Vous êtes sur la page 1sur 8

We have that

(V (1) ) (V ) (1) + (2),


where (1) is the sum of the |Ej | for (1) , and so on. The first inequality is a simple consequence of the fact that V V (1) , while the second
inequality (which holds even if V (2) 0) is an easy exercise using the minmax principle. One simply uses the eigenfunctions for () as variational
functions for (1) () and for (2) () . We know from Theorem 12.4 that
R
(2)
(2) L1,n n/2 Rn (V )1+n/2 ,and hence (2) L1,n ()1+n/2 . We also
have that (1) = (1 )((1 )1 V (1) ). Now assume that we can prove
the theorem for the potentials V (1) and (1 )1 V (1) . We would then have

(1 )n/2 (V (1) )class + L1,n

lim sup(1+n/2) (V )

lim inf (1+n/2) (V (1) ) (V (1) )class .

Finally, assuming that for every > 0 we can find a decomposition of V


into V (1) + V (2) as above, inequality (7) would then imply the
theorem,namely equation (1).
Step2. The first application of this argument is to cut off V (but not V+ )
at some large value u and some large radius in such a way that the
deleted part of V has small L(1+n/2) (n ) norm. In other words, it suffices
to prove our theorem when V is bounded and has compact support an
assumption that we shall make from now on.

Step3.To resolve problem a) above we first write V = V + V (2) . Note that


V is also bounded below and has compact support. We can ensure that
(2)
kV k1+n/2 < for any > 0 by choosing R = R() small enough (why?).
Unfortunately, V (2) is not negative, but the right side of (7) remains true
and we can obtain the one-sided bound
(1-)(n/2 (V ) + L1,n lim sup(1+n/2) (V ))
A bound in the other direction is obtained as follows.
that V is a
R Note
convex combination of translated copies of V , since Rn G2 = 1. In other
words, if we replace the integral that defines V by a discrete Riemann sum
of cross-section , we would have that = y G2 (y)y , where y is
326
326 More About Eigenvalues the original energy function shifted by yn ,
i.e., with V (x) replaced by V (x y) . By iterating the right side of (7), and
noting that all the y have the same negative eigenvalues, we conclude that
(V ) (V ).
Despite the sketchy presentation here, this convexity argument, which is
quite general, deserves to be noted. A much more direct proof along the
lines of the proof of (6) is discussed in the Exercises. Problem a) will be
solved using these results, but in a slightly different manner
than(7).Combining (9) and (4) we have

lim inf (1+n/2) (V ) (V )class Cn2 R()2 lim sup(1+n/2) N (V ).

Likewise, combining(8)and(5)wehave

lim sup (1+n/2) (V ) L1,n +(1)n/2 [(V )class +Cn2 R()2 lim sup(1+n/2) N (V )]

Equations (10) and (11) will prove the theorem if we can show that

N (V ) 0 as A . This is problem b), and we turn to that


next. Step4. As stated in the Exercises, if we find a potential U such that
U V , then N (V ) will not exceed N (U ). The U we shall choose is
U (x) = v for x and U (x) = 0 otherwise. Here, is a cube of some
length l that supports V . and v is a lower bound forV . The Exercises
also show that the number of negative eigenvalues for U in H l (n ) is, in
turn, not greater than for H 1 (). The latter are the Neumann eigenvalues.
All these facts come from the min-max principle.
(1+n/2)

What we have to compute now is the number of Neumann eigenvalues of


that lie below v. Another exercise shows that the large N asymptotics
(which is the same as the large asymptotics) is the same as for the
Dirichlet problem. According to 12.3(6), with EN = v, we have that there
is a constant n so that the number of eigenvalues satisfies
N (V ) < n ln (v)n/2 .

Recall that l and v are independent of . We conclude that the error term
in (10, 11) goes to zero as 1 .
324
324 More About Eigenvalues PROOF. We use the same coherent states as
in the proof of Theorem 12.11 with G in HRBR) for some radius R. For the
moment, let us replace V by V := V * G2. In this case we have that for any
in Hlat fRV(x)1012(x)dx = f V(y)ITP12(k,y)dk dy. (2) n n Consequently,
8(0 = fan fan 112(k7 121r/C12 sul/(y) dkdy (3) We can now proceed as in
the proof of Theorem 12.11, steps 1-3. To derive an upper bound to E.Ei
we use the min-max principle and choose M(k, y) to be the characteristic
function of the set (k, y) : p2+AV(y) 0. In this way we deduce (recalling that IIVIGfl = Cn2R-2) E Ei(07) = E(29),, Euivrlasscin2R
2N (A9), whereN (pV )isthenumberof negativeeigenvaluesof V.Similarly, asin12.11, weobtainth
E(n)< EGLI/rimsCn217 2N (p.M.(4)(5)N otethe Cin(4)andthe +
Cin(5).Equations(4)and(5)presenttwoproblems :
a)Howdoweestimatethedif f erencebetweenE(p, V )andE(p, V )?b)Howcanweestimatethenumbe
1/(1) + V (2), whereV (2) < 0andsatisf ies1117(2)1114 n/2 < e <
1.W ewritetheenergyandasE = E(1) + E(2), wheregmOP ) =
f ci oivorAv(1)100)(o = f eivorpv(2)10.
323
3

