Alphabeta Math
Pipeline-generated
How statement and proof provenance work

The first chip identifies the source of the statement or construction; the second identifies the source of its local proof or verification.

  • Literature-sourced: the exact statement appears in a cited source; only wording and notation differ.
  • AI-adapted: a semantically identical restatement of literature-sourced material, modulo indexing, notation, and boundary cases adopted by the library.
  • AI-generated: a genuinely novel statement formulated by AI, with no source for the claim itself.

These labels describe origin, not correctness: citations and verification chips remain separate evidence.

Unbounded Self Adjoint Operators and Stones Theorem

1 · Prerequisites

2 · Summary

This page extends the bounded PVM calculus of the preceding pair to unbounded self-adjoint operators and ends with Stone's theorem and the min-max principle. It fixes at the outset that a linear operator carries its domain as part of its data, that ST means containment of graphs, and that closedness is closedness of the graph in HH; the graph norm makes closedness a completeness statement. The adjoint is defined only for densely defined operators, by representability of xTx,y through Hilbert space Riesz representation, and is proved closed with ran(Tz)=ker(Tz); a graph rotation identifies closability with density of D(T) and yields T=T. Symmetric, self-adjoint and essentially self-adjoint operators are then defined by TT, T=T and self-adjointness of T, and the minimal derivative id/dx with vanishing endpoint conditions is proved closed and symmetric but not self-adjoint, with the periodic operator in between.

The resolvent RT(z)=(zT)1 is used in the library's sign convention. For self-adjoint T the identity (Tz)x2=(Ta)x2+b2x2 off the real axis gives CRρ(T) with RT(z)1/Imz, and the resulting range criterion characterises self-adjointness by ran(T±i)=H. The Cayley transform CT=(Ti)(T+i)1 is unitary with ker(ICT)={0} and is proved to be a bijection from self-adjoint operators onto such unitaries, the inverse recovering T from ran(IU). The unbounded PVM integral f(E) has domain {x:f2dEx<} and is closed, with adjoint f(E); the spectral theorem is proved by transporting the bounded normal spectral theorem along λ(λi)(λ+i)1, and completeness of the calculus is recorded as the product, sum and spectral-mapping rules on the correct domains.

Stone's theorem is proved in both directions: a self-adjoint T generates U(t)=eitT by the Borel calculus, with derivative domain exactly D(T), while the generator of a strongly continuous unitary group is reconstructed from the Laplace resolvents 0eλtU(±t)xdt, proved to be bounded inverses of λG; the generator is closed, symmetric and skew-adjoint, and a group is determined by its generator, so TeitT is a bijection. The deficiency subspaces K±=ker(Ti) and the von Neumann parameterization of self-adjoint extensions by unitaries K+K are proved, with the explicit domain D(T){u+Vu} and action, and equal deficiency indices are equivalent to existence of a self-adjoint extension. The page closes with the canonical decomposition H=HppHacHsc by spectral type, with relative boundedness, the second resolvent identity, the Kato-Rellich theorem, the discrete and essential spectrum with the Weyl criterion and Weyl's invariance theorem, norm and strong resolvent convergence with the continuous calculus and unitary-group corollary, and the form domain and min-max principle below the essential spectrum.

3 · Logical flowchart

4 · Definitions, theorems and proofs

DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Unbounded linear operators: domain, graph and extension

Definition

Throughout this page H is a complex Hilbert space (Hilbert space) with inner product linear in the first variable and conjugate-linear in the second (Real and complex inner-product spaces and their induced length).

A (possibly unbounded) linear operator on H is a linear map T:D(T)H (Linear map between vector spaces over the same field) whose domain D(T)H is a linear subspace (Linear subspace of a vector space). The domain is part of the data: T=S means D(T)=D(S) and Tx=Sx for all xD(T). One writes TS, and calls S an extension of T, when D(T)D(S) and Sx=Tx for every xD(T).

The graph of T is Γ(T):={(x,Tx):xD(T)}HH. The direct sum HH is read as the complex vector space of pairs with coordinatewise operations and the inner product (x,y),(u,v)=x,u+y,v, whose induced norm is (x,y)=(x2+y2)1/2; it is complete because H is (Complete metric space: every Cauchy sequence converges in the space). The graph norm on D(T) is xT:=(x2+Tx2)1/2, so that x(x,Tx) is an isometric isomorphism of (D(T),T) onto Γ(T) with the norm restricted from HH.

T is closed when Γ(T) is a closed subset of HH. A linear operator is determined by its graph, and the following elementary translations are used silently below: Γ(T) is a linear subspace of HH; Γ(T)({0}H)={(0,0)}; the image of Γ(T) under the first coordinate projection is D(T); and TS if and only if Γ(T)Γ(S). In particular a closed operator is exactly one whose graph is a closed subspace of HH.

Notation. No boundedness of T is assumed, and T is frequently called unbounded to stress this; a bounded everywhere defined operator on H is the special case D(T)=H in which Γ(T) is a closed subspace by continuity. The letter I denotes the identity operator on H with domain H.

Completeness of HH. If (xn,yn) is a Cauchy sequence in HH, then xnxm(xnxm,ynym) and ynym(xnxm,ynym) show that (xn) and (yn) are Cauchy in H; let x,y be their limits. Then (xn,yn)(x,y)2=xnx2+yny20, so HH is complete. Consequently a sequence in HH converges exactly when its two coordinate sequences converge in H.

DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Densely defined, closed and closable operators, and cores

Definition

Assume Countable Choice (The Axiom of Countable Choice (ACω)). Let T be a linear operator on H with domain D(T) and graph Γ(T) (Unbounded linear operators: domain, graph and extension).

T is densely defined when D(T) is a dense subset of H (Dense, nowhere dense and codense subsets of a topological space, and the criterion by basic open sets); closed when Γ(T) is a closed subset of HH; and closable when T has a closed extension. Recall that the graph norm is xT=(x2+Tx2)1/2.

The graph-norm dictionary. The following three statements are part of the definition's content and are proved in the remarks below rather than assumed:

  1. T is a norm on D(T), and x(x,Tx) is an isometric isomorphism of (D(T),T) onto Γ(T);
  2. T is closed if and only if (D(T),T) is a complete metric space (Complete metric space: every Cauchy sequence converges in the space);
  3. if T is closed, then D(T) is a Hilbert space for the inner product x,yT=x,y+Tx,Ty whose induced norm is T.

A linear subspace D0D(T) is a core for a closed operator T when D0 is dense in (D(T),T), equivalently when Γ(TD0)=Γ(T), where the closure is taken in HH and TD0 denotes the restriction of T to D0.

Remarks

Claim 1. T= holds on kerT, and xxT always. The map J:x(x,Tx) is linear, and xT=JxHH. Hence homogeneity and the triangle inequality for T are the corresponding norm properties in HH pulled back along J. It is definite because xT=0 forces x=0 and x=0. The map J is therefore linear and isometric, and its image is Γ(T) with (x,Tx)2=x2+Tx2=xT2; a linear isometry is injective, so it is a bijection onto Γ(T).

Claim 2. The space HH is complete with (x,y) (Complete metric space: every Cauchy sequence converges in the space, Hilbert space), while Γ(T) carries the subspace metric. If Γ(T) is closed, every Cauchy sequence in it converges in HH and its limit remains in Γ(T), so the graph is complete. Conversely suppose Γ(T) is complete and qΓ(T). Countable Choice selects qnΓ(T) with qnq<1/(n+1) for every nN. Then (qn) is Cauchy, so it converges to a point of Γ(T); uniqueness of metric limits makes that point q. Thus Γ(T) is closed. Since the map of claim 1 is an isometry onto Γ(T), the space (D(T),T) is complete exactly when Γ(T) is closed, that is, exactly when T is closed.

Claim 3. If T is closed then Γ(T) is a closed subspace of the Hilbert space HH, hence a Hilbert space in its own right, and the inner product (x,Tx),(y,Ty)=x,y+Tx,Ty transfers to D(T) through the isometry of claim 1, making D(T) a Hilbert space with inner product x,yT and induced norm T.

Core. The isometry J:x(x,Tx) of claim 1 is a homeomorphism from (D(T),T) onto Γ(T) and maps D0 onto Γ(TD0). Hence it carries the closure of D0 onto the closure of Γ(TD0) inside Γ(T). Therefore D0 is graph-norm dense exactly when Γ(TD0)=Γ(T); this topological argument does not replace density by sequential density.

TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Closure of a closable operator

Statement

Assume Countable Choice (The Axiom of Countable Choice (ACω)). Let T be a linear operator on H with domain D(T) and graph Γ(T) (Unbounded linear operators: domain, graph and extension). Then the following are equivalent:

  1. T is closable (Densely defined, closed and closable operators, and cores);
  2. whenever xnD(T), xn0 and Txny, one has y=0.

If either condition holds, then the closure of Γ(T) in HH is the graph of a linear operator T, the closure of T; it is the least closed extension of T, and T is closed if and only if T=T. Necessity of the hypothesis is never claimed: both conditions hold automatically for a closed operator.

Facts & Assumptions

[A1]

Γ(T) is a linear subspace of HH; TS exactly when Γ(T)Γ(S); T is closed exactly when Γ(T) is closed; and Γ(T)({0}H)={(0,0)} (Unbounded linear operators: domain, graph and extension).

[A2]

T is closable when it has a closed extension; the closure A of a subset A of a metric space is closed, is contained in every closed set containing A, and every point of A is the limit of a sequence in A (Interior, closure, boundary, limit point, isolated point and dense subset of a metric space, Densely defined, closed and closable operators, and cores).

[A3]

A sequence in HH converges exactly when its two coordinate sequences converge in H (Unbounded linear operators: domain, graph and extension).

Proof

technique · direct

Given: A linear operator T on H and the two conditions (1) and (2).

1.1

If T has a closed extension S, then Γ(T)Γ(S) with Γ(S) closed, and for xnD(T) with xn0, Txny we get (xn,Txn)(0,y) by [A3] with (xn,Txn)Γ(T)Γ(S); closedness of Γ(S) gives (0,y)Γ(S), and by the last clause of [A1] applied to S this forces y=0. Thus (1) implies (2).

A1A2A3given
1.2

Now assume (2), and let G:=Γ(T)HH. Then G is closed and, being the closure of the linear subspace Γ(T), is itself a linear subspace. If (0,y)G then by [A2] there are (xn,Txn)Γ(T) with (xn,Txn)(0,y), hence xn0 and Txny by [A3], so y=0 by (2). Therefore G({0}H)={(0,0)}.

A1A2A3given
2.1

By step 1.2, G determines at most one second coordinate per first coordinate: if (x,y),(x,y)G then (0,yy)=(x,y)(x,y)G since G is a subspace, so y=y. Hence D:={xH:there is y with (x,y)G} is a linear subspace of H, the formula Tx:=y for (x,y)G defines a linear operator T:DH with Γ(T)=G, and T is closed with TT because Γ(T)G.

A1step 1.2
3.1

By step 2.1 the operator T is a closed extension of T, so T is closable, and (2) implies (1). With 1.1 this proves the equivalence of (1) and (2), and it shows that whenever either holds the closure Γ(T) is the graph of the closed extension T.

A2step 1.1step 2.1
3.2

If R is any closed extension of T, then Γ(R) is closed and contains Γ(T), so G=Γ(T)Γ(R) by [A2], that is, TR. Thus T is the least closed extension of T.

A1A2step 2.1
4.1

If T is closed then Γ(T) is already closed, so G=Γ(T) and T=T; conversely if T=T then T is closed because T is closed by step 2.1. Hence T is closed if and only if T=T, and the closure of the graph is the graph of the least closed extension.

A1step 2.1step 3.2
DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Adjoint of a densely defined operator

Definition

Assume the Axiom of Countable Choice (The Axiom of Countable Choice (ACω)). Let T be a densely defined linear operator on H, with domain D(T) dense in H (Densely defined, closed and closable operators, and cores, Dense, nowhere dense and codense subsets of a topological space, and the criterion by basic open sets). Since D(T) is dense, and since H is complete (Hilbert space), every bounded linear functional on the normed space (D(T),) (A bounded linear operator between normed spaces) extends uniquely to a bounded linear functional φ~ on H of the same norm, by setting φ~(z)=limnφ(zn) for any sequence znD(T) with znz in H; the limit exists because φ is bounded on a dense subspace and H is complete, and it does not depend on the sequence because a bounded functional is uniformly continuous.

A vector yH belongs to the adjoint domain D(T) when the linear functional φy:D(T)C, φy(x)=Tx,y, is bounded on (D(T),). In that case Hilbert space Riesz representation (Riesz representation for Hilbert spaces) applied to the extension φy~ produces a unique vector TyH with Ty is the unique wH such that Tx,y=x,wfor all xD(T), equivalently φy~(z)=z,Ty for all zH. The map T:D(T)H, yTy, is the adjoint of T.

Well-definedness. The functional φy is linear in x for each y, so its domain of boundedness D(T) is a linear subspace: if φy,φy are bounded so is φay+by for scalars a,b, because the first-variable-linear convention gives φay+by=aφy+bφy. On D(T) the map T is linear: if y,y are represented by w,w, then ay+by is represented by aw+bw, since x,aw+bw=ax,w+bx,w; the representing vector is unique. By Riesz representation for the Hilbert space H a vector wH is determined by the values z,w with z ranging over H, and those values are determined by the functional φy~. Equivalently, if w,w both represent φy, then x,ww=0 for every xD(T); density of D(T) and continuity of the inner product extend this equality to every xH, and taking x=ww gives w=w. Finally only ambient-norm boundedness of φy on D(T) is required: no extension, no closure and no closedness of T is presupposed.

LemmaStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

The adjoint is well defined, closed, and reverses inclusions

Statement

Assume Countable Choice. For a densely defined linear operator T on H the adjoint T is well defined and linear with linear domain D(T); T is closed; for every zC ran(Tz)=ker(Tz), and if ST and both are densely defined, then TS.

Facts & Assumptions

[A1]

yD(T) holds exactly when xTx,y is bounded on D(T), and then Ty is the unique w with Tx,y=x,w for all xD(T); T is linear and D(T) is a linear subspace (Adjoint of a densely defined operator, The Axiom of Countable Choice (ACω)).

[A3]

T is closed exactly when Γ(T) is closed, and Γ(T)={(y,w):Tx,y=x,w for all xD(T)} (Unbounded linear operators: domain, graph and extension, Adjoint of a densely defined operator).

Proof

technique · direct

Given: A densely defined linear operator T on H.

1.1

By [A1] the adjoint is well defined, D(T) is a linear subspace and T is linear.

A1
1.2

Let ynD(T) with yny and Tynw. For every xD(T) we have Tx,y=limnTx,yn=limnx,Tyn=x,w, both limits being scalar limits; hence xTx,y is bounded, with Tx,y=x,wxw. So yD(T) and Ty=w by [A1], and therefore T is closed.

A1
1.3

Let yH and zC. If yran(Tz), then (Tz)x,y=0, that is Tx,y=zx,y=x,zy, for every xD(T); this exhibits xTx,y as a bounded functional, so yD(T) and Ty=zy by uniqueness in [A1]. Conversely if Ty=zy, then (Tz)x,y=x,zyzx,y=0 for every xD(T), so yran(Tz). Hence ran(Tz)=ker(Tz).

A1
1.4

If ST are densely defined, then Sx,y=Tx,y=x,w for every xD(S), so every pair (y,w)Γ(T) lies in Γ(S) by [A3]; hence TS.

