Miscellanea
7.4 Analytic Vectors, Quasi-analytic Vectors, Stieltjes Vectors, and Self-adjointness of Symmetric Operatorsand Self-adjointness of Symmetric Operators
In this section we develop criteria for the self-adjointness of symmetric operators and the strong commutativity of self-adjoint operators by means of analytic vectors, quasi-analytic vectors, and Stieltjes vectors. These are classes of vectorsx which are defined by growth conditions on the sequence(Tnx)n∈N0.
Let us begin with some basics on quasi-analytic functions. Let(mn)n∈N0 be a positive sequence, andJ ⊆Rbe an open interval. We denote byC{mn}the set of functionsf ∈C∞(J)for which there exists a constantKf >0 such that
f(n)(t )≤Kfnmn for allt∈J, n∈N0. (7.20) A linear subspaceC ofC∞(J)is called a quasi-analytic class, provided that the following holds: iff ∈C and there is a pointt0∈J such thatf(n)(t0)=0 for all n∈N0, thenf (t )≡0 onJ.
The quasi-analytic classes among the linear subspacesC{mn}are characterized by the following famous Denjoy–Carleman theorem.
Proposition 7.6 C{mn}is a quasi-analytic class if and only if ∞
n=1
inf
k≥nm1/ kk −1
= ∞. (7.21)
A proof of Proposition7.6can be found in [Hr, Theorem 1.3.8]. In the case of log convex sequences a proof is given in [Ru1, Theorem 19.11].
For the results obtained in this section, it suffices to have the following:
Corollary 7.7 Let(mn)n∈N0 be a positive sequence such that ∞
n=1
m−n1/n= ∞. (7.22)
Suppose thatf ∈C∞(J)and there is a constantKf >0 such that (7.20) is sat-isfied. If there exists at0∈J such thatf(n)(t0)=0 for alln∈N0, thenf (t )≡0 onJ.
Proof Since obviously m1/nn ≥infk≥nm1/ kk , (7.22) implies (7.21), soC{mn}is a quasi-analytic class by Proposition7.6. This gives the assertion.
Example 7.2(mn=n!)n∈N Sincen! ≤nn, the sequence(n!)n∈N0 satisfies (7.21) and (7.22), soC{n!} is a quasi-analytic class. But any function f ∈C{n!}has a holomorphic extension to some strip{z:Rez∈J, |Imz|< δ},δ >0 (see, e.g., [Ru1, Theorem 19.9]). Using the latter fact, the assertion of Corollary7.7is well
known in this special case. ◦
7.4 Analytic Vectors, Quasi-analytic Vectors, Stieltjes Vectors 145 Example 7.3 (log convex sequences) Let(mn)n∈N0be a positive sequence such that m0=1 and
m2n≤mn−1mn+1 forn∈N. (7.23) Condition (7.23) means that(logmn)n∈N0 is a convex sequence. It is well known and easily checked that thenm1/nn ≤m1/ kk forn≤k, som1/nn =infk≥nm1/ kk . There-fore, by Proposition7.6, in this caseC{mn}is a quasi-analytic class if and only if
nm−n1/n= ∞, that is, if (7.22) is satisfied. ◦ Next, we define and investigate various classes of vectors for a linear operatorT on a Hilbert spaceH. A vectorx∈His called aC∞-vector ofT ifx belongs to
D∞(T ):=∞
n=1D Tn
.
The reason for this terminology stems from the following fact which follows easily from Proposition 6.1: IfT is self-adjoint, then a vector x∈Hbelongs to D∞(T )if and only ift→eit Txis aC∞-map ofRinto the Hilbert spaceH.