Sections 12.11-12.12 323 greater than fan fan M(k, y)(VFk,y, VFk,y)dk dy.
This follows by consider-ing the significance of the inequality K - K1, 0 in
Fourier space and we leave it as an easy exercise. It is now easy to compute
fR n fn M(k, y)(17 Fky , V Fky) dk dy = itn fRn 127k12M(k,y)dk dy + (N
+ 6)1117 GII2 . (10) This formula looks just like (3) except for the change
of sign in the last term. (10) gives an upper bound and (3) a lower bound.
We can, of course, take the limit e -) 0. As we did for the lower bound, we
utilize the bathtub principle and choose M(k, y) = X Bn(k)X12** (y), with
lisn-lovn. the radius k = nNlinagn The result has the same form, except for
the sign of the error term, and agrees with (1). A second illustration of
coherent states concerns the eigenvalues of p2 + V(x). To obtain a large
N limit we have to consider a sequence of potentials with many
eigenvalues. We give the theorem for the nonrelativis-tic case and leave the
corresponding relativistic case to the reader. This time we will not give an
estimate of the error term because to obtain one would require us to impose
some kind of regularity condition on the potential V; the following contains
no
assumption other than VE Li+n/2(Rn).W enotethesimplescaling : eti(1+
n/2)E0inclassisindependentof p.(11)12.12T HEOREM (LargeN asymptoticsof SchrOclingere
n/2(Rn)LetE(p, V ) := Ei >
0lEi(p, V )I, whereEj(pV )arethenegativeeigenvaluesof A +
pV (x)(countedwiththeirmultiplicity).T hen, limp 0.4 70)E(pv) =
p 0A n/2)E(pwassit)002ISP 11, 27,n .f (v)lin/2(x)dxk)n(n +
2)anasgivenin12.6(8).(1)
322
322 More About Eigenvalues Step I. Let M(k, y) be a function on phase
space with the properties that 0 M(k, y) 1 and fre fRnM(k,y)dk dy = N
+ e (6) for some e 0. Construct the integral kernel (see 12.8(3,4)) K(x, z)
= A 1(k7 Y)7k,y(x, z)dkdy. (7) Rn From Theorem 12.8 we have (since
M(k, y) 1) that (f,f)Jf(x)K(x,z)f(z)dxdz =: (f,Kf) 01 Rn N+e= I K (x, x)
dx. (8) Rn Next, we construct the eigenvalues Ai A2, ... of K, with
corresponding eigenfunctions fi(x). These eigenvalues and eigenfunctions
are constructed in the usual way by first maximizing (f, K f) under the
condition that Ilf112 = 1. One shows that a maximizer 11 exists and then
looks for a maximum of (1, K f) under the additional condition that (f, fi)
= 0, and so on. All this is particularly easy in this case because K is a nice
kernel (see Exercises). The eigenfunctions form an orthonormal set, as
usual, and from (8) we have that 0A1. For each integer J 0 we can define
4

