Main Problem. (Kazdan-Warner) Given (M, g) a compact Riemmanian manifold without boundaries, and a function eR 2 C•(M), is there any comformal metric eg=e2ug, such that the scalar curvature with respect to g is exactly ee R?
In this note, we deal with the case of surfaces (n=dim M=2)
Observation. In the case dimM = 2, we have R = 2K and eR = 2 eK, where K is the sectional curvature (Gaussian curvature). By Gauss-Bonnet
theorem, Z
MKdµ= Z
Me2uKdµe =2pc(M), thus we have the necessary condition:
1. If c(M) < 0, then eK is negative at some point.
2. If c(M) = 0, then either eK⌘0 or eK changes sign.
3. If c(M) > 0, then eK is positive at some point.
However these conditions are in general not sufficient.
1
Conformal changes of curvatures. Under local coordinate, we have eGkij = 1
2 egkl(∂geil
∂xj +∂egjl
∂xi
∂geij
∂xl)
=Gkij+1
2(dik∂u
∂xj +djk∂u
∂xi gijgkl ∂u
∂xl), Reij = ∂eG
tij
∂xt
∂eGtit
∂xj +eGsijeGtst eGsiteGtsj
=Rij (4u)gij, (n=2) Re =geijReij =e 2u(R 24u)
Thus Kazdan-Warner problem for surfaces is equivalent to solving the fol-lowing PDE on M.
4u K+e2uKe=0. (1)
2 Case I: c ( M ) < 0
We may try to study this case by sub- and sup- solution method.
Proposition 2.1. On a compact Riemannian manifold(M, g), consider the semilinear PDE4u+ f(x, u) = 0, where f 2 C•(M⇥R). If there exists f, y 2 C2(M), such that fyand
4f+ f(x, f) 0, 4y+ f(x, y) 0,
(We call f and y a sub-solution and a sup-solution for the PDE respec-tively.) Then there exists u 2 C• s.t. f yand (1) holds.
Proof. Since M is compact, we can choose A, c > 0 such that A < f y<A and ct+ f(x, t)increasing in t2 [ A, A]. We rewrite (1) into
Lu =F(x, u),
where L is an elliptic operator defined by Lu , 4u+cu, and F(x, u) , ct+ f(x, t). By maximum principle (looking at the minimum), we can see that L is a ”positive” operator, in the sense Lu 0 ) u 0, or equiva-lently, Lu1 Lu2)u1 u2.
On the other hand, we have the following result of Schauder estimate:
2
Proposition 2.2. As an operator L : C2,a ! C0,a between Holder spaces (a2 (0, 1)), L has a compact inverse L 1.
Consider two sequences of sub- and sup- solutions{fk},{yk}given by f0 =f, fk+1 =L 1(F(x, fk))
y0 =y, yk+1 =L 1(F(x, yk))
We inductively show that f < f1 < f2 < · · · < y2 < y1 < y. First, notice that Lfk+1 = F(x, fk) Lfk, so fk fk+1 by positiveness of L.
Now if fk yk, then
Lfk+1 =F(x, fk) F(x, yk) = Lyk+1, thus fk+1yk+1again by positiveness of L.
Thus we have pointwise convergence
{fk} ! u,{yk} ! u, f uu y.
Since fk, yk, Lfk, Lykare all bounded, using the following Lp-estimate, we can show that{fk},{yk}are bounded in the Sobolev space W2,pfor p>n:
Proposition 2.3. For u 2 W2,p(M) with p > 1, we have the following inequality1:
kukW2,p C(kukLp +k4ukLp) for some constant C >0.
Hence by Sobolev embedding theorem2, {fk},{yk} are also bounded in C1,a with a = 1 np. As a result, {fk},{yk} converges in C1,a to u, u respectively, and
Lu =F(x, u), Lu=F(x, u),
and then by elliptic regularity theorem, u, u are smooth solutions to (1).
1Theorem 9.13 of [2]
2Theorem 7.26 of [2]
3
Now we may apply the method to the problem (n=dim M=2) Corollary 2.4. In the case of c(M) < 0, if (1) has a sup-solution y in C2, then it has a smooth solution.
Proof. We only have to show that there is also a sub-solution f such that fy. Choose an f 2 C•(M)such that
4f =K K0, where K0 = R
Kdµ/R
dµ is the mean value of K, which is negative by Gauss-Bonnet theorem. Note that R
(K K0)dµ = 0, so such f exists by Hodge decomposition. Now we consider f = f c for some constant c>0. We have
4f K+ ˜Ke2f = K0+ ˜Ke2 f 2c >0
for c sufficient large, and we also take c large enough so that f c y, then hence we find the f we want.