A3
2.1

Claims collected: well-definedness and linearity from step 1.1, closedness from step 1.2, the kernel-range identity from step 1.3, and the inclusion reversal from step 1.4. ∎

TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Closability is equivalent to density of the adjoint domain

Statement

Assume Countable Choice. Let T be a densely defined linear operator on H. Then T is closable if and only if D(T) is dense in H. In that case T=T and T=T; moreover if TT, then T is closable.

Facts & Assumptions

[A1]

For densely defined T the adjoint is defined by Tx,y=x,Ty for xD(T), yD(T), and Γ(T)={(y,w):Tx,y=x,w for all xD(T)}. Putting W(x,y):=(y,x) defines a bijective isometry of HH with W2=I, and W maps orthocomplements to orthocomplements: W(M)=(WM) (Adjoint of a densely defined operator, Hilbert space, Orthogonality and the orthogonal complement).

[A2]

If M is a linear subspace of the Hilbert space HH, then M=M (The double orthogonal complement of a subspace is its closure).

[A3]

For a densely defined T the operator T is closed. If in addition D(T) is dense, then T is defined, is closed, and contains T (The adjoint is well defined, closed, and reverses inclusions, Adjoint of a densely defined operator).

[A4]

T is closable when it has a closed extension; if T is closable then T is the least closed extension of T and TT (Closure of a closable operator, Densely defined, closed and closable operators, and cores).

Proof

technique · direct

Given: A densely defined linear operator T on H.

1.1

By [A1] we have Γ(T)=WΓ(T): indeed (y,w)Γ(T) means Tx,y=x,w for all xD(T), which is exactly (x,Tx),W(y,w)=0 for all xD(T), that is W(y,w)Γ(T).

A1
2.1

Assume in addition that D(T) is dense, so that T is defined. Replacing T by T in step 1.1 gives Γ(T)=WΓ(T), so by step 1.1 and unitarity of W, with W(M)=(WM) and W2=I, one has Γ(T)=W(WΓ(T))=WWΓ(T)=Γ(T)=Γ(T).

step 1.1A1A2
2.2

Conversely assume T is closable, and let T be its closure, a closed densely defined operator with Γ(T)Γ(T) and Γ(T)=Γ(T) by [A4]. Then D(T) is dense: if yD(T) then (y,0)Γ(T), so by step 1.1 applied to T and closedness of T we get (y,0)Γ(T)=WΓ(T)=WΓ(T), that is (y,0)=W(x,Tx)=(Tx,x) for some xD(T), forcing x=0 and y=0.

A1A2A4step 1.1
3.1

Under the hypothesis of step 2.1 the set Γ(T) is the graph of the operator T, which is closed by [A3]; hence T has a closed extension and is closable, and its closure is T by minimality in [A4].

step 2.1A3A4
3.2

With T closable as in step 2.2, apply step 1.1 to T and to T: using Γ(T)=Γ(T) and [A2] one gets Γ(T)=WΓ(T)=WΓ(T)=WΓ(T)=Γ(T), so T=T and D(T)=D(T) is dense by step 2.2.

A1A2step 1.1step 2.2
4.1

The two implications are steps 3.1 (density of D(T) gives closability, with closure T) and 3.2 (closability gives density of D(T), and T=T). If TT, then T is a closed extension of T by [A3], so T is closable.

step 3.1step 3.2
DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Symmetric, self-adjoint and essentially self-adjoint operators

Definition

Assume Countable Choice. Let T be a densely defined linear operator on H.

  • T is symmetric when TT, that is, D(T)D(T) and Tx=Tx for all xD(T); equivalently Tx,y=x,Ty for all x,yD(T).
  • T is self-adjoint when T=T, that is, the domains and the values agree.
  • T is essentially self-adjoint when its closure T is self-adjoint; by Closability is equivalent to density of the adjoint domain this presupposes that T is closable.

Consequences, with proofs. These are part of the content of the definition.

  1. A symmetric T is closable. TT and T is closed (The adjoint is well defined, closed, and reverses inclusions), so T is a closed extension of T. In particular the closure T exists and TT.
  2. The closure of a symmetric operator is symmetric. If TT, the inclusion reversal of The adjoint is well defined, closed, and reverses inclusions applied to TT gives (T)T, that is TT. Applying it once more to the inclusion TT gives (T)(T), that is TT. Since T=T and (T)=T (Closability is equivalent to density of the adjoint domain), this reads T(T): the closure is symmetric. (Each application is legitimate because the adjoint is defined once its operator is densely defined, and D(T) is dense because TT=T, so that D(T) contains the dense domain D(T).)
  3. A self-adjoint operator has no proper symmetric extension. If T=T and TSS, then ST by inclusion reversal, so TSST=T and all inclusions are equalities.
  4. A self-adjoint operator is closed, being equal to the adjoint T of the densely defined T, and an essentially self-adjoint operator has exactly one self-adjoint extension, namely T: a self-adjoint extension S of T is closed, hence contains the least closed extension T, and then SS(T)=T by item 2 and symmetry, so S=T.
CounterexampleConstruction: Literature-sourcedVerification: AI-adaptedaudited 2026-09-22 rests on later materialOpen item page →

A symmetric closed operator that is not self-adjoint

Statement refuted

Assume the Axioms of Countable Choice and Dependent Choice. On the Hilbert space H=L2((0,1);C) (The space Lp(μ) as the quotient by null functions) let T:=id/dx with domain D(T)={fAC[0,1]:fL2(0,1), f(0)=f(1)=0}, where complex-valued absolute continuity and the derivative are read on real and imaginary parts (Absolute continuity on a compact interval). Domain notation means the L2 classes having the indicated absolutely continuous representative; endpoint values refer to that representative. Then:

  1. T is densely defined, closed and symmetric, so T refutes the reading "closed and symmetric implies self-adjoint";
  2. D(T)={gAC[0,1]:gL2(0,1)} and Tg=ig, so T is a proper closed extension of T; T itself is not symmetric, and TT;
  3. the periodic domain D2={gAC[0,1]:gL2(0,1), g(0)=g(1)} carries a closed symmetric extension of T, strictly between T and T.

Facts & Assumptions

[A1]

A densely defined operator is symmetric when TT and self-adjoint when T=T; the adjoint of a densely defined operator is always closed (Symmetric, self-adjoint and essentially self-adjoint operators, The adjoint is well defined, closed, and reverses inclusions).

[A2]

For real-valued absolutely continuous functions the fundamental theorem of calculus and integration by parts hold, and the indefinite integral of an L1 function is absolutely continuous; applied to real and imaginary parts this gives the same calculus for complex-valued absolutely continuous functions (Absolute continuity on a compact interval, Fundamental theorem of calculus for absolutely continuous functions, Integration by parts for absolutely continuous functions, The indefinite integral of an L1 function is absolutely continuous, The indefinite integral of an L1 function is differentiable almost everywhere).

[A3]

Cc(R) is dense in L2(R) under Countable Choice, applied componentwise for complex functions. There exists a smooth χ on R with 0χ1, equal to one on [1,1] and zero off [2,2]. Dominated convergence applies under an integrable majorant (Cc(Rn) is dense in Lp(Rn) for 1p<, Explicit compactly supported smooth cutoffs , Dominated convergence).

[A4]

yD(T) exactly when xTx,y is bounded on D(T), and then Tx,y=x,Ty for all xD(T) (Adjoint of a densely defined operator, The adjoint is well defined, closed, and reverses inclusions).

[A5]

Under Countable Choice, complex L2 is a Hilbert space of almost-everywhere classes with first-variable-linear pairing f,g=fg (L2 with the integral pairing is a Hilbert space, The space Lp(μ) as the quotient by null functions). The pairing satisfies Cauchy–Schwarz; taking h and 1 on an interval of length at most one gives hh2 (The complex L2 pairing is well-defined and satisfies Cauchy–Schwarz).

Counterexample

technique · direct

Given: H=L2((0,1);C) and T=id/dx on the domain D(T) above.

1.1

An absolutely continuous representative is continuous and unique in its almost-everywhere class: a nonzero difference at a point would stay nonzero on a relative interval of positive length. It is bounded on [0,1] and thus belongs to L2; its a.e. derivative is independent of the representative. The stated domains are linear and define linear operators. Complex absolutely continuous calculus is obtained componentwise from the real theory. If F,G are complex-valued with absolutely continuous real and imaginary parts, then F(x)F(0)=0xF for all x, and 01FG=F(1)G(1)F(0)G(0)01FG; if hL2(0,1) then x0xh is of this kind with derivative h almost everywhere by the first L1 fundamental theorem, using the inclusion L2L1 in [A5]. These are the uses of the stated Countable Choice and Dependent Choice through the calculus and Hilbert-space suppliers.

A2A5
2.1

Given uH, extend it by zero to u0 on R. By [A3] and Countable Choice, choose complex φnCc(R) with φnu02<1/n, using real and imaginary approximations if necessary. For n5 put ζn(x)=(1χ(nx))(1χ(n(x1))) on (0,1) and extend it by zero outside. It vanishes in neighborhoods of both endpoints, so this extension is smooth with compact support in (0,1). Also 0ζn1 and ζn(x)1 for every x(0,1). Then ψn=ζnφnCc(0,1)D(T) and ψnu21/n+(ζn1)u20 by domination by u2 for the squared error. Thus D(T) is dense.

A3A5step 1.1
2.2

Symmetry: for f,gD(T) the boundary term in 1.1 vanishes because f(0)=f(1)=0, so Tf,g=i01fg=01fig=f,Tg; hence TT, and T is symmetric.

A1step 1.1
2.3

Closedness: let fnD(T) with fnf and Tfnw in L2. Then fniw in L2, so by 1.1 the functions Gn(x):=0xfn converge uniformly to G(x):=0xiw on [0,1], since supxGn(x)G(x)fniw2 by [A5], with Gn=fnfn(0)=fn; hence fnG pointwise and in L2, so f=G, that is, f is absolutely continuous with f=iw almost everywhere, f(0)=G(0)=0 and f(1)=G(1)=limnfn(1)=0. Thus fD(T) and Tf=if=i(iw)=w, so T is closed. The sequential graph criterion applies in this metric product under the assumed Countable Choice.

A5step 1.1
2.4

Adjoint computed. If gAC[0,1] has gL2(0,1), then for every fD(T) the identity in step 1.1 read backwards gives Tf,g=f,ig, so gD(T) and Tg=ig.

A4step 1.1
2.5

Conversely let gD(T) and put h:=Tg; by 1.1 the function H(x):=0xh is absolutely continuous with H=h and H(0)=0. For every fD(T) integration by parts in the form of 1.1 gives 01fh=f(1)H(1)f(0)H(0)01fH=01fH, while 01fh=f,Tg=Tf,g=i01fg. Hence 01f(Hig)=0 for every fD(T).

A4step 1.1
3.1

The derivatives of elements of D(T) are exactly E={φL2(0,1):01φ=0}: one inclusion follows from step 1.1, and conversely f(x)=0xφ has derivative φ, vanishes at both endpoints, and belongs to D(T). Set k=H+igL2 and m=01k. Step 2.5 says φk=0 for every φE. Taking φ=km gives 0=(km)k=km2, since (km)=0. Thus k=m as a class and g=iHim. This supplies an absolutely continuous representative of g with g=ih a.e., so h=ig. No unproved orthogonal-hyperplane assertion is needed.

A5step 1.1step 2.5
4.1

By steps 2.4 and 3.1, D(T)={gAC[0,1]:gL2(0,1)} and Tg=ig. This is strictly larger than D(T): the constant function 1 lies in D(T) but not in D(T). Hence TT, and T is a proper closed extension of T by [A4]. Moreover T is not symmetric: for g1(x)=x, g2=1 one computes from step 1.1 that Tg1,g2g1,Tg2=i(g1(1)g2(1)g1(0)g2(0))=i0.

A4step 2.2step 2.4step 3.1
5.1

Let Sg=ig on the periodic domain D2. It is densely defined because it extends T. For f,gD2 the boundary form in step 1.1 vanishes, so S is symmetric. If gD(S), the adjoint identity restricted to D(T) gives gD(T) and Sg=Tg=ig by [A4] and step 4.1. For arbitrary fD2, integration by parts therefore gives 0=i(f(1)g(1)f(0)g(0)). Choose f=1D2: then g(1)=g(0), so gD2. Conversely the boundary form vanishes for every periodic g, giving gD(S) and Sg=Sg. Hence S=S with equality of domains; S is self-adjoint and therefore closed by [A1]. The constant function lies in D2D(T), and xx lies in D(T)D2, proving both strict inclusions.

A1A4step 1.1step 2.1step 4.1
6.1

Every claim is witnessed: T is densely defined and closed by steps 2.1 and 2.3, symmetric by step 2.2, and TT by step 4.1; the failure of "symmetric implies self-adjoint" is therefore established, and no claim is made that a self-adjoint extension does not exist: the periodic domain of step 5.1 provides one by its explicitly computed adjoint.

step 2.1step 2.2step 2.3step 4.1step 5.1
DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Resolvent and spectrum of an unbounded operator

Definition

Let H be a complex Hilbert space and T a linear operator on H with domain D(T). Write (zT)x=zxTx for xD(T). The resolvent set of T is ρ(T):={zC: zT:D(T)H is bijective and (zT)1B(H)}, where B(H) is the space of bounded everywhere defined operators on H (A bounded linear operator between normed spaces); the spectrum is σ(T):=Cρ(T). For zρ(T) the bounded operator RT(z):=(zT)1B(H) is the resolvent of T at z. This is the library convention used throughout the page, matching the bounded resolvent of Spectrum and resolvent of a bounded operator: the shift is zIT, with coefficient 1 on T.

The resolvent determines T back. If zρ(T) and R:=RT(z), then D(T)=ranR and TRy=zRyy for yH; equivalently T=zIR1 with D(T)=ranR, and R(zT)x=x for xD(T), (zT)Ry=y for yH.

Nonempty resolvent set forces closedness, without the closed graph theorem. Assume ρ(T) and fix zρ(T) with R:=RT(z). Choose a bound C0 such that RvCv. The map F(v,w)=wRv is continuous, since F(v,w)F(v,w)ww+Cvv. Its zero set Γ(R)=F1({0}) is closed. Because D(T)=ranR and (zT)Rv=v for every vH, the graph of T is Γ(T)={(x,Tx):xD(T)}={(Rv,zRvv):vH}=Φ(Γ(R)), where Φ(v,w):=(w,zwv) is a continuous linear bijection of HH with continuous inverse Ψ(u,v)=(zuv,u), since Φ(Ψ(u,v))=(u,zu(zuv))=(u,v) and Ψ(Φ(v,w))=Ψ(w,zwv)=(v,w). A homeomorphism carries closed sets to closed sets, so Γ(T) is closed and T is closed. No closed graph theorem and hence no choice principle is used here.

For the zero operator on a nonzero H, ρ(0)=C{0}: the nonzero shifts have inverse z1I, whereas the zero shift is not bijective. If H={0}, every shift is the unique bijection of the zero space, so ρ(T)=C and σ(T)=.

TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Resolvent of a self-adjoint operator: nonreal resolvents and the estimate

Statement

Assume Countable Choice. Let T be a self-adjoint operator on H. Then every nonreal number belongs to ρ(T): CRρ(T). More precisely, for z=a+ib with a,bR, b0, and every xD(T), (Tz)x2=(Ta)x2+b2x2, and consequently RT(z)1/Imz. In particular σ(T)R.