Definition 7.1 Letx∈D∞(T ). We write
x∈Db(T )andxis called bounded forT if there is a constantBx>0 such that
Tnx≤Bxn forn∈N, (7.24)
x∈Da(T )andxis called analytic forT if there exists a constantCx>0 such that Tnx≤Cxnn! forn∈N0, (7.25) x∈Dqa(T )andxis called quasi-analytic forT if
∞ n=1
Tnx−1/n= ∞, (7.26)
x∈Ds(T )andxis called a Stieltjes vector forT if ∞
n=1
Tnx−1/(2n)= ∞. (7.27)
Obviously, eigenvectors ofT are bounded vectors. The setsDa(T )andDb(T ) are linear subspaces ofD∞(T ), and we have
Db(T )⊆Da(T )⊆Dqa(T )⊆Ds(T ). (7.28) The second inclusion in (7.28) follows from the inequalityTnx−1/n≥(Cxn)−1, sincen! ≤nn. A vectorxis inDqa(T )(resp.Ds(T )) if and only ifλxis forλ=0, but the sum of two vectors fromDqa(T )(resp.Ds(T )) is in general not inDqa(T ) (resp.Ds(T )).
Example 7.4 (Example7.3continued) Suppose that the operatorT is symmetric andx is a unit vector ofD∞(T ). Setmn:=Tnx. Then the assumptions of Exam-ple7.3are fulfilled, sincem0=1 and
m2n=Tnx2=
Tnx, Tnx
=
Tn−1x, Tn+1x
≤mn−1mn+1 forn∈N. Hence, as stated in Example7.3,C{Tnx}is a quasi-analytic class if and only if
nTnx−1/n= ∞, that is,x∈Dqa(T ).
In Example 7.3 it was also noted that T2n+1x1/(2n+1)≥ T2nx1/(2n) for n∈N. This in turn yields that
kTkx−1/ k≤2
nT2nx−1/(2n). Therefore, x∈Dqa(T )if and only if
nT2nx−1/(2n)= ∞. ◦
Example 7.5 (Examples of analytic vectors) LetAbe a self-adjoint operator onH, and letα >0. Clearly,e−α|A|ande−αA2 are bounded operators onH. Suppose that u, v∈H.
Statement x:=e−α|A|uandy:=e−αA2vare analytic vectors forA.
Proof By Stirling’s formula [RA, p. 45] there is a null sequence(εn)such that n! =√
2π nn+1/2e−n(1+εn) forn∈N. (7.29) Hence, there is a constantc >0 such that n2n≤ce2n(n!)2 for n∈N. By some elementary arguments one shows thatλ2ne−2α|λ|≤α−2nn2nforλ∈R. Thus,
Rλ2ne−2α|λ|d
EA(λ)u, u
≤
Rα−2nn2nd
EA(λ)u, u
≤cu2α−2ne2n(n!)2. Hence,u∈D(Ane−α|A|)by (4.26), that is, x∈D(An). Since the first integral is equal toAne−α|A|u2= Anx2by (4.33), the preceding shows thatx∈Da(A).
Sinceλ2ne−2αλ2 ≤λ2ne−2α|λ| for λ∈R,|λ| ≥1, a slight modification of this
reasoning yieldsy∈Da(A). ◦
Proposition 7.8 Suppose that the operatorT is self-adjoint andx∈Da(T ). For anyz∈C,|z|< Cx−1, whereCxis given by (7.25), we havex∈D(ezT)and
ezTx= lim
n→∞
n k=0
zk
k!Tkx. (7.30)
Proof Fixz∈C, |z|< Cx−1. Let E be the spectral measure of T, andEx(·)the measureE(·)x, x. Letn∈N. ThenTn:=T E([−n, n])is a bounded self-adjoint operator onHwith spectrum contained in[−n, n]. Since the series∞
k=0zk k!λk con-verges to the functionezλuniformly on[−n, n], we haveezTnx=∞
k=0zk
k!Tnkx(see Proposition 4.12(v)). Using properties of the functional calculus, we derive
7.4 Analytic Vectors, Quasi-analytic Vectors, Stieltjes Vectors 147 n
−n
ezλ2dEx(λ) 1/2
=ezTE
[−n, n]x=ezTnx= ∞
k=0
zk k!Tnkx
≤∞
k=0
|z|k
k! Tnkx≤∞
k=0
|z|k k! Tkx
≤ ∞ k=0
|zCx|k=
1− |zCx|−1
. Lettingn→ ∞, we get
R|ezλ|2dEx(λ) <∞. Therefore,x∈D(ezT)by (4.26).