Theorem 2.5. In the case c(M) < 0, if eK 0, then (1) has a smooth solu-tion.
Proof. As previous, we only have to show that there is also a sup-solution yfor (1). Choose an f 2 C•(M)such that
4f =Ke Ke0, where eK0 = R e
Kdµ/R
dµ < 0. Consider y = a f +b for some constant a, b>0. We have
4y K+ ˜Ke2y = (a eK0 K) + (e 2a f+2b a)Ke <0 for a, b sufficient large, so we find the y we want.
For general eK, the problem remains unsolved.
4
3 Case II: c ( M ) = 0
For the case c(M) = 0, we have a complete criterion for eK that (1) has a smooth solution:
Theorem 3.1. For the case c(M) < 0, (1) has smooth solution if and only if one of the following condition holds:
(1) eK ⌘0, or
(2) eK changes sign andR e
Ke2 fdµ<0,
where fR 2 C•(M) is a solution of4f = K. (Note that such f exists since Kdµ=2pc(M) =0.)
Proof. Necessity. Suppose u is a solution of (1), then consider v = u f , then
4v =4u K = e2v+2 fKe (2) Times e2von both side and integrate, we obtain
Z Kee 2 fdµ=
Z e 2v4v=
Z 2e 2v|rv|2dµ0
by Green’s identity. If the equality holds, then rv ⌘ 0, so v must be a constant, which implies eK ⌘ 0 by (2). Thus if eK 6⌘ 0, then we must have R e
Ke2 fdµ<0, and also eK have to change sign by Gauss-Bonnet theorem.
Sufficiency. If eK ⌘ 0, then obviously we can take u = f , so we only have to deal with the second case. Consider the subset of W2,p(M),
S , {u 2W2,p(M),Z udµ=
Z Kee 2u+2 fdµ=0}
It is nonempty since eK changes sign. We want to minimize inS the Dirich-let energy
J(u), 12Z |ru|2dµ
Suppose there is a minimizer u02 S, then by the theory of Lagrange mul-tiplier, u0is an emtreme of
Z 1
2|ru|2+au+b eKe2u+2 f. 5
Compute the variation in u gives
4u0+a+2b eKe2u0+2 f =0.
Integrating over M gives
aA= 2bZ Kee 2u0+2 fdµ=0,
where A is the area of M. Thus e 2u04u0+2b eKe2 f =0. Again integrating over M gives
2bZ Kee 2 fdµ=
Z e 2u04u0dµ.
= 2Z e 2u0|ru0|2 0 which implies b >0 by the condition.
Thus we have v0 = u0+1
2log 2b is a weak solution of (2), i.e., u = u0+ 1
2log 2b+ f is a weak solution of (1). By elliptic regularity, if we can show that eu 2 Lp(M)for all p > 1, then u is a smooth solution. To do this we need the following Lemmas.
Lemma 3.2 (Trudinger). For any compact Riemannian surface(M, g), there exists b, C >0 such that any u 2 W1,2satisfying
Z udµ =0,Z |ru|2dµ1
hasR
ebu2dµ<C.
Proof. We fix a partition of unity{(Ui, fi)}ki=1with each Ui diffeomorphic to a 2 dimensional Euclidean disc D, and let ui =fiu, thus u=Âiui.
We first prove that on each Ui ⇠= D with standard Euclidean metric, there exists c0 >0 such that
kuikp c0pp
kruik2, 8p 2.
6
For v2 C1c(D), we have the following identity, v(x) = 1
2p Z
D4v(y)·log(|x y|)dy
= 1 2p
Z
Drv(y)· x y
|x y|2dy By Holder inequality(1p+1q =1),
|v(x)| 2p1 Z
D
⇣|rv(y)|2|x y| q⌘
1p
|x y| 2q|rv(y)|1 2pdy
2p1
✓Z
D|rv(y)|2|x y| qdy
◆1p ✓Z
D|x y| qdy
◆12 ✓Z
D|rv(y)|2dy
◆12 1p
the middle term is controlled by Z
D|x y| qdy Z
2D|y| qdy=21 qp(p+2). Taking pth power and integrate over x and we have
Z
D|v(x)|pdxc1(p+2)p2+1
✓Z
D|rv(y)|2dy
◆p2
)
✓Z
D|v(x)|pdx
◆1p
c2pp✓Z
D|rv(y)|2dy
◆12
since C1c is dense in W1,2, we have on each(Ui, g), kuikp c0pp
kruik2, for some c0depend on g.