Facts & Assumptions

[A1]

T=T; thus D(T) is dense, TT, and T is closed, T being closed (Symmetric, self-adjoint and essentially self-adjoint operators, The adjoint is well defined, closed, and reverses inclusions).

[A2]

For xD(T) one has Tx,y=x,Ty whenever yD(T), and Tx,x is a real number: it equals x,Tx=Tx,x (Symmetric, self-adjoint and essentially self-adjoint operators, Real and complex inner-product spaces and their induced length).

[A3]

For zC the identity ran(Tz)=ker(Tz)=ker(Tz) holds (The adjoint is well defined, closed, and reverses inclusions, [A1]).

[A4]

If M is a linear subspace of a Hilbert space, then M=M (The double orthogonal complement of a subspace is its closure, Orthogonality and the orthogonal complement).

[A5]

zρ(T) means that zT is a bijection of D(T) onto H with bounded inverse, and then RT(z) is the operator norm of that inverse (Resolvent and spectrum of an unbounded operator, The operator norm as the least bound and as the unit-sphere or unit-ball supremum).

Proof

technique · direct

Given: Countable Choice, a self-adjoint operator T on H, and a number z=a+ib with b0.

1.1

For xD(T), expanding (Tz)x2=(Tz)x,(Tz)x gives Tx2zTx,xzx,Tx+z2x2; by [A2] the two middle terms combine to 2aTx,x, so (Tz)x2=Tx22aTx,x+z2x2.

A2
2.1

Adding and subtracting a2x2 and using z2a2=b2, step 1.1 becomes (Tz)x2=(Ta)x2+b2x2.

step 1.1algebra
3.1

By step 2.1, (Tz)xbx for every xD(T); in particular Tz is injective and its range is closed: if (Tz)xny, then (xn) is Cauchy, hence xnx for some xH and Txnzx+y, and closedness of T gives xD(T) and Tx=zx+y, that is y=(Tz)x.

A1step 2.1
4.1

Also by [A3] applied to z, ran(Tz)=ker(Tz), and step 2.1 with z replaced by z shows Tz is injective, so the kernel is {0}. Hence the closed range of step 3.1 satisfies ran(Tz)=ran(Tz)=(ran(Tz))=H.

A3A4step 2.1step 3.1
5.1

By steps 3.1 and 4.1 the map zT:D(T)H is a bijection, and step 2.1 gives (zT)1yb1y for every yH: applying step 2.1 to x=(zT)1y yields y2=(Ta)x2+b2x2b2x2. Thus zρ(T) and RT(z)1/Imz by [A5].

A5step 2.1step 3.1step 4.1
6.1

Since z was an arbitrary nonreal number, CRρ(T), that is, σ(T)R; the identity and the bound of the statement are steps 2.1 and 5.1.

step 2.1step 5.1
TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Range criterion for self-adjointness

Statement

Assume Countable Choice. Let T be a densely defined symmetric operator on H. Then the following are equivalent:

  1. T is self-adjoint;
  2. T is closed and ker(Tz)={0} for every zCR;
  3. T is closed and ker(Ti)=ker(T+i)={0};
  4. ran(Tz)=H for every zCR;
  5. ran(Ti)=ran(T+i)=H;
  6. σ(T)R.

In particular a closed symmetric operator is self-adjoint if and only if ran(T±i)=H, and the same criterion holds with i replaced by iλ for any real λ>0.

Facts & Assumptions

[A1]

TT, D(T) is dense, and T is self-adjoint exactly when T=T; a self-adjoint operator is closed, and for x in the domain of a symmetric operator S the number Sx,x is real (Symmetric, self-adjoint and essentially self-adjoint operators, The adjoint is well defined, closed, and reverses inclusions, Real and complex inner-product spaces and their induced length).

[A2]

ran(Tz)=ker(Tz) for every zC (The adjoint is well defined, closed, and reverses inclusions).

[A3]

For a linear subspace M of a Hilbert space, M=M; in particular if M is closed and M={0} then M=H (Orthogonality and the orthogonal complement, The double orthogonal complement of a subspace is its closure).

[A4]

For self-adjoint T one has CRρ(T), hence σ(T)R; conversely every zρ(T) gives that zT is a bijection D(T)H (Resolvent of a self-adjoint operator: nonreal resolvents and the estimate, Resolvent and spectrum of an unbounded operator).

[A5]

If ρ(T) then T is closed (Resolvent and spectrum of an unbounded operator).

Proof

technique · direct

Given: A densely defined symmetric operator T on H.

1.1

Preparatory identity. Let S be densely defined and symmetric, let z=a+ib with b0 and let xD(S). Expanding (Sz)x2 and using that Sx,x is real by [A1] gives, as in the proof of Resolvent of a self-adjoint operator: nonreal resolvents and the estimate, (Sz)x2=(Sa)x2+b2x2  b2x2.

A1
1.2

(1) implies (2): if T=T then T is closed by [A1]. If z is nonreal and vker(Tz), then vD(T)=D(T) and zv2=Tv,v=zv2 by [A1], so v=0.

A1
1.3

(4) implies (5) is immediate, since ±i are nonreal.

given
1.4

(5) implies (1): let vD(T). Since ran(Ti)=H, choose wD(T) with (Ti)w=(Ti)v. Then (Ti)(vw)=0, because Tw=Tw for wD(T); and by [A2] with z=i the kernel of Ti is ran(T+i)={0}. Hence v=wD(T). So D(T)D(T), and with TT this gives T=T.

A1A2
1.5

(1) implies (6) by [A4]. Conversely (6) implies (4): if σ(T)R then every nonreal z lies in ρ(T), so zT is surjective and ran(Tz)=H.

A4
2.1

Closed range for closed symmetric S. If in addition S is closed, then ran(Sz) is closed for nonreal z: given (Sz)xny, step 1.1 makes (xn) Cauchy with limit x, so Sxnzx+y, and closedness of S gives xD(S) with Sx=zx+y, that is y=(Sz)x.

A1step 1.1
2.2

(5) implies (3): by 1.1 with S=T and z=i the map T+i satisfies (T+i)xx, so its inverse on its range is bounded by 1; since the range is H by (5), iρ(T) because iT=(T+i), and T is closed by [A5]. The two kernels vanish by [A2] and (5).

A2A5step 1.1
3.1

(2) implies (4): for nonreal z, [A2] and (2) give ran(Tz)=ker(Tz)={0}, while (2) and step 2.1 show ran(Tz) is closed; hence the range is all of H by [A3].

A2A3step 2.1
3.2

(3) implies (5): by [A2], ran(Ti)=ker(T±i)={0}, and by step 2.1 applied to the closed T with z=i the two ranges are closed; hence they equal H by [A3]. So (3) and (5) are equivalent.

A2A3step 2.1
4.1

Collecting: (1)(2)(4)(5)(1) by steps 1.2, 3.1, 1.3 and 1.4; (5)(3) and (3)(5) by steps 2.2 and 3.2; and (1)(6)(4) by step 1.5. Thus all six statements are equivalent. The final clause follows because only nonreality of the parameters was used, so iλ with λ>0 may replace i.

step 1.2step 3.1step 1.3step 1.4step 2.2step 3.2step 1.5
DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Cayley transform of a self-adjoint operator

Definition

Assume Countable Choice (The Axiom of Countable Choice (ACω)). Let T be a self-adjoint operator on H. By Resolvent of a self-adjoint operator: nonreal resolvents and the estimate the points ±i lie in ρ(T), so T+i and Ti are bijections of D(T) onto H with bounded inverses, and CT:=(Ti)(T+i)1B(H) is a bounded everywhere defined operator, the Cayley transform of T. In the resolvent convention RT(z)=(zT)1 of Resolvent and spectrum of an unbounded operator one has (T+i)1=RT(i),CT=I+2iRT(i)=I2i(T+i)1.

Its properties, with proofs. Writing C:=CT:

  1. C is isometric. For xH put y:=(T+i)1x, so that x=(T+i)y with yD(T); then Cx=(Ti)y, and the symmetry computation (Ti)y2=Ty2+y2 of Resolvent of a self-adjoint operator: nonreal resolvents and the estimate gives Cx=x.
  2. C is unitary, with C1=C=(T+i)(Ti)1. Using C=I2i(T+i)1 and the adjoint rule ((T+i)1)=((T+i))1=(Ti)1 for the self-adjoint T one gets C=I+2i(Ti)1=(T+i)(Ti)1; the elementary resolvent identity then gives CC=CC=I, as follows also from applying the computation of item 1 to C and to C and using that a surjective isometry of H onto H is unitary (A bounded linear operator between normed spaces). Concretely CCx=x for xD(T) by direct substitution, and both sides are continuous.
  3. ker(IC)={0}, and IC=2i(T+i)1, I+C=2T(T+i)1 as maps on H. Indeed IC=2i(T+i)1 is injective with inverse (2i)1(T+i), while (I+C)(T+i)x=(T+i)x+(Ti)x=2Tx for xD(T).
  4. Domain recovery. ran(IC)=D(T): by item 3, ran(IC)=ran(2i(T+i)1)=D(T).
TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Cayley correspondence between self-adjoint operators and unitaries

Statement

Assume Countable Choice. The map TCT=(Ti)(T+i)1 is a bijection from the set of self-adjoint operators on H onto the set of unitary operators U on H with ker(IU)={0}. The inverse map assigns to such a U the operator D(TU)=ran(IU),TU(IU)y=i(I+U)y(yH). For this TU both TUi and TU+i map D(TU) onto H, so TU is self-adjoint, and CTU=U.

Facts & Assumptions

[A1]

For self-adjoint T the Cayley transform CT is unitary, ker(ICT)={0}, and ran(ICT)=D(T); also (Ti)x2=Tx2+x2 for xD(T) (Cayley transform of a self-adjoint operator, Resolvent of a self-adjoint operator: nonreal resolvents and the estimate).

[A2]

A densely defined symmetric operator T is self-adjoint if ran(Ti)=ran(T+i)=H (Range criterion for self-adjointness).

[A3]

For unitary U one has UU=UU=I and U is bijective, with U1=U; moreover ran(A)=ker(A) for bounded A, so ran(IU) is dense exactly when ker(IU)={0} (Hilbert-adjoint identities, Kernel–range orthogonality for Hilbert adjoints, Orthogonality and the orthogonal complement, Dense, nowhere dense and codense subsets of a topological space, and the criterion by basic open sets).

Proof

technique · direct

Given: A unitary U with ker(IU)={0}, and the map TCT of [A1].

1.1

The assignment is well defined: if y,yH satisfy (IU)y=(IU)y, then y=y because ker(IU)={0}; hence D(TU):=ran(IU) is well defined and TU(IU)y:=i(I+U)y defines a map on D(TU).

A3given
1.2

D(TU) is dense: by [A3], ran(IU)=ker(IU), and IU=IU1 has kernel {0} because (IU1)y=0 means y=Uy, that is yker(IU)={0}.

A3given
1.3

Conversely TCT=T for every self-adjoint T: by [A1] ran(ICT)=D(T) and ICT=2i(T+i)1, so for xD(T) one has (ICT)(T+i)x=2ix and i(I+CT)(T+i)x=i(2Tx); hence the inverse construction sends CT to T.

A1
2.1

TU is symmetric: for y,yH, expanding both pairings and using Uy,Uy=y,y, one gets TU(IU)y,(IU)y=i(Uy,yy,Uy)=(IU)y,TU(IU)y.

A3step 1.1
2.2

Ranges: for every yH one has (TU+i)(IU)y=2iy and (TUi)(IU)y=2iUy; hence ran(TU+i)=H and, since U is onto by [A3], ran(TUi)=H.

A1A3step 1.1
3.1

By steps 1.2, 2.1 and 2.2 the operator TU is densely defined, symmetric, and has both ranges equal to H, so TU is self-adjoint by [A2].

A2step 1.2step 2.1step 2.2
4.1

CTU=U: by [A1] applied to the self-adjoint TU and by step 2.2 one has (TU+i)1(2iy)=(IU)y, so CTU(2iy)=(TUi)(IU)y=2iUy; since y2iy is onto H, CTU=U.

A1step 2.2step 3.1
5.1

By steps 1.3 and 4.1 the two constructions are mutually inverse, and every assignment above is a bijection by construction; hence TCT is a bijection onto the stated class.

step 1.3step 3.1step 4.1
DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Integral of a measurable function against a projection-valued measure

Definition

Assume Countable Choice. Let E be a projection valued measure on the measurable space (X,Σ) acting on the complex Hilbert space H (Projection valued measure), and let f:XC be Σ-measurable. Put D(f(E)):={xH:Xf2dEx<},Ex(B)=E(B)x,x, where the scalar measures Ex are those of Scalar and complex measures from a pvm, and for xD(f(E)) set f(E)x:=limnΦE(f1{fn})x, For H{0}, the bounded integrals ΦE are those of Bounded borel pvm integral, and the limit is taken in the norm of H. For H={0}, define ΦE(g) to be the unique operator on H for every bounded measurable g. Then E0 is the zero measure, D(f(E))=H, and f(E)0=0. Thus this case is defined directly without applying a theorem requiring a nonzero space.

Well-definedness. D(f(E)) is a linear subspace: the estimate Ex+y2Ex+2Ey for scalar measures and Eαx=α2Ex follow from E(B)(x+y)22E(B)x2+2E(B)y2 and E(B)αx2=α2E(B)x2. Writing fn:=f1{fn} the sets {fn} are measurable, so fn is bounded and measurable. For m,nN, fnfm2f21{f>N}. The integral of the right side against Ex tends to zero by Dominated convergence, with the integrable majorant f2 and pointwise limit zero since f is finite-valued. Linearity of the bounded calculus (Pvm integral is a star homomorphism) and its quadratic identity give ΦE(fn)xΦE(fm)x2=fnfm2dEx. These clauses hold directly on the zero space as well. Hence the truncation vectors are Cauchy and converge uniquely by Hilbert-space completeness (Hilbert space). The limit is linear in x because every truncation operator is linear and addition and scalar multiplication are norm-continuous. Countable Choice is inherited from the bounded PVM suppliers; no further choice is needed to take these specified limits.

Finally, if two Σ-measurable functions f,g agree outside a measurable set N with E(N)=0, then Ex(N)=E(N)x2=0 for every x. Thus the integrals of f2 and g2 agree, so their domains agree. At every truncation level their bounded truncations agree outside N; the quadratic identity applied to the difference gives equal truncation vectors for every x. Taking limits gives f(E)x=g(E)x on their common domain.

LemmaStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

The unbounded PVM integral is densely defined, closed and normal

Statement

Assume Countable Choice. Let E be a projection valued measure on (X,Σ) on the complex Hilbert space H and let f:XC be Σ-measurable. Then the operator f(E) of Integral of a measurable function against a projection-valued measure is densely defined and closed, and it is normal, in the sense D(f(E))=D(f(E)) and equal norms for the operator and its adjoint; in particular D(f(E))=D(f(E)) and f(E)x=f(E)x for all such x. Moreover f(E)x2=f2dEx,x,f(E)x=fdEx(xD(f(E))), and (f(E))=f(E); in particular f(E) is self-adjoint whenever f takes real values. Finally, if (gn) are bounded Σ-measurable functions with gnCf pointwise and gnf pointwise, then gn(E)xf(E)x for every xD(f(E)).