Replacingzby|z|and by−|z|in the preceding, we conclude that the functions e2|z|λande−2|z|λareL1(R, Ex). Hence,e2|zλ|is inL1(R, Ex). The sequence
ϕn(λ):=
ezλ− n k=0
zk k!λk
2
, n∈N,
converges to zero onRand has the integrable majorante2|zλ|, because ϕn(λ)=
k≥n+1
zk k!λk
2≤e2|zλ|.
Therefore, it follows from Lebesgue’s dominated convergence theorem (Theo-rem B.1) that
ezTx−
n k=0
zk k!Tkx
2
=
R
ezλ−
n k=0
zk k!λk
2
dEx(λ)=
Rϕn(λ) dEx(λ)→0 asn→ ∞. This proves formula (7.30).
Corollary 7.9 Suppose thatT is a self-adjoint operator onH. For anyx∈H, the following statements are equivalent:
(i) x∈Da(T ).
(ii) There is ac >0 such thatx∈D(ezT)for allz∈C,|Rez|< c.
(iii) There is ac >0 such thatx∈D(ec|T|).
Proof Clearly, by formula (4.26),x∈D∞(T )if (ii) or (iii) holds. The implication (i)→(ii) was proved by Proposition7.8. The equivalence of (ii) and (iii) follows at once from (4.26). Suppose that (iii) is satisfied and setu=ec|T|x. Thenx=e−c|T|u,
sox∈Da(T )by Example7.5.
The following corollary contains some explanation for the terminology “analytic vector”; another reason is given by Example7.8below.
Corollary 7.10 LetT be a self-adjoint operator onH,y∈H, andx∈Da(T ). Then f (z):= eizTx, yis a holomorphic function on the strip{z: |Imz|< Cx−1}, where Cxis given by (7.25).
Proof Lets∈R. By formula (7.30), applied withzreplaced by i(z−s), the series f (z)=
eizTx, y
=
e(iz−is)Tx, e−isTy
=∞
n=0
(i(z−s))n n!
Tnx, e−isTy
converges if|z−s|< Cx−1. Hence,f is holomorphic on the set{|Imz|< Cx−1}.
Before we turn to the main results of this section, we prove three simple lemmas.
Lemma 7.11 Let(rn)n∈Nbe a positive sequence and setmn=αrn+β forn∈N, whereα >0 andβ >0. If
nrn−1/n= ∞, then
nm−n1/n= ∞.
Proof LetMbe the set{n:αrn≤β}. Forn∈M, we havemn≤2βand m−n1/n≥(2β)−1/n≥(1+2β)−1,
so
nm−n1/n= ∞when the setMis infinite. IfMis finite, then there is ak∈N such thatmn≤2αrnfor alln≥k, and so
n≥k
m−n1/n≥
n≥k
(2α)−1/nrn−1/n≥
n≥k
(1+2α)−1rn−1/n= ∞.
Lemma 7.12 IfT if a symmetric operator, thenDs(T )⊆Ds(T+zI )forz∈C.
Proof Letx∈Ds(T ). Since λx∈Ds(T ) forλ∈C, we can assume without loss of generality thatT x=0 andx =1. As noted in Examples7.3and7.4, we then haveTlx1/ l≤ Tnx1/nforl≤n. Using this inequality, we derive forn∈N,
(T+zI )nx=
n l=0
n l
zn−lTlx
≤
n l=0
n l
|z|n−lTlx
≤ n
l=0
n l
|z|n−lTnxl/n=Tnx1/n+ |z|n
=Tnx1+ |z|Tnx−1/nn
≤ Tnx
1+ |z|T x−1n
. Sincex∈Ds(T ), we have
nTnx−1/(2n)= ∞. By the preceding inequality this implies that
n(T+zI )nx−1/(2n)= ∞, that is,x∈Ds(T+zI ).
Lemma 7.13 IfT is a self-adjoint operator onH, thenDb(T )is dense inH. Proof LetEbe the spectral measure ofT. Ifa >0 andx∈E([−a, a])H, then
Tnx2= a
−a
λ2nd
E(λ)x, x
≤a2nx2≤
a(1+ x)2n
forn∈N, sox∈Db(T ). For anyy∈H,y=lima→∞E([−a, a])y. Therefore, the set
a>0E([−a, a])His contained inDb(T )and dense inH.
