Proposition) is an answer to Hiro Ochiai's
suggestion:
Date: Fri, 6 Nov 1998 15:27:56 +0900 (JST)
From: Hiroyuki Ochiai
Actually, your counterexample has only two terms.
Then, | we should | clarify the | meaning |
of this counterexample. |
| ^^^ you | | ``````` |
|
(The counterexample mentioned in Hiro's e-mail is the example
introduced above. I first found the example as an counterexample
against a Conjecture which I stated in Oct. 1998.)