Facts & Assumptions

[A1]

D(f(E))={x:f2dEx<} is a linear subspace, f(E)x=limnfn(E)x with fn=f1{fn}, and the limit is linear in x; here and below ΦE(h)=h(E) for bounded Σ-measurable h (Integral of a measurable function against a projection-valued measure).

[A2]

For H{0} and bounded Σ-measurable h: h(E)x2=h2dEx, h(E)x,y=hdEx,y, h(E)=h(E), products of bounded Σ-measurable functions multiply as ΦE(h1h2)=ΦE(h1)ΦE(h2), and ΦE is linear (Bounded borel pvm integral, Pvm integral is a star homomorphism, Scalar and complex measures from a pvm).

[A3]

Dominated convergence on the finite measure Ex: if unu pointwise and unG with GdEx<, then unudEx0. In particular, if hnh and hnh2G with integrable G, then hnh2dEx0 by applying the theorem to the squared differences (Dominated convergence, Scalar and complex measures from a pvm).

[A4]

Scalar monotone convergence: for nonnegative measurable unu the integrals undEx increase to udEx (Monotone convergence for the integral).

[A5]

The adjoint of a densely defined operator is closed, and D(T) consists of those y for which zTz,y is bounded, with Ty the representing vector (Adjoint of a densely defined operator, The adjoint is well defined, closed, and reverses inclusions).

[A6]

Projection values satisfy E(B)2=E(B)=E(B), E(BC)=E(B)E(C), E(X)=I and E(B)xx (Projection valued measure). Closedness means the graph is closed, and density means the domain is dense (Densely defined, closed and closable operators, and cores).

Proof

technique · direct

Given: A PVM E and a Σ-measurable function f as in the statement.

1.1

If H={0}, [A1] defines all integrals on H as the unique zero-space operator. Its domain is all H, its graph is the whole HH, its adjoint is itself by the representing identity, and every scalar integral and norm in the statement is zero. All approximation vectors are zero. This proves every assertion in that case; henceforth assume H{0}, as required by [A2].

A1A5A6
1.2

For any measurable f and xD(f(E)), norm convergence of the defining truncations and monotone convergence give f(E)x2=limmfm2dEx=f2dEx. Also f1+f2 is integrable against the finite measure Ex. Thus dominated convergence in the bounded quadratic pairings gives x,f(E)x=limmfmdEx=fdEx, with the conjugate required by the first-variable-linear inner product.

A1A2A3A4
1.3

Fix a bounded measurable g and xD(f(E)). For each m, bounded linearity and the quadratic identity give g(E)xfm(E)x2=gfm2dEx. Let m; the left side converges by [A1], while the right side converges by [A3], since gfm22g2+2f2, an integrable majorant. Consequently g(E)xf(E)x2=gf2dEx. For the sequence gn in the statement, choose its bound C0; the majorant (C+1)2f2 and [A3] now imply the claimed convergence. This argument applies to any measurable target function in place of f.

A1A2A3
2.1

If h is bounded measurable, [A2] and [A6] give E(B)h(E)=(1Bh)(E)=h(E)E(B), and hence Eh(E)x(B)=(1Bh)(E)x2=Bh2dEx. Thus dEh(E)x=h2dEx for all xH. For xD(f(E)) this implies h(E)xD(f(E)), and also xD((hf)(E)), by the bound hf2h2f2. Bounded multiplication gives fm(E)h(E)x=h(E)fm(E)x=(hfm)(E)x. The first two limits follow from [A1] and bounded continuity; the last tends to (hf)(E)x by step 1.3 applied to the target hf, because hfmhf. Therefore f(E)h(E)x=h(E)f(E)x=(hf)(E)x on D(f(E)).

A1A2A6step 1.3
3.1

Put Ωn={fn} for n1. For every xH, step 2.1 gives dEE(Ωn)x=1ΩndEx, so E(Ωn)xD(f(E)). Moreover xE(Ωn)x2=1XΩndEx0 by [A2], [A6] and dominated convergence, since f is finite-valued. Hence the domain is dense. For every wH, the defining truncations on E(Ωn)w are constant for mn: fm(E)E(Ωn)w=fn(E)w. Therefore f(E)E(Ωn)w=fn(E)w.

A1A2A3A6step 2.1
4.1

The domains of f(E) and f(E) coincide because their defining squared moduli agree. For x,y in this domain, bounded adjoints and the defining limits yield f(E)y,x=limmfm(E)y,x=limmy,fm(E)x=y,f(E)x. Since density is established in step 3.1, [A5] gives f(E)f(E).

A1A2A5step 3.1
5.1

Conversely let yD(f(E)) and z=f(E)y. For every wH, step 3.1 and the adjoint identity give fn(E)y,w=y,fn(E)w=y,f(E)E(Ωn)w=z,E(Ωn)w=E(Ωn)z,w. Thus fn(E)y=E(Ωn)z. By [A2], [A4] and the projection bound, f2dEy=limnfn(E)y2z2<. Hence yD(f(E)), and step 4.1 proves f(E)=f(E) with equal domains.

A1A2A4A5A6step 3.1step 4.1
6.1

Apply step 5.1 to the measurable function f, whose integral has dense domain by step 3.1. It gives (f(E))=f(E), so [A5] proves f(E) closed directly. Step 1.2 and f=f give equal norms for f(E) and its adjoint on their common domain, establishing normality. If f is real-valued, step 5.1 gives f(E)=f(E). Together with steps 1.2, 1.3 and 3.1 this proves every assertion.

A5step 1.2step 1.3step 3.1step 5.1
TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Spectral theorem for unbounded self-adjoint operators (PVM form)

Statement

Assume the Axiom of Choice. Let T be a self-adjoint operator on the nonzero complex Hilbert space H. Then there is a unique regular projection valued measure E on the Borel σ-algebra of R such that

D(T)={xH:Rλ2dEx(λ)<},Tx=RλdE(λ)x(xD(T)).

Conversely, if E is a regular projection valued measure on R, then the operator λdE with that domain is self-adjoint and its spectral projection valued measure is E again.

Facts & Assumptions

[A1]

For a Borel function f the integral f(E) of Integral of a measurable function against a projection-valued measure has domain {x:f2dEx<}, is closed, satisfies (f(E))=f(E) and f(E)x2=f2dEx, and is self-adjoint for real f (The unbounded PVM integral is densely defined, closed and normal).

[A2]

Every bounded normal operator C has a unique regular spectral PVM F with C=zdF; the support of F is σ(C), and F is unique among regular PVMs representing C on a compact set (Spectral theorem for bounded normal operators pvm form, Support and uniqueness of the spectral measure).

[A3]

CT=(Ti)(T+i)1 is unitary with ker(ICT)={0}, ran(ICT)=D(T), and T=i(I+CT)(ICT)1 on D(T) (Cayley transform of a self-adjoint operator, Cayley correspondence between self-adjoint operators and unitaries).

[A4]

For a unitary C one has σ(C)S1={z:z=1}: z>C forces zρ(C) by the Neumann series, so σ(C) lies in the closed unit disk, and 0σ(C) with zσ(C) equivalent to z1σ(C1) (Neumann series, Spectrum and resolvent of a bounded operator).

[A5]

For bounded Borel h on S1 the identity hdF=ΦF(h) holds. If a PVM F on S1 satisfies F({1})=0 and ψ:RS1{1} is a Borel isomorphism, then E(B):=F(ψ(B)) defines a PVM on R with (hψ)dE=hdF; regularity is preserved by this transport (Bounded borel pvm integral, Projection valued measure, Regular Borel measure on an LCH space).

Proof

technique · direct

Given: A self-adjoint operator T on H and C:=CT.

1.1

By [A3] the operator C is unitary with ker(IC)={0}, and by [A4] its spectrum lies in S1. Let F0 be the regular spectral PVM of C on σ(C) supplied by [A2], and extend it to S1 by F(B):=F0(Bσ(C)) for Borel BS1. Then F is a regular PVM on S1, is carried by σ(C), and C=S1zdF(z).

A2A3A4
1.2

F({1})=0: for a bounded normal operator the spectral projection at a point λ is the orthogonal projection onto ker(Cλ). Indeed, if F({λ})x=x, then Fx is carried by {λ}, so the bounded calculus gives (Cλ)x2=S1zλ2dFx(z)=0. Conversely, if Cx=λx, the same identity shows that Fx is carried by {λ}; hence (IF({λ}))x2=Fx(S1{λ})=0 and F({λ})x=x. Thus ranF({λ})=ker(Cλ); with λ=1 this kernel is {0} by [A3], so the projection F({1}) is zero.

A2A3A5step 1.1
2.1

Let ψ(λ)=(λi)(λ+i)1. Then ψ is a homeomorphism of R onto S1{1}, and E(B):=F(ψ(B)) is a PVM on the Borel sets of R with E(R)=F(S1{1})=IF({1})=I; it is regular because F is and ψ is a homeomorphism, and C=zdF=ψ(λ)dE(λ).

A5step 1.2
3.1

Let A be the self-adjoint operator λdE with domain D(A)={x:λ2dEx<}, given by [A1]; write h(λ)=(λ+i)1, a bounded Borel function. Then 1C=(1ψ)dE=2ihdE=2ih(E), and (A+i)h(E)=I because (λ+i)h(λ)=1 and Ah(E)=(λh)(E) on the natural domain; hence h(E)=(A+i)1 and (A+i)(1C)=2iI, so A+i=2i(1C)1 on ran(1C)=D(A).

A1A3A5step 2.1
4.1

Using step 3.1 in the formula T=i(I+C)(IC)1 of [A3] gives T=(1/2)(I+C)(A+i); now I+C=(1+ψ)dE=2λh(λ)dE=2λh(λ)(E), so T=(λh)(E)(A+i)=(λ2h)(E)+i(λh)(E)=(λ)(E)=A on D(A), because λ2h+iλh=λh(λ+i)=λ.

A1A3step 3.1
4.2

Uniqueness: let E be a regular PVM on R with D(T)={x:λ2dEx<} and Tx=λdEx for xD(T), and put F(B):=E(ψ1(B{1})) for Borel BS1. Then F is a regular PVM on S1, F({1})=0, and zdF=ψdE=I2ih(E)=I2i(T+i)1=C, where (T+i)1=h(E) is proved as in step 3.1 with E in place of E. By the uniqueness clause of [A2] applied to the bounded normal operator C, F is carried by σ(C) and agrees there with F0, hence F=F on S1. Therefore E(B)=F(ψ(B))=F(ψ(B))=E(B).

A2A5step 2.1step 3.1
5.1

Conversely, if E is a regular PVM on R, then A=λdE is self-adjoint by [A1] and E represents A; by step 4.2 the representing PVM is unique, so E is the spectral PVM of A.

A1step 4.2
TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Unbounded Borel functional calculus: domains, products, spectral mapping

Statement

Assume the Axiom of Choice. Let T be a self-adjoint operator with spectral projection valued measure E on R acting on a complex Hilbert space H (Spectral theorem for unbounded self-adjoint operators (PVM form)) and let f,g:RC be Borel. Write u(T)=u(E), with the truncation definition below. If H={0}, use its unique PVM and unique full-domain operator directly. Then:

  1. f(T)=f(T);
  2. f(T)g(T) has domain D(g(T))D((fg)(T)) and equals the restriction of (fg)(T) to that domain, and its closure is (fg)(T);
  3. on D(f(T))D(g(T)) the sum f(T)+g(T) equals the restriction of (f+g)(T), and the closure of f(T)+g(T) is (f+g)(T);
  4. the spectrum of f(T) is the essential range {zC:E(f1(Bε(z)))0 for every ε>0} of f with respect to E;
  5. if f is continuous then that essential range is the closure of f(σ(T)), with the closure redundant when f(σ(T)) is closed.

Facts & Assumptions

[A1]

For every finite-valued measurable u, u(E) has domain Du={x:u2dEx<}, is densely defined and closed, and satisfies u(E)x2=u2dEx and u(E)=u(E). It is the norm limit of um(E)x, where um=u1{um}. Bounded approximants dominated by Cu and converging pointwise to u converge on Du (The unbounded PVM integral is densely defined, closed and normal, Integral of a measurable function against a projection-valued measure).

[A2]

On nonzero H the bounded measurable PVM calculus is linear, unital, multiplicative and adjoint preserving, and h(E)x2=h2dEx. The scalar measure Ex is finite with total mass x2 (Bounded borel pvm integral, Pvm integral is a star homomorphism, Scalar and complex measures from a pvm).

[A3]

Dominated convergence holds for an integrable majorant, and monotone convergence holds for increasing nonnegative measurable functions (Dominated convergence, Monotone convergence for the integral).

[A4]

A spectral PVM represents T as the integral of the identity function with its exact squared-integrability domain; the cited spectral theorem assumes nonzero H and AC (Spectral theorem for unbounded self-adjoint operators (PVM form), The Axiom of Choice). The resolvent convention is (zIS)1, required bounded and everywhere defined (Resolvent and spectrum of an unbounded operator).

[A5]

E(BC)=E(B)E(C), projection values are contractive, and E is strongly countably additive with E(R)=I (Projection valued measure). Rational intervals form a countable base of R, by countability and density of the rationals (Q is countably infinite, Both Q and RQ are dense in R, and every nonempty open subset of R is uncountable).

Proof

technique · direct

Given: The self-adjoint operator and spectral PVM in the statement, under AC.

1.1

For H={0} all integrals are the unique full-domain operator by [A1]. Sums, products, adjoints and closures are that operator; its resolvent is all C and its spectrum is empty. Every spectral projection is zero, so every essential range here is empty, as is f(σ(T)). Thus all claims hold. Henceforth H{0}, and all bounded calculus uses have the hypothesis required by [A2].

A1A2A4A5
1.2

For bounded Borel h and arbitrary xH, bounded multiplication gives E(B)h(E)=(1Bh)(E), so Eh(E)x(B)=Bh2dEx. For xDg, truncate g: norm convergence and boundedness of E(B) give Eg(E)x(B)=limmEgm(E)x(B)=Bg2dEx by monotone convergence. Hence dEg(E)x=g2dEx, including when g is unbounded.

A1A2A3A5
1.3

We will also use a dominated approximation with a general square-integrable majorant. Fix xDu and bounded Borel h. Comparing with um using [A2] and then passing m gives h(E)xu(E)x2=hu2dEx; the scalar limit is dominated by 2h2+2u2. Therefore if bounded hnu pointwise and hnG with G2dEx<, then u is square integrable and hn(E)xu(E)x, by dominated convergence with hnu24G2.

A1A2A3
2.1

For xDfDg, xDf+g since f+g22f2+2g2. Bounded linearity and step 1.3 applied to fm+gm with G=f+g show f(E)x+g(E)x=(f+g)(E)x. In particular Duz=Du and (uz)(E)=u(E)zI, since constants are integrable against finite Ex.

A1A2step 1.3
2.2

If h is bounded and xDg, step 1.2 shows h(E)xDg, while xDhg. The bounded identities h(E)gm(E)x=gm(E)h(E)x=(hgm)(E)x pass to limits: the last uses [A1] with target hg since hgmhg. Thus h(E)g(E)x=g(E)h(E)x=(hg)(E)x for xDg.

A1A2step 1.2
3.1

The definition of composition and step 1.2 yield D(f(E)g(E))=DgDfg. For x in that domain, step 2.2 gives fm(E)g(E)x=(fmg)(E)x. The left side tends to f(E)g(E)x by [A1]. To control the right side without pretending fmg bounded, note that (fmg)(E)x=E({fm})(fg)(E)x: apply step 2.2 with g there replaced by fg and the bounded indicator. The projections converge strongly to I, since their complementary squared norms are integrals of decreasing indicators against finite scalar measures. Hence the right side tends to (fg)(E)x, proving the product value.