7.4 Analytic Vectors, Quasi-analytic Vectors, Stieltjes Vectors 149 The following three theorems are the main results in this section. They all contain sufficient conditions for the self-adjointness of symmetric operators in terms of the density of special classes of vectors. In the proof of Theorem 7.15we shall use Theorem 10.17 below. Theorems7.14and7.15are due to A.E. Nussbaum.
Theorem 7.14 IfT is a symmetric operator on a Hilbert spaceHsuch that the linear span ofDqa(T )is dense inH, thenT is essentially self-adjoint.
Proof First, we note thatT is densely defined, because the linear span ofDqa(T )is dense and obviously contained inD(T ).
We prove that N(T∗−iI )= {0}. Assume to the contrary that there exists a nonzero vectory∈N(T∗−iI ). Letx∈Dqa(T ),x=0. By Proposition 3.17 the symmetric operator T has a self-adjoint extension Ain a possibly larger Hilbert spaceG⊇H. Define a functionf onRbyf (t )= eit Ax, y −etx, y.
SinceT ⊆A, we haveD∞(T )⊆D∞(A), so x∈D∞(A). Therefore, by a re-peated application of Proposition 6.1(iii), it follows thatf∈C∞(R)and
f(n)(t )=
eit A(iA)nx, y
−etx, y fort∈R, n∈N. (7.31) SinceT∗y=iy, we haveTnx, y = Tn−1x,iy = · · · = x,iny, and hence
f(n)(0)=
(iA)nx, y
− x, y =in Tnx, y
− x, y =0 forn∈N0. (7.32) Leta >0 andJ =(−a, a). Putα= y,β=eaxyandmn=αTnx+βfor n∈N0. Sincex∈Dqa(T ), we have
nTnx−1/n= ∞, and hence
nm−n1/n=
∞by Lemma7.11. From (7.31) we obtain
f(n)(t )≤Anxy +eaxy =αTnx+β=mn fort∈J, n∈N0, sof satisfies Eq. (7.20) withKf =1. Further,f(n)(0)=0 for alln∈N0by (7.32).
By the preceding we have shown that all assumptions of Corollary7.7are fulfilled.
From Corollary7.7we conclude thatf (t )≡0 on(−a, a)and hence onR, because a >0 was arbitrary. That is, we haveeit Ax, y =etx, yfor allt∈R.
But the functioneit Ax, yis bounded on R, whileet is unbounded on R. So the latter equality is only possible when x, y =0. Thus, we have proved that Dqa(T )⊥y. Since the linear span ofDqa(T )is dense inH, we gety=0, which is the desired contradiction. Therefore,N(T∗−iI )=0.
A similar reasoning shows that N(T∗+iI )=0. Hence, T is essentially
self-adjoint by Proposition 3.8.
Theorem 7.15 Suppose thatT is a semibounded symmetric operator onH. If the linear span ofDs(T )is dense inH, thenT is essentially self-adjoint.
Proof The proof follows a similar pattern as the proof of Theorem 7.14. By Lemma7.12we can assume without loss of generality thatT ≥I. Then 0∈π(T ) by Lemma 3.3. Therefore, by Proposition 3.9, it suffices to prove thatN(T∗)= {0}.
Assume to the contrary thatN(T∗)= {0}and take a unit vector y∈N(T∗). Let
x∈Ds(T ),x=0. By Theorem 10.17 (which we take for granted for this proof!) there exists a self-adjoint operatorA≥I onHwhich extendsT.
By Proposition 6.7,u(t ):=(cosA1/2t )xsolves the Cauchy problemu(0)=x, u(0)=0 for the wave equationu(t )= −Au(t )andu(t )= −(A1/2sinA1/2t )xfor t∈R. Define a functionf onRbyf (t ):= u(t ), y − x, y.