Now we have
kukp
Â
kuikpc0ppÂ
kruik2c3pp(kruk2+kuk2)
by equivalence of sobolev norms. We further use the following estimates:
7
Proposition 3.3 (Poincare-Wirtinger). For any u 2 W1,2(M), there exists C >0 such that
ku u0k2 Ckruk2, where u0=R
udµ/A is the average of u.
Since we have u0=R
udµ =0, we obtain kukp c4pp
kruk2
Now consider the Taylor expansion of ebu2, with kruk 1, we have Z ebu2dµ=
Z
Â
k=1
1
k!(b|u|2)k+A
Â
k=1
1
k!(2kbc24)k+AC for b sufficiently small (Take p =2k.)
Lemma 3.4. There existsC, h >0 such that for any u2W1,2, Z eudµC exp(hkruk22+ 1
A
Z udµ)
Proof. We may assumeru6⌘0, and let u0=u A1 R
udµ. We write u0 b✓ u0
kruk2
◆2
+ 1
4bkruk22,
where b is the constant appeared in the last lemma. Note that u0
kruk2 satis-fies the condition of the last lemma, thus taking exponent and integrating
over M gives Z
eu0dµC exp( 1
4bkruk22), and hence the original statement. (Take h = 1
4b.)
Lemma 3.5. Foru 2W1,2(M), we have eu 2 Lpfor all p >1.
Proof. Replace u by pu in the above lemma.
8
We still need to show the existence of the minimizer of J. Suppose{uk} is a minimizing sequence of J inS, i.e.,
J(ui) !c0 = inf
u2SJ(u) Then by Poincare-Wirtinger inequality we have
kuik2 Ckruik2 =2CJ(ui)12
So {ui} is bounded in W1,2, and there is a subsequence {ui} weakly con-verging to some u0 2 W1,2. We have J(u0) c0by the weak semi-lower-continuity of J. On the other hand, we can show thatR e
Ke2udµ is a contin-uous functional of u in W1,2-weak topology, and hence
Z Kee 2u0+2 fdµ= lim
i!•
Z Kee 2ui+2 fdµ=0
and alsoR
u0dµ = 0, hence u0 2 S and J(u0) c0 by the definition of c0. As a consequence, u0 is a minimizer of J inS, and we complete the proof of the theorem.
Proof of the continuity ofR e
Ke2udµ. Suppose {ui}weakly converges to u, then by Rellich–Kondrachov theorem, they are strongly converge to u in Lpfor p>1, thus
Z Ke(eui eu)dµ= Z 1
0
Z
(ui u)Kee u+t(ui u)dµdt!0 as i !0, by Lemma 3.4. and Holder’s inequality.
9
4 Case III: c ( M ) > 0
Only partial results are known for c(M) > 0. In this case, M must be diffeomorphic to S2 or RP2. We first consider the case M = S2 with the standard 2-sphere metric g (K ⌘1).
Proposition 4.1. On(S2, g), If f is a first eigenfunction of4, i.e., 4f+2f=0,
then any solution u of (1) must satisfies Z
(rKe· rf)e2udµ =0
Note that f can be seen as a linear function on R3restricted to S2. Proof. Multiply (1) byru· rfand integrate over S2, we have
Z
(ru· rf)4udµ Z
(ru· rf)dµ+ Z
(ru· rf)Kee 2udµ=0 we deal with these three terms respectively. Note that f,ij = fgij, so we have
Z
(ru· rf)4udµ= Z
r(ru· rf)· rudµ
= 1 2
Z
r(|ru|2)· rfdµ+ Z
|ru|2fdµ
= 1 2
Z
|ru|2(4f+2f)dµ=0, Z
(ru· rf)dµ= Z
f4udµ = Z
f(Kee 2u 1)dµ= Z
f eKe2udµ, Z
(ru· rf)Kee 2udµ= 1 2
Z
(re2u· rf)Kdµe
= 1 2
Z
(r(Kee 2u)· rf)dµ 1 2
Z
(rKe· rf)e2udµ
= 1 2
Z Kee 2u4fdµ 12 Z
(rKe· rf)e2udµ summing up the three terms then the proposition follows.
10
With this, consider eK = 1+efwith e > 0 sufficinet small so that eK > 0, then the above proposition says
e Z
|rf|2e2udµ =0,
which impliesrf = 0, or f is a constant, contradiction. This shows that even eK>0 is not sufficient for (1) to have a solution.