A1A2A3A5step 1.2step 2.2
3.2

For xDf+g use xm=E({fm,gm})x. Step 1.2 shows xmDfDg, and gives xmx and (f+g)(E)(xmx)2={f>m}{g>m}f+g2dEx0. Step 2.1 and closedness in [A1] prove the sum closure. These cutoff ranges also show the sum domain dense.

A1A2A3A5step 1.2step 2.1
4.1

For xDfg put xm=E({gm})x. Step 1.2 shows xmDgDfg. The same step and dominated convergence give xmx and (fg)(E)(xmx)2={g>m}fg2dEx0. Thus every point of the graph of (fg)(E) is a limit of graph points of f(E)g(E). The reverse graph inclusion follows from step 3.1 and closedness in [A1], proving the product closure. The composition domain is dense as well: the ranges of E({fm,gm}), contained in that domain, approximate every vector by the same indicator estimate.

A1A2A3A5step 1.2step 3.1
4.2

Suppose E({fz<ε})=0 for some ε>0. Define the Borel function w piecewise: w(λ)=1/(zf(λ)) when f(λ)zε, and w(λ)=0 otherwise. It is bounded by 1/ε, and (zf)w=1 off an E-null set. By the domain and value identities of step 3.1, (zf)(E)w(E) has domain all H and equals I; w(E)(zf)(E) has domain Dzf and equals the identity there. Null-set invariance is supplied by [A1]. By step 2.1, Dzf=Df and (zf)(E)=zIf(E) (linearity with constants, or multiplication by 1 in step 2.2). Thus w(E) is its bounded inverse and zρ(f(E)) in the convention of [A4]. Only the inverse is asserted bounded.

A1A2A4step 2.1step 2.2step 3.1
5.1

Conversely suppose E({fz<ε})0 for every ε>0. Using AC in [A4], choose unit xn in the range of E({fz<1/n}) for each n1. The scalar measure of xn is supported there by [A5], and fz+1/n there, so xnDf. Hence (zIf(E))xn2=zf2dExn1/n2 by [A1] and step 2.1. A bounded inverse with bound C would give 1C/n for every n, impossible. This proves the essential-range formula.

A1A4A5step 2.1step 4.2
6.1

Apply steps 4.2 and 5.1 to u(λ)=λ, which represents T by [A4]. Nonreal z have a ball disjoint from R, so σ(T)R; for real λ, membership in σ(T) is equivalent to every interval about λ having nonzero projection. The union U of all rational intervals with zero projection is exactly Rσ(T), by the rational base in [A5] and projection monotonicity from E(BC)=E(B)E(C). Enumerate pairs of a fixed rational enumeration by increasing sums of their indices, giving an enumeration of rational intervals. In that enumeration retain such intervals and replace the rest by the empty set. Disjointify this sequence by subtracting previous intervals. Each resulting set has zero projection, and their union is U, so strong countable additivity gives E(U)=0. In particular U is open and measurable.

A4A5step 4.2step 5.1
7.1

Let f be continuous. If zf(σ(T)), a ball about z has preimage contained in U of step 6.1, so that preimage has zero projection. Conversely, for zf(σ(T)) and ε>0, choose λσ(T) with f(λ)z<ε/2. Continuity supplies an interval about λ whose image lies in Bε(z); its projection is nonzero by step 6.1, so the whole preimage has nonzero projection. This proves the continuous spectral-mapping formula. The adjoint identity is [A1], and steps 3.1, 4.1, 3.2, 4.2 and 5.1 prove the remaining claims.

A1A5step 3.1step 4.1step 3.2step 4.2step 5.1step 6.1
DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Strongly continuous one-parameter unitary group

Definition

A strongly continuous one-parameter unitary group on the complex Hilbert space H is a map U:RB(H) such that U(0)=I, U(s+t)=U(s)U(t) for all s,tR, each U(t) is unitary (that is, onto and norm preserving, equivalently U(t)U(t)=U(t)U(t)=I, A bounded linear operator between normed spaces, The operator norm as the least bound and as the unit-sphere or unit-ball supremum), and the orbit maps tU(t)x are continuous at every t for every xH, from the usual metric on R to the norm metric of H (Continuity of a map between metric spaces, at a point and globally, in the ε-δ form).

Continuity at the origin suffices, and weak continuity is equivalent to strong continuity. These two clauses are part of the definition's content and are proved here. If tU(t)x is continuous at t=0 and t0R, then U(t)xU(t0)x=U(t0)(U(tt0)xx) and U(t0) is isometric, so U(t)xU(t0)x=U(tt0)xx0 as tt0; the group law and isometry turn continuity at one point into continuity everywhere. Likewise, if tU(t)x is merely weakly continuous at 0, then for xH the expansion U(t)xx2=2x22ReU(t)x,x shows that norm convergence at t=0 follows from U(t)x,xx,x; and at an arbitrary t0 one has U(t)yU(t0)y=U(t0)(U(tt0)yy), so weak continuity at t0 for every y follows from the case t=0. Conversely, norm continuity of an orbit implies its weak continuity, since for each fixed yH, Cauchy--Schwarz gives U(t)xU(t0)x,yU(t)xU(t0)xy0. Finally, a group with U(0)=I is automatically invertible with U(t)1=U(t), so the unitarity and group clauses are symmetric in t.

DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Infinitesimal generator of a unitary group

Definition

Let U be a strongly continuous one-parameter unitary group on H (Strongly continuous one-parameter unitary group). Its infinitesimal generator is the linear operator G with domain D(G):={xH: limt01t(U(t)xx) exists in H},Gx:=limt01t(U(t)xx), where this is the two-sided norm limit over real t0: it has value y precisely when for every ε>0 there is δ>0 such that 0<t<δU(t)xxty<ε. Equivalently, defining the quotient's value at t=0 to be y makes it continuous there from the usual metric on R to the norm metric on H (Continuity of a map between metric spaces, at a point and globally, in the ε-δ form). Such a y is unique by the triangle inequality.

D(G) is a linear subspace and G is linear: if x,yD(G) and a,b are scalars, then 1t(U(t)(ax+by)(ax+by))=a1t(U(t)xx)+b1t(U(t)yy) converges with limit aGx+bGy, because the operations of H are continuous and both estimates use the same punctured real parameter t; the restriction to D(G) is therefore well defined and linear (Linear subspace of a vector space, Linear map between vector spaces over the same field).

Sign convention. This page writes T=iG for the generator, so that Stone's theorem reads U(t)=eitT with T self-adjoint; equivalently G=iT. In the convention of Teschl's book one has U(t)=eitA with A self-adjoint, so G=iA and hence A=iG=T; every formula below is written in the U(t)=eitT convention and the translation is recorded where a source is cited.

LemmaStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

A self-adjoint operator generates a strongly continuous unitary group

Statement

Assume the Axiom of Choice. Let T be a self-adjoint operator on H with spectral projection valued measure E, and set U(t):=eitT=ReitλdE(λ) for tR. Then U is a strongly continuous one-parameter unitary group, U(t)D(T)=D(T) and TU(t)=U(t)T on D(T), and the derivative limt01t(U(t)xx) exists exactly for xD(T), where it equals iTx. Thus the generator of U is G=iT, with D(G)=D(T).

Facts & Assumptions

[A1]

For bounded Borel h the operator h(T) is bounded with h(T)x2=h2dEx, h(T)=h(T), and h1(T)h2(T)=(h1h2)(T); the truncation definition gives g(T)x=limmgm(T)x for xDg (Unbounded Borel functional calculus: domains, products, spectral mapping, The unbounded PVM integral is densely defined, closed and normal, Integral of a measurable function against a projection-valued measure).

[A2]

D(T)={x:λ2dEx<}, and for xD(T) one has (Ta)x2=(λa)2dEx for real a; also U(t)D(T)=D(T) and TU(t)=U(t)T on D(T), since U(t)=eitλ(T) commutes with E(B) and products of functions multiply (Spectral theorem for unbounded self-adjoint operators (PVM form), Unbounded Borel functional calculus: domains, products, spectral mapping).

[A3]

Scalar dominated convergence and Fatou's lemma apply to the finite measures Ex (Dominated convergence, Fatou's lemma).

[A4]

The generator is defined by the difference quotients of the statement, and D(G) is a linear subspace (Infinitesimal generator of a unitary group).

Proof

technique · direct

Given: A self-adjoint T with spectral PVM E, and U(t)=eitT.

1.1

Unit and group law: U(t) is bounded, U(t)U(s)=(eitλeisλ)(T)=U(t+s) and U(0)=I by [A1], and U(t) is unitary because U(t)x2=eitλ2dEx=dEx=x2 and U(t)=U(t)=U(t)1.

A1
1.2

Strong continuity: for xH and tt0, U(t)xU(t0)x2=eitλeit0λ2dEx0 by [A3], the integrand being bounded by 4 and tending to 0 pointwise.

A1A3
1.3

Derivative at 0 for xD(T): 1t(U(t)xx)iTx2=1t(eitλ1)iλ2dEx0 by [A3], since 1t(eitλ1)iλ2λ and λ2 is Ex-integrable exactly because xD(T) by [A2].

A2A3
2.1

Converse: if z=limn1tn(U(tn)xx) for some sequence tn0, then by [A2] and Fatou's lemma λ2dExlim infn1tn(eitnλ1)2dEx=lim infn1tn(U(tn)xx)2=z2<, so xD(T); step 1.3 then identifies the full limit as iTx.

A2A3step 1.3
3.1

By steps 1.1, 1.2, 1.3 and 2.1 the family U is a strongly continuous one-parameter unitary group whose generator satisfies D(G)=D(T) and G=iT; the invariance and commutativity claims for U(t) on D(T) are [A2].

A2A4step 1.1step 1.2step 2.1
LemmaStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Laplace resolvents of a unitary group

Statement

Assume Countable Choice and Dependent Choice. Let U be a strongly continuous one-parameter unitary group on H with infinitesimal generator G, and let λ>0. Then Q±(λ)x=0eλtU(±t)xdt is a Bochner integral depending linearly and boundedly on x, with Q±(λ)1/λ and ranQ±(λ)D(G); moreover (λG)Q±(λ)=I,Q±(λ)(λG)=I on D(G), and Q+(λ)+Q(λ)=2λQ+(λ)Q(λ).

Facts & Assumptions

[A1]

Strong measurability means pointwise almost-everywhere norm approximation by measurable simple functions. Such a function is Bochner integrable when its norm is integrable, and ff. The integral is the limit of integrals of simple approximations in integral norm (Strongly measurable Banach-valued function, Banach-valued simple function and integral, Bochner-integrable function, Bochner integrability criterion, Bochner integral norm inequality).

[A2]

Bounded linear maps commute with Bochner integrals and Bochner dominated convergence holds under Countable Choice (Bounded linear maps commute with Bochner integration, Bochner dominated convergence theorem, The Axiom of Countable Choice (ACω)).

[A5]

Lebesgue measurability and measure are invariant under translation (Lebesgue outer measure, Lebesgue measurability and Lebesgue measure are unchanged by translation).

Proof

technique · direct

Given: Countable Choice, Dependent Choice, a strongly continuous unitary group U with generator G, and λ>0.

1.1

For either sign put Fx(t)=eλtU(±t)x on [0,). It is continuous. For integer n1, approximate it on [0,n) by its values at the left endpoints of intervals of length 2n and put zero elsewhere. These are measurable simple functions with finite support measure, converging at every fixed t0 to Fx(t) by continuity. Thus Fx is strongly measurable. Compact Newton--Leibniz with primitive eλt/λ and the bridge in [A4] give 0Neλtdt=(1eλN)/λ. Monotone convergence as integer N gives 0eλtdt=1/λ. Since Fx(t)=eλtx, [A1] defines Q±x and gives Q±xx/λ. Linearity follows first for simple integrals and then by adding their approximating sequences in integral norm. Therefore these are bounded linear operators.

A1A3A4
1.2

For xD(G) write vh=(U(h)xx)/hGx for real h0. Norm preservation and the inner-product expansion give 0=2Revh,x+hvh2. Since a convergent family is bounded near zero, the limit gives ReGx,x=0. Hence Re(λG)x,x=λx2. If (λG)x=0, positivity of λ gives x=0; both shifts are injective. This argument requires no density or closedness theorem for G.

A3given
2.1

Translation of a Bochner integral is legitimate here: for simple integrable functions it follows termwise from [A5]; for their integral-norm limits the scalar change-of-variables identity follows first for nonnegative simple functions, then by monotone convergence, and shows that translation preserves the approximation error. Thus the simple identities pass to the Bochner integral by [A1]. Subdivision and linearity follow in the same way from simple integrals. Also h10hFx(s)dsxsup0shFx(s)x0 as h0. The tail hFx tends to Q±x, since the norm of the omitted integral is at most hx.

A1A4A5step 1.1
3.1

For the plus sign and h>0, commuting U(h) with integration and translating gives U(h)IhQ+x=eλh1hheλsU(s)xds1h0heλsU(s)xdsλQ+xx by step 2.1. To obtain the required two-sided derivative, if yH has right quotient vh=(U(h)yy)/hv, then U(h)yyh=U(h)vhv: its error is bounded by vhv+U(h)vv. Thus Q+xD(G) and (λG)Q+x=x. For V(t)=U(t), substitution s=t in the two-sided derivative definition gives D(GV)=D(G) and GV=G. Applying the proved plus-sign argument to V gives QxD(G) and (λ+G)Qx=x.

A2A3step 1.1step 2.1
4.1

For xD(G) let y=Q±(λG)x. Step 3.1 places y in D(G) and gives (λG)y=(λG)x. Injectivity from step 1.2 implies y=x, proving Q±(λG)=I on D(G). Together with step 3.1 this shows ranQ±=D(G) and both inverse identities with their stated domains.

step 3.1step 1.2
5.1

From (λ+G)Q=I obtain GQ=IλQ and (λG)Q=2λQI. Multiplication on the left by Q+ is legitimate on every vector because ranQD(G). Step 4.1 gives Q=2λQ+QQ+, hence Q++Q=2λQ+Q.

step 3.1step 4.1
6.1

The norm bound, range and inverse claims are steps 1.1, 3.1 and 4.1, and the sum identity is step 5.1. For H={0} the same formulas directly concern its unique full-domain operator; zero vectors give zero integrals. The strict condition λ>0 ensures integrability and injectivity. Countable Choice supplies the compact integration bridge and the Bochner framework; the declared Dependent Choice is not additionally needed by this proof. No half-line fundamental theorem for a merely bounded derivative is invoked.

A1A2A4step 1.1step 3.1step 1.2step 4.1step 5.1

Source notes

Schnaubelt, Lemma 1.18 and Proposition 1.20(a)-(b), printed pp.11-13, supplies the translated-integral route to the resolvent. Here unitarity proves injectivity directly, so the left inverse follows from the right inverse without any half-line scalar fundamental theorem or a prior closedness theorem for the generator. Both signs and the two-sided derivative are checked explicitly.

LemmaStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

The generator of a unitary group is closed and skew-adjoint

Statement

Assume Countable Choice and Dependent Choice. The generator G of a strongly continuous one-parameter unitary group U is densely defined, closed, and skew-adjoint: G=G. Consequently T=iG is self-adjoint with D(T)=D(G).