Since x ∈D∞(T )⊆D∞(A), from the functional calculus of the self-adjoint operatorAwe easily derive that
u(2k)(t )=(−A)ku(t ) and u(2k−1)(t )=(−A)(k−1)u(t ) fork∈N. Therefore,u∈C∞(R,H), and hence f ∈C∞(R). Since y =1 andA≥I, it follows from the preceding and the formulas foruanduthat
f(2k−1)(t )=(−A)k−1u(t ), y≤Ak−1A1/2x≤Akx=Tkx, f(2k)(t )=(−A)ku(t ), y≤Akx=Tkx
fork∈Nand|f (t )| ≤2xonR. That is, settingm0=2x,m2k= Tkx, and m2k−1= Tkx, the preceding shows that|f(n)(t )| ≤mnfor alln∈N. Clearly,
n
m−n1/n≥
n
Tnx−1/(2n)= ∞ by (7.27). Sinceu(0)=0,u(0)=xandT∗y=0, we obtain
f(2k−1)(0)=
(−A)ku(0), y
=0, f (0)= u(0), y
− x, y =0, f(2k)(0)=
(−A)ku(0), y
=
(−T )kx, y
=
−(−T )k−1x, T∗y
=0 for k∈N, that is, we have shown that f(n)(0)=0 for alln∈N0. Therefore, all assumptions of Corollary7.7are satisfied, so we conclude thatf (t )≡0 onR.
Sincef (t )=0 and cosA1/2t is a bounded self-adjoint operator fort ∈R, we have
x, y = u(t ), y
=
cosA1/2t x, y
= x,
cosA1/2t y
.
Therefore,x, y−(cosA1/2t )y =0 for all vectorsx∈Ds(T ). By assumption the linear span ofDs(T )is dense inH. Hence, it follows thaty−(cosA1/2t )y=0. If Eis the spectral measure ofA, from the functional calculus we derive
0=y−
cosA1/2t y2=
∞
1
1−cos
λ1/2t2d
E(λ)y, y .
This implies thatλ1/2t∈2π·Zfor anyλin the support of the measureE(·)y, y. Becauset∈Rwas arbitrary, this only possible when this measure is zero. Since y =1, we arrive at a contradiction. This completes the proof.
The following result is the famous Nelson theorem.
Theorem 7.16 LetT be a symmetric operator onH. If the spaceDa(T )of analytic vectors forT is dense inH, thenT is essentially self-adjoint. IfT is closed, thenT is self-adjoint if and only ifDa(T )is dense inH.
7.4 Analytic Vectors, Quasi-analytic Vectors, Stieltjes Vectors 151 Proof SinceDa(T )is a linear space andDb(T )⊆Da(T )⊆Dqa(T )by (7.28), the assertions follow immediately from Theorem7.14and Lemma7.13.
Nelson’s and Nussbaum’s theorems are powerful tools for proving self-adjoint-ness by means of growth estimates ofTnx. We illustrate this by an example.
Example 7.6 (Jacobi operators) Let{en:n∈N0}be an orthonormal basis of a Hilbert spaceH. Given a complex sequence(an)n∈N and a real sequence(bn)n∈N, we define the linear operatorT with domainD(T )=Lin{en:n∈N0}by
T en=an−1en−1+bnen+anen+1, n∈N0, (7.33) wheree−1:=0. It is easily seen thatT is symmetric.
Operators of the form (7.33) are called Jacobi operators. They play a crucial role in the study of the moment problem in Sect. 16.1, see, e.g., (16.7).
Statement Suppose that there exist two positive real constantsαandβ such that
|ak| ≤αn+β and|bk| ≤αn+β for allk≤n,k, n∈N. Then we haveDa(T )= D(T ), and the operatorT is essentially self-adjoint.
Proof The main part of the proof is to show that each vectorek is analytic forT. From formula (7.33) we concludeTnek is a sum of at most 3n summands of the formγmem, wherem≤k+nandγm is a product ofnfactorsaj,aj, orbj with j < k+n. Therefore, by the above assumption,Tnek ≤3n(α(n+k)+β)n.