Nevertheless, we have the following result:
Theorem 4.2. On(S2, g), if eK(x) = Ke( x)for all x2 S2, and eK is positive somewhere, then there is a solution u 2 C•(S2) such that u(x) = u( x) for all x 2S2.
Thus for standard RP2, the condition that eK is positive somewhere is necessary and sufficient.
We need the fact that in this case (S2 and symmetry) we can actually take the constant h in Lemma 3.4 to be a number close to 1/32p, that is, we can take b = 8p # with small # in Lemma 3.2. We first prove for b=4p #/2 if we drop the condition u( x) = u(x).
The first step is to symmetrize u into a radially symmetric function u# using the following Polya-Szego inequality [4].
Lemma 4.3 (Weil-Polya-Szego). Let (M, g)be a Riemannian surface such that the sectional curvature of M is bounded from above by k, with B ⇢ M an open subset diffeomorphic to R2 and with smooth boundary. For u2 W1,2(B)and u#is a radially symmetric function on a geodesic ball B# on S = k 1/2S2 (2-sphere of radius k 1/2) monotone in latitude, such that
|B#| = |B|and|(u#) 1((t, •))| = |u 1(t, •)|for any t 2 R, then we have u#2 W1,2(B#)and
Z
B#|ru#(x)|2dµS Z
B|ru(x)|2dµM. (and hence we can replace u with u#if M=S2and k =1.)
Now we rewrite things into spherical coordinate. Let q, f be the longi-tude and altilongi-tude of the 2-sphere, then the metric can be written as
ds2 =df2+cos q dq2, 11
and the area element is
dµ=cos q dq df.
The conditions on u is then Z
|ru|2dµ= Z
(u2q+cos 2q u2f)cos q dq df
=2pZ p
pu2q cos q dq 1, Z u dµ=2pZ u cos q dq=0.
Parametrize q by t such that e t/2 =tan
✓q 2
p 4
◆
, and let
w(t) = (4p)1/2u(q), r(t) = 1
4sech2t 2 then the conditions become
Z
|ru|2dµ = Z •
•w0(t)2dt 1, Z u dµ =
Z •
•w(t)r(t)dt =0, and we want a bound for
Z e(4p #/2)u2dµ=4pZ •
•e(1 #/8p))2w(t)2r(t)dt.
Note that r(t)has the properties
r(t) < C0e |t|, Z •
•r(t)dt=1 for some constant C0. By Cauchy’s inequality we have
(w(r) w(s))2 =
✓Z r
s w0(t)dt
◆2
✓Z r
s w0(t)2dt
◆ ✓Z r
s 1dt
◆
|r s|. Write(1 #/8p)2 =1 t, then
Z •
•e(1 #/8p)2w(t)2r(t)dtC0e(1 t)C1 Z •
•e t|t|dt=2C0e(1 t)C1t 1 12
if we have the following estimate
w(t)2<|t| +C1
for some constant C1. To show this inequality, we write r(s)(w(t) w(s))r(s)|(t s)|1/2 integrate over s and we have
w(t) Z •
•r(s)|(t s)|1/2 ds (|t| +C1)1/2 for some C1 (R•
•r(s)|(t s)|1/2ds)2 |t|for all t.
Remark. In fact the boundedness is also true for b = 4p, but that requires a longer argument controlling the case when
d,1 maxr>s (w(r) w(s))2/|r s| is small.
We return to the case b = 8p e with u( x) = u(x). In Lemma 4.3, we take B to be a hemisphere on S2, and k=p
2, then the lemma implies Z
S|ru#(x)|2dµS 12 Z
S2|ru(x)|2dµS2, where S = p1
2S2, and thus Z
S2|ru#(p
2x)|2dµS2 12 Z
S2|ru(x)|2dµS2
if we scale u#to the standard 2-sphere. Now we can run the same estimate for the functionp
2u#(p
2x), and we haveR
e(8p #)u2dµ bounded.
Now consider the setS , {u 2 W1,2(S2),R
udµ = 0,R e
Ke2udµ > 0}. This is nonempty since eK is positive somewhere. We want to minimize
13
J(u) = 1
2kruk22 2p logZ Kee 2udµ Note that c0 ,infu2S J(u) ✓1
2
2p 8p #
◆
kruk22+c1 • by Lemma 3.4 and h = 1
32p 4#. A similar argument of case II shows that J has a minimizer u0 2 S. The theory of Lagrange multiplier and calculus of variation then gives
4u0+R4p eeKe2u0
Ke2u0dµ l =0
for some multiplier l. Integrate over S2gives 4p 4pl, hence l =1 and we have u =u0+12log(4p1 R e
Ke2u0dµ)is a solution of (1).