Facts & Assumptions

[A1]

Both Q±(λ) have range D(G) and satisfy Q±(λ)y=0eλtU(±t)ydt with Q±(λ)1/λ; also (λG)Q±(λ)=I and Q±(λ)(λG)=I on D(G) (Laplace resolvents of a unitary group).

[A2]

U(t)x,U(t)y=x,y and tU(t)x,U(t)y is differentiable at 0 with derivative Gx,y+x,Gy when x,yD(G), by the definition of G and sesquilinearity and continuity of the inner product (Infinitesimal generator of a unitary group, Strongly continuous one-parameter unitary group, Hilbert space, Cauchy–Schwarz: x,yxy, with equality exactly for dependent pairs).

[A3]

A densely defined symmetric operator S with ran(S±i)=H is self-adjoint (Range criterion for self-adjointness, Symmetric, self-adjoint and essentially self-adjoint operators).

[A4]

A bounded everywhere-defined inverse to 1G puts 1 in the resolvent of G and forces its graph closed, with convention RG(1)=(IG)1 (Resolvent and spectrum of an unbounded operator, Densely defined, closed and closable operators, and cores). The exact limit argument is also given below.

[A5]

The Bochner integral is linear by passage from simple integral approximations, and its norm is bounded by the integral of the norm (Bochner-integrable function, Bochner integral norm inequality). Compact Newton--Leibniz, the Countable Choice Riemann/Lebesgue bridge and monotone convergence compute aλeλtdt=eλa for a0, λ>0, by the primitive eλt on [a,N] followed by N (Newton–Leibniz needs only continuity on [a,b], differentiability on (a,b), and a Riemann-integrable extension of the interior derivative, A bounded Riemann integrable function on a closed bounded interval is Lebesgue measurable and has the same integral, Monotone convergence for the integral).

[A6]

The adjoint domain consists of vectors making xSx,y bounded on D(S), with Sx,y=x,Sy in the first-variable-linear convention (Adjoint of a densely defined operator).

Proof

technique · direct

Given: Countable Choice, Dependent Choice, and a strongly continuous unitary group U with generator G.

1.1

D(G) is dense: for yH and λ>0, [A1] gives λQ+(λ)yD(G) and λQ+(λ)yy=0λeλt(U(t)yy)dt. Given ε>0, strong continuity supplies δ>0 such that U(t)yy<ε for 0tδ; the integral norm is then at most ε+2yeλδ. Letting λ and then ε0 proves λQ+(λ)yy. Taking positive integer λ supplies an approximating sequence in D(G) for every y. Thus D(G) is dense.

A1A5given
1.2

G is skew-symmetric: the function tU(t)x,U(t)y is constant with value x,y for x,yD(G), so its derivative at 0 vanishes, that is Gx,y+x,Gy=0; equivalently Gx,y=x,Gy.

A2
1.3

G is closed: with R=Q+(1), [A1] and [A4] already imply closedness. Explicitly, if xnD(G), xnx and Gxny, then xn=R(xnGxn)R(xy) since R is bounded. Uniqueness of limits gives x=R(xy)D(G) and (IG)x=(IG)R(xy)=xy, hence Gx=y. The sequential graph criterion is valid in the norm metric under the assumed Countable Choice.

A1A4
2.1

On D(T)=D(G) set T=iG. For x,y in this domain, step 1.2 gives Tx,y=iGx,y=ix,Gy=x,Ty, so T is symmetric. It is densely defined by step 1.1. It is closed: convergence of xn and Txn implies convergence of Gxn=iTxn, so step 1.3 applies. The correct signed formulas are T+iI=i(IG) and TiI=i(I+G). Both ranges equal H by the two signs of [A1] at λ=1; multiplication by a nonzero scalar preserves surjectivity. Hence [A3] makes T self-adjoint.

A1A3step 1.1step 1.2step 1.3
3.1

The adjoint domain of T=iG equals that of G: multiplication of the scalar functional in [A6] by i preserves boundedness in both directions. For y in this domain, Tx,y=iGx,y=ix,Gy=x,iGy; uniqueness of the representing vector gives T=iG. Step 2.1 gives T=T=iG including domains, hence D(G)=D(G) and G=G.

A6step 2.1
4.1

The conclusions are density, closedness and skew-adjointness from steps 1.1, 1.3 and 3.1, and self-adjointness of iG from step 2.1. The zero space and constant identity group satisfy the same identities directly. The declared Countable Choice and Dependent Choice match the Laplace supplier; Countable Choice also licenses the range/adjoint and integration interfaces. Only strictly positive Laplace parameters are used.

A1A3A5A6step 1.1step 1.3step 2.1step 3.1
TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Stone's theorem: unitary groups and self-adjoint generators

Statement

Assume the Axiom of Choice. The map TU(t)=eitT, computed by the Borel functional calculus of T (Unbounded Borel functional calculus: domains, products, spectral mapping), is a bijection from the set of self-adjoint operators on H onto the set of strongly continuous one-parameter unitary groups on H. Its inverse assigns to U its generator G and the self-adjoint operator T=iG; here D(T)=D(G) is exactly the set of vectors at which tU(t)x is differentiable at 0, and 1t(U(t)xx)iTx for xD(T).

Facts & Assumptions

[A1]

For every self-adjoint T, U(t)=eitT is a strongly continuous one-parameter unitary group with generator G=iT and D(G)=D(T) (A self-adjoint operator generates a strongly continuous unitary group).

[A2]

The generator G of a strongly continuous unitary group is densely defined and skew-adjoint, so S:=iG is self-adjoint with D(S)=D(G) (The generator of a unitary group is closed and skew-adjoint, Infinitesimal generator of a unitary group).

[A3]

If xD(G) then U(t)xD(G), GU(t)x=U(t)Gx, and tU(t)x is differentiable with derivative U(t)Gx (Laplace resolvents of a unitary group, Infinitesimal generator of a unitary group).

[A4]

A skew-symmetric operator G satisfies ReGw,w=0 for wD(G), because Gw,w=w,Gw=Gw,w. The generator in [A2] is skew-adjoint and hence skew-symmetric. The generator of a unitary group is closed and skew-adjoint Hilbert space

Proof

technique · direct

Given: A self-adjoint T, and a strongly continuous unitary group U with generator G.

1.1

Applying [A1] to T produces a strongly continuous unitary group with generator iT, so the map TeitT is well defined, and its derivative at 0 exists exactly on D(T) where it equals iTx.

A1
1.2

Applying [A2] to U produces the self-adjoint operator S=iG with D(S)=D(G); applying [A1] to S gives the strongly continuous unitary group V(t)=eitS, whose generator is iS=G.

A1A2
2.1

Uniqueness for a fixed generator: if U,V are strongly continuous unitary groups with the same generator G and xD(G), then w(t):=U(t)xV(t)x is differentiable with w(t)=Gw(t) by [A3], so ddtw(t)2=2Rew(t),Gw(t)=0 by [A4], and w(0)=0 gives w0 on D(G); since D(G) is dense by [A2] and U(t),V(t) are isometries, U(t)=V(t) for every t.

A2A3A4step 1.2
3.1

Hence U=V in step 1.2, that is U(t)=eitT for the self-adjoint T=S=iG; combined with step 1.1 this makes TeitT a bijection with the stated inverse.

step 1.1step 1.2step 2.1
4.1

The derivative characterisation is the one from [A1] applied to the self-adjoint T=iG: the limit exists exactly on D(T)=D(G) and equals iTx.

A1step 3.1
DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Deficiency subspaces and deficiency indices

Definition

Assume the Axiom of Choice. Let T be a densely defined closed symmetric operator on a complex Hilbert space H (Symmetric, self-adjoint and essentially self-adjoint operators, Densely defined, closed and closable operators, and cores). Its deficiency subspaces are K+:=ker(Ti),K:=ker(T+i), and its deficiency indices are the Hilbert dimensions d±(T):=dimK±, that is, the cardinalities of orthonormal bases of K± (Orthonormal families, complete orthonormal systems and Hilbert bases).

The subspaces are the orthocomplements of the ranges. By The adjoint is well defined, closed, and reverses inclusions ran(T±i)=ker(Ti), so K+=ran(T+i) and K=ran(Ti). Each K± is a closed linear subspace by Orthogonal complements are closed, hence is itself a Hilbert space. For xD(T), symmetry and conjugate symmetry make r:=Tx,x=x,Tx real. With the inner product linear in its first variable Real and complex inner-product spaces and their induced length, (T±i)x2=Tx2+x2ir±ir=Tx2+x2. Thus T±i is injective. If (T±i)xny, applying the identity to xnxm makes (xn) Cauchy. Completeness gives xnx, and Txn=(T±i)xnixnyix. Closedness of T now gives xD(T) and (T±i)x=y, proving both ranges closed. The closed-subspace decomposition theorem Orthogonal decomposition by a closed subspace therefore gives H=ran(T+i)K+=ran(Ti)K. Also K+K={0}, since membership forces Tu=iu=iu. The two deficiency spaces need not be orthogonal to each other in H.

Dimension convention and well-definedness. An orthonormal basis exists in each K± by Existence of a maximal orthonormal family, and maximality as completeness, using full AC. Its cardinality is independent of the basis, as follows. Let (ei)iI and (fj)jJ be two orthonormal bases of the same Hilbert space. By The Bessel inequality for an arbitrary orthonormal family, for each i and integer n1 at most n indices satisfy ei,fj21/n. Thus the support in J of each row is countable. Each column has a nonzero entry: otherwise fj is orthogonal to the dense span of all ei, hence to itself, contradicting norm one. Here orthogonality passes to the closure by Orthogonal complements are closed. If I is infinite, full AC lets us enumerate the row supports and assign each j to one row containing it; this gives an injection JI×N. Cardinal absorption Absorption: for cardinals κ,λ with κ infinite and λκ, κλ=κ, and κλ=κ when λ0 and cardinal comparison Commutativity, associativity, distributivity and monotonicity of and , the unit laws, the two exponent laws, and κλ if and only if κ injects into λ give JI. If I has finite size n, the residual fjiIfj,eiei is orthogonal to the dense span of the ei and so vanishes. Taking its norm gives iIfj,ei2=1. For every finite FJ, Bessel in the other direction yields F=jFiIfj,ei2=iIjFei,fj2n. Hence Jn. Exchanging the two bases proves equality of cardinalities in all cases. In the finite case each basis also spans algebraically by the same residual argument, so this is the ordinary linear dimension. For the zero space the basis is empty and the dimension is zero. AC is used for basis existence and the simultaneous choices in the infinite comparison.

Cayley sign and domain convention. Define CT:ran(T+i)ran(Ti),CT((T+i)x)=(Ti)x. Injectivity of T+i makes this well defined; the norm identity makes it an isometry onto the stated range. On its domain, (ICT)(T+i)x=2ix, whence ran(ICT)=D(T), since D(T) is a complex linear subspace. If a unitary U:HH extends CT, it maps the orthogonal complement K+ of the initial range onto the orthogonal complement K of the final range, by preservation of inner products and surjectivity. Conversely, any unitary W:K+K gives the unitary extension U=CTW on the two displayed orthogonal decompositions. This describes the free part of a unitary extension and fixes the signs; it does not assert that such a W always exists.

TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Von Neumann parameterization of self-adjoint extensions

Statement

Assume the Axiom of Choice. Let T be a densely defined closed symmetric operator on a complex Hilbert space H, with first-variable-linear inner product, with deficiency subspaces K± (Deficiency subspaces and deficiency indices), and let V:K+K be a unitary operator. Then D(TV)=D(T){u+Vu:uK+},TV(x+u+Vu)=Tx+iuiVu defines a self-adjoint extension TV of T; the sum is direct and D(TV) is dense. The map VTV is a bijection from the set of unitary operators K+K onto the set of self-adjoint extensions of T.

Facts & Assumptions

[A1]

For the given closed densely defined symmetric T, K+=ker(Ti) and K=ker(T+i) are closed, and H=ran(T+i)K+=ran(Ti)K orthogonally. The linear map CT((T+i)x)=(Ti)x is an isometric isomorphism between these ranges, and (ICT)(T+i)x=2ix. Full AC licenses this deficiency-space interface, including its Hilbert-dimension convention. Deficiency subspaces and deficiency indices

[A2]

Under Countable Choice the Cayley correspondence sends a unitary U with ker(IU)={0} to the self-adjoint operator S(IU)y=i(I+U)y, with domain ran(IU), and recovers CS=U. For self-adjoint S its Cayley transform satisfies CS(S+i)x=(Si)x on D(S). Cayley correspondence between self-adjoint operators and unitaries Cayley transform of a self-adjoint operator

[A3]

D(T) is norm dense in H. Orthogonal decompositions have zero intersection and their squared norms add. A vector orthogonal to a dense subspace is zero: continuity of pairings follows from Cauchy-Schwarz. Densely defined, closed and closable operators, and cores Orthogonality and the orthogonal complement Cauchy–Schwarz: x,yxy, with equality exactly for dependent pairs

[A4]

Full AC is assumed to use [A1]. It implies the Countable Choice required by [A2] directly: AC supplies a choice function for the range family of any given sequence of nonempty sets, and composing that choice function with the sequence gives the required indexed choices. No additional family of choices is made in the construction from the supplied unitary V. The Axiom of Choice The Axiom of Countable Choice (ACω)

Proof

technique · direct

Given: T as in the statement and a unitary V:K+K. Put M+=ran(T+i) and M=ran(Ti).

1.1

Define U(m+u)=CTmVu for mM+ and uK+. The orthogonal decomposition in [A1] makes this a uniquely defined linear map on H. Its two output terms lie in the orthogonal subspaces M and K, so U(m+u)2=CTm2+Vu2=m2+u2=m+u2. Since both component maps are onto their corresponding summands, U is onto H. Thus U is unitary and extends C_T; the minus sign on K_+ is necessary for the displayed plus sign in u+Vu.

A1A3A4
2.1

For xD(T), (IU)(T+i)x=2ix, hence ran(IU) contains D(T) and is dense. If Uz=z, then for every y in H, z,(IU)y=z,yUz,Uy=0, since a unitary preserves inner products. Consequently z is orthogonal to the dense D(T), so z=0 by [A3]. This proves ker(IU)={0}.

A1A3step 1.1
3.1

Apply [A2], licensed by [A4], to get the self-adjoint operator S(IU)y=i(I+U)y on ran(IU), with CS=U. Because (IU)((T+i)x+u)=2ix+u+Vu, that domain equals D(T)+{u+Vu:uK+}; scalar multiplication by 2i maps D(T) onto itself. To prove the sum direct, suppose x=u+VuD(T). Then (IU)(T+i)x=2ix=(IU)(2iu), and injectivity from step 2.1 gives (T+i)x=2iu. The two sides lie in M_+ and K_+, whose intersection is zero. Hence u=0 and x=0. Also u+Vu=(I-U)u shows the parametrization of the second summand is injective. Thus every vector has a unique representation x+u+Vu, and the domain contains the dense D(T).

A1A2A3A4step 1.1step 2.1
4.1

On D(T), S(2ix)=i(I+U)(T+i)x=2iTx, so Sx=Tx. On the second summand, S(u+Vu)=S(IU)u=i(I+U)u=iuiVu. Linearity yields S(x+u+Vu)=Tx+iuiVu. Thus S is exactly the well-defined operator T_V in the statement and is a self-adjoint extension of T.