Setc:=k+βα−1. Clearly,c0:=supn(n+nc)n<∞. As noted in Example7.5, it follows from Stirling’s formula that there is a constantc1>0 such thatnn≤c1enn! forn∈N. From the preceding we obtain
Tnek≤3n
α(n+k)+βn
=(3α)n(n+c)n≤(3α)nc0nn≤c0c1(3αe)nn!
forn∈N0. This implies thatek∈Da(T ). Hence,D(T )=Da(T ), becauseDa(T ) is a linear space. By Nelson’s Theorem7.16,T is essentially self-adjoint. ◦ Example 7.7 (Annihilation and creation operators) Recall from (7.11) that the an-nihilation operatorAand the creation operatorA+act on the standard orthonormal basis{en:n∈N0}ofH=l2(N0)by
Aen=√
n en−1 and A+en=√
n+1en+1, n∈N0, wheree−1:=0. SinceA∗=A+, setting
P0= 1
√2 i
A−A+
and Q0= 1
√2
A+A+ , P0andQ0are symmetric operators onl2(N0), and we have
P0en= 1
√2 i(√
n en−1−√
n+1en+1), Q0en= 1
√2(√
n en−1+√
n+1en+1).
Since the assumptions of the statement in Example7.6are satisfied, the restrictions of P0 and Q0 to D0:=Lin{en:n∈N0} are essentially self-adjoint. It is easily checked that the operatorsP0andQ0satisfy the canonical commutation relation
P0Q0−Q0P0= −iID0. (7.34) The operatorP = −idxd and the multiplication operatorQ=Mx by the variablex onL2(R)satisfy this relation as well. It is well known that the Hermite functions
hn(x):=(−1)n 2nn!√
π−1/2
ex2/2 dn
dxne−x2, n∈N0,
form an orthonormal basis ofL2(R). Hence, there is a unitary operatorUofL2(R) ontol2(N0)defined byU hn=en. Then it can be shown thatU P U∗x=P0x and
U QU∗x=Q0xfor allx∈D0. ◦
Example 7.8 (Analytic vectors for the self-adjoint operatorT = −idxd onL2(R)) LetF denote the Fourier transform, andMxthe multiplication operator by the in-dependent variable x on L2(R). For this example, we will take for granted that T =F−1MxF. (This follows from formula (8.2) in Sect. 8.1).
Statement f ∈L2(R)is an analytic vector forT if and only iff is the restriction to Rof a holomorphic functionF on a strip{z: |Imz|< c}for somec >0 satisfying
|supy|<c
R
F (x+iy)2dx <∞. (7.35) Proof Since T and Mx are unitarily equivalent, f ∈Da(T ) if and only if fˆ= F(f )is inDa(Mx). Obviously,Da(Mx)=Da(|Mx|)=Da(M|x|). Thus, by Corol-lary7.9,f ∈Da(T )if and only iffˆis inD(ecM|x|), or equivalently,ec|x|f (x)ˆ is in L2(R)for somec >0. By a classical theorem of Paley–Wiener [Kz, Theorem 7.1]
the latter is equivalent to the existence of a functionF as stated above.
From this statement it follows that the null vector is the only analytic vector for T contained inC0∞(R). Nevertheless, the restriction ofT toC0∞(R)is essentially self-adjoint, sinceC0∞(R)is a core forT as shown in Example 6.1. However, by Theorem7.16this cannot happen if the symmetric operatorT is closed.
Let(εn)n∈Nbe a positive null sequence. Then the function g(z):=∞
n=1
2−n 1 z−n−εni
is holomorphic on the setC\{n+εni:n∈N}. In particular,gis analytic onR. One easily checks thatg∈D∞(T ). Butgis not an analytic vector forT, because
condi-tion (7.35) is not fulfilled. ◦
Our next main theorem states that if two commuting symmetric operators have sufficiently many joint (!) quasi-analytic vectors, their closures are strongly com-muting self-adjoint operators. For this, we need a preliminary lemma.
7.4 Analytic Vectors, Quasi-analytic Vectors, Stieltjes Vectors 153 IfT is a linear operator with domainDandTD⊆D, we denote by{T}cthe set of all densely defined linear operatorsS satisfyingD⊆D(S)∩D(S∗),SD⊆D, andT Sx=ST xfor allx∈D.