the kernel .1 KAX, Z) := Ai; (x)-f-,(z) j=1 (9) and it is easy to see from
the definition of the eigenvalues of K that (i) K Kj 0, in the sense that
(1, Kf) (f,Kjf) for all f, and (ii) as J goes to infinity El=i Aj converges to
fir K(x, x) dx = N + E. Hence, for some finite integer L, we have that
Ejli1 Ai >
N.Step2.W ewanttomakesurethatthef unctionsf ihavesupportinthedomainQ.Letusdef ineQDQ
:= EQ : dist(x, SF ) > R, sothat1Q 1 >
IS2I4RA(Q)f orsmallR.T hesupportconditioncanbesatisf iedif, f oreachkinW I, wechoosesuppM
.Step3.W enowusethef unctionsf l, 12, , f Linthegeneralizedmin
maxprinciple12.2(1)toconcludethatEiN. : 01Ei < V f i) =
f RnV V zKL(xlz)lx = zdx.Ontheotherhand, thislastintegralisnot
321
Sections 12.10-12.11 321 PROOF. Let BR be a ball of radius R centered at
the origin. R will be chosen to depend on N, A(fl) and Q, but for the
moment it is fixed. We take G to be a spherically symmetric function in Ha
(BR) with unit norm. There is a universal constant C such that it is
possible to have 11Gfl Cn2R-2 (see Exercises). Let 00, /Pi, ... be the
orthonormal eigenfunctions of -A with corre-sponding eigenvalues E0 .Ei
. By 12.8(3) and 12.8(10) their coherent state transforms satisfy N-1 ,
P(kCY) := ) lo(k, y)12 1 j=0 and p = N. J fan (2) We also note the
important fact that supp /Pi C Q implies that supp p(k, ) C Sr := Q U Ix
E Qc : dist(x, Q) R for every k E Rn (why?). Note, also, that 101 1Q*1
IQ + 2RA(Q) when R is small. Using 12.9(1), summed over 0 j N - 1,
we have that N-1 E Ei = i L * 127k12p(k,y)dk dy - N11VGfl. lle II i=0 (3)
We can use (3) to obtain a lower bound to S(N) = EA Ei by using
conditions (2) and applying the bathtub principle to (3) just as in the
proof of Theorem 12.3. The minimizing p is XB,,(k)X(1* (y), with the
radius is = nNlinair 1P-11)1/n Thus, N-1i n 2/n S(N) := E Ei ?.. (27)2 n
N1+2/n IQ* 1 -2/n CNn2/R2. n+ 2 1Sin1-1 j=0 (4) This lower bound is
obviously not as good as Theorem 12.3, but it does give the correct answer
to leading order for large N. We merely have to choose R = n (2A(Q)11/3 (
n -2/3n N -2/3n PI ISn-1 I ) ( PI) (5) and we will then have an error as
stated in the theorem (but with a negative sign). The new feature is an
upper bound to S(N), and here we use the gen-eralized min-max principle,
Theorem 12.2. Coherent states are admirably suited for constructing the
trial functions mentioned there.
320
320 More About Eigenvalues PROOF. Recall that (2) is figt,[(127k12 +
5

m2)1/2 - m]lil)(k)12 dk and that 11(-a)1/11G112 = (27)-n f lkiia(k)12dk.


We then proceed as in 12.9 by mul-tiplying 12.8(2) by (12rk12 + m2)1/2 m and integrating over k. To prove (1), however, we use the inequality (I kl2
+ m 2) 1 / 2 m m + Okq 12 + mt/2m whichiseasilyverif iedbydef iningA =
(kl, k2, k3, m)andB =
(q1, q2, q3, 0)asvectorsinR4andusingthetriangleinequalityAl1 <
1B1 + IA
BI.(3)isprovedsimilarly.N owwearereadytoapplycoherentstatestotheeigenvalueprobleminadom
1)
dimensionalM inkowskicontentof OQ.(Itmightbeinf inity, of course, butitiswelldef ined.)1A(Q
limsupn[LnxECr : dist(x, Q) < rr10a0 +
f nXEQ : dist(x, Cr) < rd.(4)12.11T HEOREM (LargeN eigenvaluesumsinadomain)LetQCR
319
Sections 12.8-12.10 319 12.9 THEOREM (Representation of the
nonrelativistic kinetic energy) Suppose that G in 12.7(1) is in H1 (1[8n)
and either that G(x) = G(x) for all x or that G(x) is real for all x. Then for
all 4l1 E H1 (Rn) W03 = f 121r kr b(k dk dY 11VG112 110112. Rn IRn (
1 ) PROOF. Multiply both sides of 12.8(2) by 127rk12 and integrate over k
(using Fubinis theorem). Then I fRn 127rk121P(k, y) dk dy = Atn fRn
12,71.42R-to q) idiqx 12 C )1 dkdq. (2) Now write 1k12 = Ikq 121q12 +
2(q, (kq))andthenchangeintegrationvariablesin(2)f romkandqtokandkq.RecallingT heorem7.9,
n2rkilb(k)12dkandBz =
f iltn27/016(k)12dk.Bothof theseintegralsmakesensesince0 0, G, 14andIM OareinL2.However,
G(x)impliesthatG(k) = G(k)whileG(x) = G(x)impliesthat6(k) =
Ineithercase1 6(k)12 = Id(k)12andhenceB =
0.F ortherelativistickineticenergythereisnosimplef ormulaasinT heorem12.9, butthereisanef f
f Rn[(12742 + m2 <
11R.T hai + m2)1/2ml1/201122(2)f I[(127k12nz2N 1/2)(k7Y )12dkdy +
11(6)1/4Gf l110113RnRn(3)
318
318 More About Eigenvalues trivially computable to be f 10(x)12dx =
11011) since f 1G(x - y)12 dy = f1G(y)12dy = 1. Thus, we can conclude
that the function y I+ f H (x, y) dx = are *1G-12)(Y) is an Ll(Rn)
function; hence it is finite for almost every y. Another way to view this
result is that for almost every y the function x i- G(x - y)ib(x) is in L2(Rn).
It is also in Li(Rn) (since G E L2 and 0 E L2). :16(k, y) is then the Fourier
transform of this function, and our (1) is nothing more than Plancherels
theorem (Sect. 5.3). Formula (3) is an immediate consequence of (1)
6

together with Theorem 5.3. This little exercise shows the power of Fubinis
theorem. Similarly, (2) follows from (4), by interchanging k and y (and
noting that le2ni(k,y)1 = 1). We now prove (4). Parsevals identity is (A,
B) = (A, :13) for A, B E L2(11e). Let A(x) = Fky(x). Then 0(k, y) = (A,
0) while the right side of (4) is just (A, Tz-p). ........ 0...0 Not only is
110112 = 110112, as Theorem 12.8 states, but 0 can also be bounded
pointwise in terms of 110112. Using 12.7(1) we have 11Fk,y112 = 1 for all
k/ Y (8) and, since 0(k, y) = (Fk,y, 0), the Schwarz inequality implies
likk,Y)1 110112 for all k, y. (9) A more interesting fact is that if cbo, 01,
... , ON are any orthonormal functions in L2(Rn), then N E Wi(k, y)12 1.
(10) j=o The proof uses (3) and imitates 12.3(5); we leave it to the reader.
In the next theorem we show how to represent the kinetic energy 110/P 112
in terms of coherent states. The formula is similar to, but more
complicated than, the representation in terms of Fourier transforms, Sect.
7.9, namely 110112 = f 127k121-s;(k)12dk, Rn and the reader might wonder
why something requiring one integral deserves to be represented in terms of
a double integral, plus an extra negative term. The advantage is that the
potential energy (in the case of the Schr8dinger eigenvalues) or the domain
(2 can also be conveniently accommodated in this formalism, along with
the Laplacian. We ask for the readers patience at this point.
317
Sections 12.6-12.8 317 12.8 THEOREM (Resolution of the identity) Let 0 E
L2(r) and let :11) and G be the Fourier transforms of 0 and G (which are
also in L2( n)). Then (with G R(x) := G(-x) and with * denoting
convolution) f 11*, Y) 12 dk = (1012 * 1G Ri2)(Y) for a.e. Y1 (1) fRn
1;;(k7Y)124 = (I 12 *1d12)(k) for a.e. k, (2) fRn fRn 13(kY)12 dk 4 =:
liTbfl = (0)/P) = (11-7-1/1)). (3) Finally, for all k and y, 17)(k, y) =
(271-)-n e-21ri(kv) i 11.)(q)e27i(QY)G (q - k) dq. (4) Rn REMARK.
Formally, (3) says that fRnirkY dk dy = I = Identity, fRn (5) where rk,y is
the projection onto Fk,v, i.e., (7k,y0)(x) = (Fk,y, 0). This can also be
formally written as fRnFkY (X)P kY (xi) dk dy = (5(x - x). fRn (6)
Strictly speaking, (6) is meaningless because the left side appears to be a
function (if it is anything at all) while the right side is a distribution that is
not a function. The same problem arises with Fourier transforms where one
is tempted to write f exp[27i(k, x - x)] dk = (5(x - x). Eq (6) has to be
interpreted as in (3), i.e., as a weak integral (just like Parsevals identity),
namely fRn fRn(0)7rk,y1P)dk dY = f 12.(k)12 dk = (7 50)- (7) PROOF.
To prove (1), consider the function of two variables H(x, y) a-10(x)121G(x 7

y)12, which is certainly measurable and nonnegative. By Fu-binis theorem,


ff H(x, y) dx dy = f If H(x, y) dy dx oo if either of these two iterated
integrals is finite. The second of these integrals is

Vous aimerez peut-être aussi