Lemma 7.17 LetT be a symmetric operator with domainDsuch thatTD⊆D.
For any operatorS∈ {T}c, we haveSDqa(T )⊆Dqa(T ). In particular,Dqa(T )is invariant under the operatorT.
Proof Letx∈D. Using the properties ofT andS∈ {T}c, we obtain TnSx2=
T2nSx, Sx
=
ST2nx, Sx
=
T2nx, S∗Sx
≤T2nxS∗Sx. Letx∈Dqa(T ). Without loss of generality we can assume thatx =1. Then we
have
nT2nx−1/(2n) = ∞ by Example7.4. The preceding inequality implies thatTnSx1/n≤ T2nx1/(2n)(1+ S∗Sx). Hence,
n
TnSx−1/n≥
n
T2nx−1/(2n)
1+S∗Sx−1= ∞,
so thatSx∈Dqa(T ).
Theorem 7.18 LetA and B be symmetric operators acting on the same dense domainDof Hsuch thatAD⊆D,BD⊆D, andABx=BAxforx∈D. If the linear spanDQof
Q:= Sx:S∈ {A}c∩ {B}c, x∈Dqa(A)∩Dqa(B)
is dense inH(in particular, ifDa(A)∩Da(B)is dense inH), then the closuresA andBare strongly commuting self-adjoint operators onH.
Proof By the definition of the set{T}cwe haveQ⊆D. From Lemma7.17it follows thatQis contained inDqa(A)andDqa(B). Therefore,AandBare self-adjoint by Theorem7.14.
LetA0andB0denote the restrictions ofAandB toDQandEQ:=(A−iI )DQ, respectively. SinceQ⊆Dqa(A0)has a dense linear span, Theorem7.14applies also toA0, soA0is essentially self-adjoint. Hence,EQ=(A0−iI )DQ=Lin(A−iI )Q is dense inH. But(A−iI )Q⊆Dqa(B) by Lemma7.17, because Q⊆Dqa(B) and(A−iI )∈ {B}c. Thus,(A−iI )Q⊆Dqa(B0), sinceB0=BEQ. Hence, The-orem7.14applies toB0and the set(A−iI )Qas well and implies thatB0is essen-tially self-adjoint. SinceB0⊆Band henceB0⊆B, we obtainB0=B.
Since A and B commute on D, we have B0(A−iI )x =(A−iI )Bx for x∈DQ. Settingy=(A−iI )x, the latter yields(A−iI )−1B0y=B(A−iI )−1yfor y∈EQ=D(B0). From this relation we derive(A−iI )−1B0 y=B( A−iI )−1y fory∈D(B0). SinceB0=Bas noted above, we get(A−iI )−1B⊆B(A−iI )−1. Hence, by Proposition 5.27, the self-adjoint operatorsAandBstrongly commute.
Completion of the proof of Proposition 5.20 Suppose thatvis a generating vector forA. Puty=e−A2v. LetD0denote the linear span of vectorsAny,n∈N0, and let H0be its closure inH. To complete the proof, it suffices to show thatH0=H.
First, we prove thatv∈H0. Setpn(t ):=n
k=0t2k
k! andϕn(t ):=pn(t )e−t2. The sequence(ϕn(t ))n∈Nconverges monotonically to 1 onR. From the functional calcu-lus we therefore conclude thatϕn(A)v=pn(A)y→vinH. Hence,v∈H0, since pn(A)y∈D0.
Obviously,A0:=AD0 is a symmetric operator on the Hilbert spaceH0 and A0D0⊆D0. By Example 7.5, we havey ∈Da(A). SinceADa(A)⊆Da(A), the latter yields D0⊆Da(A), so that D0=Da(A0). Since A0 has a dense set D0
of analytic vectors, A0 is self-adjoint by Theorem 7.16. Hence, Proposition 1.17 applies, andA decomposes asA=A0⊕A1 with respect to the orthogonal sum H=H0⊕H⊥0. Clearly,EA
0(M)=EA(M)H0. Fromv∈H0we getEA(M)v= EA
0(M)v∈H0. Sincevis a generating vector forA, the span of vectorsEA(M)v
is dense inH. Therefore,H=H0.