A1step 1.1step 3.1
5.1

Let R be any self-adjoint extension of T and put W=C_R. For x in D(T), (R+i)x=(T+i)x, so W(T+i)x=(Ri)x=(Ti)x. Thus W agrees with C_T on M_+ and maps M_+ onto M_-. For u in K_+ and m in M_+, Wu,Wm=u,m=0, so Wu belongs to K_-. Conversely, for v in K_-, take the unique y with Wy=v. For every m in M_+, y,m=v,Wm=0, hence y belongs to K_+. This proves W(K_+)=K_-, without treating W* as the Cayley transform of R. Consequently V=-W restricted to K_+ is unitary from K_+ onto K_-, and the construction of step 1.1 returns U=W. The inverse correspondence [A2] then gives T_V=R.

A1A2A3step 1.1step 4.1
6.1

If T_V=T_{V'}, their Cayley transforms agree by [A2]. Step 3.1 identifies these transforms with the constructed U and U', whose restrictions to K_+ are -V and -V'. Hence V=V'. Along with steps 4.1 and 5.1, this proves the bijection. This includes empty parameter sets: if no such unitary exists, step 5.1 rules out every self-adjoint extension. If K_+=K_-={0}, the unique unitary of the zero spaces gives D(T_V)=D(T) and T_V=T, so T is already self-adjoint. If H={0}, every displayed map is its unique zero-space map and the same conclusion holds directly. No finite-dimensional or separability assumption is used. AC is used only through [A4].

A1A2A4step 1.1step 3.1step 4.1step 5.1
CorollaryStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Existence of self-adjoint extensions is equality of deficiency indices

Statement

Assume the Axiom of Choice. Let T be a densely defined closed symmetric operator on H, with deficiency indices d±(T) (Deficiency subspaces and deficiency indices). Then T has a self-adjoint extension if and only if d+(T)=d(T). Moreover T is self-adjoint if and only if d+(T)=d(T)=0, and a densely defined symmetric (not necessarily closed) operator S is essentially self-adjoint if and only if d±(S)=0, equivalently ker(Si)={0}; if d+(S)=d(S), then S has self-adjoint extensions.

Facts & Assumptions

[A1]

Unitary operators K+K correspond bijectively to self-adjoint extensions of T; every such unitary is onto by definition, and the Cayley transform US of a self-adjoint extension restricts to a unitary K+K (Von Neumann parameterization of self-adjoint extensions).

[A2]

Every Hilbert space has a complete orthonormal family, and two Hilbert spaces are unitarily isomorphic exactly when their orthonormal bases have the same cardinality; a unitary K+K exists exactly when dimK+=dimK (Existence of a maximal orthonormal family, and maximality as completeness, A Hilbert space with a given orthonormal basis is 2 of the index set, Orthonormal families, complete orthonormal systems and Hilbert bases).

[A3]

A closed symmetric operator is self-adjoint if and only if ker(Ti)={0}, equivalently ran(T±i)=H (Range criterion for self-adjointness).

[A4]

For a densely defined symmetric S one has S=S and S=(S); S is essentially self-adjoint exactly when S is self-adjoint (Closability is equivalent to density of the adjoint domain, Symmetric, self-adjoint and essentially self-adjoint operators).

Proof

technique · direct

Given: A densely defined closed symmetric operator T, and a densely defined symmetric operator S.

1.1

If d+(T)=d(T), then by [A2] there is a unitary V:K+K (equal Hilbert dimensions), and [A1] produces a self-adjoint extension TV of T. Conversely, if T has a self-adjoint extension S, then by [A1] its Cayley transform restricts to a unitary K+K, so dimK+=dimK by [A2].

A1A2
1.2

T is self-adjoint if and only if d+=d=0: if both deficiency subspaces are zero then ker(Ti)={0} and [A3] applies; conversely a self-adjoint T has T=T and Ti injective by the estimate (Ti)xx, so both kernels vanish.

A3
2.1

For S symmetric, S is closed and symmetric by [A4], and ker(Si)=ker((S)i) because S=(S); applying steps 1.1-1.2 to S gives: S is essentially self-adjoint, meaning S self-adjoint, if and only if d±(S)=0; if d+(S)=d(S) then S has a self-adjoint extension, which is also an extension of S.

A4step 1.1step 1.2
3.1

All the stated equivalences are steps 1.1, 1.2 and 2.1. ∎

DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Pure point, absolutely continuous and singular continuous spectral subspaces

Definition

Assume the Axiom of Choice. Let T be a self-adjoint operator on H. If H{0}, let E be its spectral projection valued measure on R from Spectral theorem for unbounded self-adjoint operators (PVM form); if H={0}, let E(B)=0 for every Borel B, the unique PVM on the zero space. In either case put Ex(B)=E(B)x,x (Integral of a measurable function against a projection-valued measure). Call a finite Borel measure on R purely atomic (equivalently, discrete) when it is concentrated on a countable subset of R. Then Hpp={xH: Ex is purely atomic}, Hac={xH: Exλ (Lebesgue measure)}, Hsc={xH: Ex is atomless and singular with respect to λ}, where atoms are as in An atom of a measure on R, absolute continuity is that of Absolute continuity of a signed or complex measure with respect to a positive measure and singularity that of Mutual singularity for signed or complex measures.

Well-definedness. Every finite Borel measure on R has a unique decomposition into a discrete (hence, by the convention above, purely atomic), an absolutely continuous and an atomless singular part (Every finite Borel measure on R has a unique absolutely continuous, discrete, and singular-continuous decomposition), and the three classes of nonzero measures are mutually exclusive; hence each x either satisfies one of the three defining conditions or none, and the zero vector lies in all three subspaces. Whether the three subspaces do cover H and are closed is not part of this definition and is the content of the canonical spectral type decomposition theorem below.

Conventions. For each type one writes σtype(T)=σ(THtype) for the restriction of T to the closed invariant subspace Htype constructed in the canonical spectral type decomposition theorem below; these restrictions are self-adjoint there. The point spectrum is not σpp: σpp(T) is the closure of the set of eigenvalues of T, and the three sets σpp,σac,σsc may overlap, so they do not partition σ(T).

TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Canonical decomposition into pure point, absolutely continuous and singular continuous parts

Statement

Assume the Axiom of Choice. Let T be a self-adjoint operator on a complex Hilbert space H with spectral projection valued measure E on R and let Hpp, Hac, Hsc be the subspaces of Pure point, absolutely continuous and singular continuous spectral subspaces. Then Hpp,Hac,Hsc are closed, mutually orthogonal, T-reducing subspaces with H=HppHacHsc, canonically determined by T; the restrictions of T to them are self-adjoint and their spectral measures are respectively purely atomic, absolutely continuous with respect to Lebesgue measure, and atomless and singular. If H is separable and μ is a maximal scalar spectral measure with disjoint Borel supports Bpp,Bac,Bsc of its discrete, absolutely continuous and singular continuous parts, then Htype=ranE(Btype).

Here a support means a Borel carrier (zero mass off the set), not necessarily topological support. A maximal scalar spectral measure is a finite positive Borel measure mu with Exμ for every x (in particular a scalar spectral measure with this domination property qualifies). The last assertion is conditional on the supplied mu. On the zero Hilbert space use the unique PVM and full-domain zero operator directly.

Facts & Assumptions

[A1]

The three types are defined by the scalar measures Ex(B)=E(B)x2: discrete means concentrated on a countable set, absolutely continuous means vanishing on Lebesgue-null Borel sets, and singular continuous means atomless and carried by a Lebesgue-null Borel set. Pure point, absolutely continuous and singular continuous spectral subspaces Absolute continuity of a signed or complex measure with respect to a positive measure Mutual singularity for signed or complex measures An atom of a measure on R

[A2]

Each finite positive Borel measure on the line has a unique decomposition into discrete, absolutely continuous and atomless singular measures. Its atoms form a countable set. Lebesgue measure gives zero mass to a singleton, hence to a countable set by countable subadditivity. Every finite Borel measure on R has a unique absolutely continuous, discrete, and singular-continuous decomposition Every finite Borel measure on R splits as an atomic part plus an atomless part A box in Rn with parameters aibi is Lebesgue measurable of measure i<n(biai), whichever of its faces are included

[A3]

E is a regular PVM, its projection values commute and satisfy E(B)E(C)=E(B intersect C), and each is contractive and self-adjoint. Orthogonality uses the first-linear inner product. Projection valued measure Orthogonality and the orthogonal complement Spectral theorem for unbounded self-adjoint operators (PVM form)

[A4]

D(T)={x:λ2dEx<} and Tx=limnΦE(λ1[n,n])x. Bounded integrals are operator-norm limits of integrals of uniform simple approximations. Conversely the integral of the real coordinate against a regular PVM is self-adjoint on this domain. Integral of a measurable function against a projection-valued measure Bounded borel pvm integral Spectral theorem for unbounded self-adjoint operators (PVM form)

[A5]

AC supplies the measure-decomposition and spectral interfaces and directly supplies the countable choices of null carriers and all dependent or countable witness choices used below. The Axiom of Choice

Proof

technique · direct

Given: AC, a complex Hilbert space H, and self-adjoint T with its regular spectral PVM E; use the direct zero-space convention when needed.

1.1

For any finite measure nu, take its decomposition [A2]. Let P be the countable set of atoms, carrying the discrete part, and choose a Lebesgue-null Borel carrier N for the singular continuous part. The three disjoint Borel sets Spp=P, Ssc=NP, Sac=R(PN) partition the line and carry their corresponding parts: both atomless parts vanish on P, the absolutely continuous part vanishes on P union N, and the discrete part vanishes off P. Thus νSr=νr for each type r. More generally, for any disjoint carriers S_r of the three components, their union carries nu and the same restriction identity holds. If eta<<nu is finite positive, its restrictions to these carriers are of the respective types: they are dominated in the sense of null sets by nu_r, so inherit a countable carrier, Lebesgue absolute continuity, or a null carrier and zero singleton masses. They sum to eta, so uniqueness in [A2] implies eta has type r exactly when eta is carried by S_r. The zero measure has all three types, consistently with this assertion.

A1A2A5
1.2

For every Borel B, [A3] gives Ex+y(B)2Ex(B)+2Ey(B) and Ecx(B)=c2Ex(B). Consequently each type set is linear: two countable carriers have countable union, two null carriers have null union, and zero masses on null sets or singletons pass through this inequality. If x_n of one type converge to x, the contraction inequality E(B)(xxn)xxn shows E_x(B)=0 whenever all E_{x_n}(B)=0. For the discrete case choose countable carriers P_n and use their countable union P; then E_x(P^c)=0. For the singular case choose null Borel carriers N_n and use their null union N. For the atomless condition apply the same argument to each singleton; for absolute continuity apply it to each fixed Lebesgue-null Borel set. Hence all three subspaces are closed. The countable carrier choices and countable unions use the declared AC [A5].

A1A2A3A5
2.1

Measures of different types are mutually singular: a discrete carrier is countable and both other types give it zero mass; a singular-continuous null carrier has zero absolutely continuous mass. Thus for x,y of different types there is a Borel S carrying E_x with E_y(S)=0. Since E(Sc)x2=Ex(Sc)=0, one has E(S)x=x, while E(S)y=0. Self-adjointness of E(S) yields x,y=E(S)x,y=x,E(S)y=0.

A1A2A3step 1.2
3.1

For arbitrary x, apply step 1.1 to E_x and set x_r=E(S_r)x. The PVM identities give Exr(B)=E(B)E(Sr)x2=Ex(BSr), so x_r belongs to the indicated type. The partition gives x=xpp+xac+xsc. By step 2.1 this decomposition is orthogonal and unique. The component maps Q_r are linear by uniqueness, contractive by the Pythagorean identity for this finite orthogonal sum, and self-adjoint because Qrx,y=xr,yr=x,Qry. They are orthogonal projections onto the closed type subspaces. As these subspaces were defined from E_x, they and the projections are canonical, independent of the carriers chosen for individual vectors.

A1A3step 1.1step 1.2step 2.1
4.1

For each Borel B the equality EE(B)x=ExB shows that E(B) preserves each type. Applying it to the unique decomposition in step 3.1 yields Q_r E(B)=E(B)Q_r. Therefore EQrx(B)=QrE(B)x2Ex(B), so x in D(T) implies Q_rx in D(T) by [A4]. Commutation with every E(B) gives commutation with every simple integral and then every bounded integral by [A4]. Passing to the coordinate truncation limit gives TQ_rx=Q_rTx for x in D(T). Thus each type subspace reduces T, with its domain carried along.

A3A4step 3.1
4.2

In the separable clause let the supplied finite maximal measure mu have disjoint carriers B_r of its three components. For every x, E_x<<mu, so step 1.1 says x has type r exactly when E_x(B_r^c)=0. The latter is equivalent to E(B_r)x=x, since E(B_r^c)=I-E(B_r) and Ex(Brc)=E(Brc)x2. Hence H_r=ran E(B_r). This proves the assertion for every supplied maximal mu, not just for a specially constructed one, and needs no circle-to-line transport.

A1A3step 1.1step 3.1
5.1

On a nonzero type subspace K, E_K(B)=E(B)|_K is a regular PVM: its projection and strong countable-additivity properties restrict from E, and its scalar measures are the same regular E_x for x in K. The coordinate integral against E_K has domain K intersect D(T); bounded simple integrals and their limits agree with the restrictions of those for E, so its value is Tx. The converse spectral theorem in [A4] makes this restriction self-adjoint. On K={0}, self-adjointness is direct since its unique densely defined operator equals its adjoint. The scalar measures of each restriction have exactly the specified type by [A1].

A1A3A4step 3.1step 4.1
6.1

If H={0}, every scalar measure is zero, the three subspaces are {0}, and all conclusions including the carrier formula hold directly. Vanishing components on a nonzero H also give zero subspaces by the same arguments, and no measure is divided by its mass. Nonseparability causes no difficulty in the preceding arguments, since only a single scalar measure or a sequence of vectors is used at a time. Full AC is used exactly as in [A5], including countable carrier choices for closedness.

A5step 1.1step 1.2step 3.1step 5.1step 4.2

Source notes

Teschl, Section 3.3, Lemma 3.18, printed pp.118–119, gives the canonical type spaces and their spectral projections from maximal-measure carriers. The direct scalar-measure argument here proves the decomposition without a separability assumption; the maximal-measure carrier formula is asserted conditionally as in the statement. No change-of-variables or Cayley transport is needed.

DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Relative boundedness with respect to an operator

Definition

Let A be a linear operator on H with domain D(A). An operator B is A-bounded, or relatively bounded with respect to A, when D(A)D(B) and there are finite constants a,b0 with BxaAx+bxfor all xD(A). The infimum of the admissible constants a is the A-bound of B, and one says the A-bound is below one when some a<1 is admissible.

Equivalent form. B is A-bounded exactly when B is bounded on D(A) for the graph norm xA=(x2+Ax2)1/2 of Unbounded linear operators: domain, graph and extension: each estimate BxaAx+bx gives Bxmax(a,b)2xA, and conversely a graph-norm bound BxCxA gives the estimate with a=b=C. The constant b is not intrinsic, and no closedness, density or resolvent hypothesis is needed for the definition. If A is closed with nonempty resolvent set, then A-bounded operators are exactly those with D(A)D(B) for which BRA(z) is bounded for some, equivalently every, zρ(A) (Resolvent and spectrum of an unbounded operator). Indeed, an A-bound makes BRA(z) bounded because ARA(z)=zRA(z)I; conversely, if BRA(z) is bounded, then Bx=BRA(z)(zA)x gives an A-bound. Thus boundedness for one resolvent implies relative boundedness and hence boundedness for every resolvent. This is used in the Kato-Rellich theorem below and recorded here as an interface.

LemmaStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Second resolvent identity for a closed perturbation

Statement

Assume Dependent Choice. Let A and C be closed operators with D(C)=D(A), put B:=CA, and assume B is bounded for the graph norm of A (Relative boundedness with respect to an operator). Then for every zρ(A)ρ(C) RC(z)RA(z)=RA(z)BRC(z)=RC(z)BRA(z), and every product here is defined on all of H and is bounded.

Facts & Assumptions

[A1]

For zρ(A) the operator RA(z) maps H bijectively onto D(A) and ARA(z)=zRA(z)I on H, since A=z(zA) (Resolvent and spectrum of an unbounded operator).

[A2]

B is bounded for the graph norm of A: there are a,b0 with BxaAx+bx for xD(A); each BRA(z) is therefore everywhere defined and bounded, since BRA(z)yaARA(z)y+bRA(z)y and both terms are bounded in y (Relative boundedness with respect to an operator, [A1]).

[A3]

(zC)RC(z)=I and (zA)RA(z)=I on H, so RA(z)(Cz)RC(z)=RA(z) and RC(z)(Az)RA(z)=RC(z), because RC(z) has range D(C)=D(A) (Resolvent and spectrum of an unbounded operator).

[A4]

The graph norms of two closed operators with the same domain are equivalent: both domains are Banach, and the identity map from the C-graph norm to the A-graph norm has closed graph, hence is bounded by the closed graph theorem (Closed graph theorem, The axiom of dependent choice: a relation in which every element is related to something admits an N-indexed chain).

Proof

technique · direct

Given: Closed A,C with D(A)=D(C), B=CA graph-norm bounded for A, and zρ(A)ρ(C).

1.1

The operator BRA(z) is bounded by [A1] and [A2]. By [A4] the A-graph norm is bounded by a constant times the C-graph norm, so the same relative bound makes B bounded for the C-graph norm; applying [A1] with C in place of A shows that BRC(z) is bounded as well.

A1A2A4
2.1

RC(z)+RA(z)BRC(z)=RA(z), that is RC(z)RA(z)=RA(z)BRC(z): by [A3], RA(z)(Cz)RC(z)=RA(z), and expanding C=A+B gives RA(z)(Cz)RC(z)=RA(z)(Az)RC(z)+RA(z)BRC(z)=RC(z)+RA(z)BRC(z), because RA(z)(Az) is minus the identity on D(A)=D(C) and RC(z) takes values in D(C).

A2A3step 1.1
2.2

By the same computation with the roles of A and C interchanged (so that the perturbation is B), RA(z)RC(z)BRA(z)=RC(z), that is RC(z)RA(z)=RC(z)BRA(z).

A2A3step 1.1
3.1

Rearranging steps 2.1 and 2.2 gives RC(z)RA(z)=RA(z)BRC(z) and RC(z)RA(z)=RC(z)BRA(z), which is the stated identity; all products are bounded by step 1.1. ∎

TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Kato-Rellich theorem

Statement

Assume the Axiom of Choice. Work on a complex Hilbert space with inner product linear in the first variable. Let A be self-adjoint and let B be symmetric with D(A)D(B) and A-bound less than one (Relative boundedness with respect to an operator). Then A+B with domain D(A) is self-adjoint. If A is merely essentially self-adjoint and B is symmetric with D(A)D(B) and A-bound less than one, then A+B on D(A) is essentially self-adjoint and its closure is the self-adjoint operator obtained by applying the first part to A and the graph-norm extension of BD(A) to D(A). If Aγ and (a,b) is an admissible pair for B with a<1, then A+Bγmax{aγ+b, b/(1a)}.

Facts & Assumptions

[A1]

Relative boundedness supplies finite a,b>=0 with a<1 and BxaAx+bx on D(A). The graph norm is (x2+Ax2)1/2; graph closure defines the operator closure. Relative boundedness with respect to an operator Unbounded linear operators: domain, graph and extension Densely defined, closed and closable operators, and cores

[A2]

Symmetry is the identity Tx,y=x,Ty on the domain; self-adjoint operators are closed and densely defined. A densely defined symmetric S is self-adjoint if both ranges of S plus and minus i mu equal H for some mu>0. Nonreal points are resolvent points of a self-adjoint operator. The convention is R_T(z)=(z-T)^{-1}. Symmetric, self-adjoint and essentially self-adjoint operators Range criterion for self-adjointness Resolvent of a self-adjoint operator: nonreal resolvents and the estimate Resolvent and spectrum of an unbounded operator

[A3]

If a bounded operator C has norm less than one, I-C has a bounded inverse given by the Neumann series. Neumann series

[A4]

On a nonzero H, a self-adjoint T has a regular PVM E, domain {x:v2dEx<}, and Tx2=v2dEx, Tx,x=vdEx. Bounded integrals satisfy the norm bound and quadratic identity. Projections multiply by intersection; strong countable additivity holds. The Borel calculus has the product rule with domain D(g(T)) intersect D((fg)(T)), and the spectrum of T is the set of v for which every neighborhood has nonzero projection (apply its essential-range assertion to f(v)=v). Spectral theorem for unbounded self-adjoint operators (PVM form) The unbounded PVM integral is densely defined, closed and normal Bounded borel pvm integral Projection valued measure Unbounded Borel functional calculus: domains, products, spectral mapping

[A5]

AC supplies the spectral theorem's choices and directly supplies the countable witness choices used by the range/adjoint interfaces and by sequences approximating a fixed point in a graph closure. The Axiom of Choice

Proof

technique · direct

Given: the operators and admissible pair (a,b) in the statement, with 0<=a<1 and b>=0.

1.1

If H={0}, all domains and graphs are zero and all conclusions hold directly. Otherwise use [A4]. For z=plus or minus i mu, mu>0, put r_z(v)=(z-v)^{-1}. The bounds |r_z(v)|<=1/mu and |v r_z(v)|<=1 show that r_z(A) maps H into D(A), that rz(A)1/μ, and that Arz(A)1. The product rule gives (z-A)r_z(A)=I on H and r_z(A)(z-A)=I on D(A), so this is R_A(z). Thus BRA(z)a+b/μ<1 when, for example, μ=1+2b/(1a).

A1A2A4given
1.2

For any nonzero-space self-adjoint T and real c, the spectral-measure equivalence Tc if and only if E((,c))=0 follows directly. If T>=c and Jn=[(n+1),c1/(n+1)] has a nonzero projection, a nonzero x in its range belongs to D(T), has E_x carried by J_n, and satisfies Tx,x(c1/(n+1))x2, a contradiction. These increasing sets exhaust (-infinity,c), so countable additivity gives zero projection. Conversely zero projection below c gives Tx,x=vdExcx2 for every x in D(T). Also if every real t<c is a resolvent point, the essential-range characterization in [A4] supplies a zero-projection open neighborhood of each t. The rational intervals contained in such neighborhoods form a countable cover of (-infinity,c); their union has zero projection by countable subadditivity of each E_x. Therefore again T>=c. This does not require a finite spectral infimum or any resolvent-distance formula.

A4
2.1

Put S=A+B on exactly D(A). For z=plus or minus i mu the identity zS=(IBRA(z))(zA) holds on D(A), since R_A(z)(z-A)x=x there. The first factor is boundedly invertible by [A3], and z-A is bijective D(A) to H. Hence both nonreal shifts of S are onto. S is symmetric by summing the two symmetry identities and densely defined because D(A) is dense. The range criterion makes S self-adjoint. No second-resolvent identity with a previously closed S is assumed.

A1A2A3step 1.1
3.1

Suppose A>=gamma and let lambda+gamma=d>0. By step 1.2 E_A is carried by [gamma,infinity). Define r(v)=(-lambda-v)^{-1} on that half-line and zero outside. For v>=gamma, r(v)1/d and vr(v)max(1,γ/d): for v>=0 the ratio v/(v+lambda) is monotone with its maximum at an endpoint or its limiting value 1; for gamma<=v<0, (-v)/(v+lambda) decreases with v since lambda>0 in that case. The spectral product rule shows r(A)=R_A(-lambda), exactly as in step 1.1, and hence BRA(λ)amax(1,γ/d)+b/d. Put C=max{aγ+b,b/(1a)}. For every d>C this last bound is strictly below one: if d>=|gamma| it equals a+b/d<1 (also when b=0); if d<|gamma| it equals (a|gamma|+b)/d<1. The factorization of step 2.1 therefore proves that every real number t<gamma-C is in rho(S), with bounded inverse RA(t)(IBRA(t))1. Apply step 1.2 to the already self-adjoint S to obtain S>=gamma-C, the stated bound including its endpoint.

A1A3A4step 1.1step 2.1step 1.2
4.1

If A is essentially self-adjoint, put T=closure(A). For each x in D(T) choose x_n in D(A) with x_n to x and Ax_n to Tx, using [A5]. The inequality in [A1] applied to x_n-x_m makes Bx_n Cauchy. Define Btilde x as its limit. Two such approximations give the same limit by the same inequality applied to their difference. Approximating x and y and their linear combinations proves linearity, B~xaTx+bx, and symmetry by passing to the limit in Bxn,yn=xn,Byn. It extends B restricted to D(A); no extension of B's possibly larger domain is claimed. Step 2.1 makes T+Btilde self-adjoint. The inclusion A+B subset T+Btilde and closedness give closure(A+B) subset T+Btilde. Conversely the same approximating sequences satisfy (A+B)x_n to (T+Btilde)x, giving the reverse graph inclusion. This proves essential self-adjointness and the exact closure formula. If A>=gamma in this case, taking limits of its quadratic inequality gives T>=gamma; step 3.1 then gives the bound for the closure and its restriction A+B.

A1A2A5step 2.1step 3.1
5.1

The choice use is exactly [A5]. The cases B=0 or a=b=0 are admitted by the same estimates and return the original lower bound. Dimension one requires no change. The strict hypothesis a<1 is used in the positive choice of mu and in b/(1-a); no conclusion at a=1 is asserted. The endpoint gamma-C is included by the zero-projection argument, without asserting that the spectrum is nonempty on the zero space.

A1A5step 1.1step 3.1step 4.1

Source notes

Teschl, Section 6.1, Lemma 6.3 and Theorem 6.4, printed pp.158–159 (PDF pp.169–170), supply the imaginary and real resolvent perturbation method. Signs here are computed for the library convention (z-A)^{-1}. The numerical bound is derived above directly from equation (6.3); the proof does not rely on a spectral-distance claim or endpoint continuity of a concave function.

DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Discrete and essential spectrum of a self-adjoint operator

Definition

Assume the Axiom of Choice. Let A be a self-adjoint operator on a complex Hilbert space H with spectral projection valued measure E on R (Spectral theorem for unbounded self-adjoint operators (PVM form)). The discrete spectrum σd(A) is the set of eigenvalues of A that are isolated points of σ(A) and whose eigenspace is finite dimensional; the essential spectrum is σess(A):=σ(A)σd(A), with σ(A) as in Resolvent and spectrum of an unbounded operator.

For H={0} use its unique PVM directly: the spectrum and both parts are empty and every projection has rank zero. Below suppose H{0}, as required by the cited spectral theorem.

Spectral-projection description, with proofs. Fix λR and write Pε:=E((λε,λ+ε)) for ε>0.

Here rank means the algebraic dimension of the range when finite; rank = means the range is not finite dimensional. The calculus Unbounded Borel functional calculus: domains, products, spectral mapping gives the support facts: E(Rσ(A))=0, and every open interval about a spectral point has nonzero projection. Projections on disjoint sets have orthogonal ranges, and E(B)E(C)=E(BC) Projection valued measure.

  1. E({λ}) is the projection onto ker(Aλ). If x=E({λ})x, its scalar measure Ex is supported on {λ} by the projection identity. Thus μ2dEx=λ2x2<, so xD(A), and (Aλ)x2=μλ2dEx=0. The domain and norm identities are supplied by the unbounded calculus and The unbounded PVM integral is densely defined, closed and normal. Hence Ax=λx. Conversely, for an eigenvector (or the zero vector) the same norm identity gives zero integral. On {μλ1/n} this bounds the measure by n2 times that zero integral; taking the countable union shows Ex(R{λ})=0. Since (IE({λ}))x2 equals this scalar measure, E({λ})x=x.
  2. A finite-rank interval contains only finitely many spectral points. If Pε has rank r< and its interval contained r+1 distinct spectral points, choose disjoint small open intervals about those finitely many points, all contained in the given interval. Each has a nonzero projection by support. Choose one unit vector in each range. They are orthonormal vectors in ranPε, hence linearly independent (take inner products with each vector), contradicting its dimension r. Thus there are at most r spectral points there. In particular any λσ(A) with such a finite-rank Pε is isolated: take a smaller interval around λ excluding the other finitely many points. For that interval the projection is E({λ}) by support, is nonzero by support, and has finite rank since its range lies in ranPε. By item 1, λ is an eigenvalue of finite multiplicity.
  3. The two rank characterizations. If λσd(A), an isolating interval has projection E({λ}), of finite rank by item 1. Conversely, if λ is an eigenvalue and some Pε has finite rank, item 2 proves it discrete. Hence λσd(A)λ is an eigenvalue and some Pε has finite rank. If λσess(A), item 2 excludes every finite-rank interval. Conversely, if every Pε has infinite rank, each is nonzero, so support puts λ in σ(A), and the just-proved discrete characterization excludes it from σd(A). Therefore λσess(A)rankPε= for every ε>0.
  4. σess(A) is closed. If λnσess(A) and λnλ, then for every ε>0 some n has (λnε/2,λn+ε/2)(λε,λ+ε), so rankPεrankPε/2(n)= and λσess(A) by item 3.

The closure conclusion is in R, and also in C since R is closed there. A spectral accumulation point has infinitely many spectral points in every surrounding interval, so item 2 forces infinite rank. An isolated eigenvalue of infinite multiplicity has its infinite-dimensional eigenspace inside every interval range by item 1. Consequently an accumulation point of σ(A) and an isolated eigenvalue of infinite multiplicity both lie in σess(A), and σ(A) is the disjoint union of σd(A) and σess(A).

TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Weyl criterion for the essential spectrum

Statement

Assume the Axiom of Choice. Let A be a self-adjoint operator on a complex Hilbert space H and let λR. Then λσess(A) if and only if there is a sequence xnD(A) with xn=1, xn0 weakly (Weak convergence of nets and sequences) and (Aλ)xn0. The sequence may be chosen orthonormal; such a sequence is called a singular Weyl sequence for λ.

Sequences below are indexed by all n in N={0,1,2,...}; the shrinking radii are 1/(n+1). Inner products are linear in the first variable.

Facts & Assumptions

[A1]

For every real lambda, membership in the essential spectrum is equivalent to infinite rank of every projection Pε=E((λε,λ+ε)), epsilon>0. This equivalence includes real resolvent points and isolated finite-multiplicity eigenvalues. On the zero Hilbert space the essential spectrum is empty. Discrete and essential spectrum of a self-adjoint operator

[A2]

On nonzero H the spectral theorem gives the domain D(A)={x:v2dEx<}. Projections multiply by intersection and Ex(C)=E(C)x2. The integral norm identity and Borel sum rule give (Aλ)x2=vλ2dEx on D(A). Spectral theorem for unbounded self-adjoint operators (PVM form) Projection valued measure The unbounded PVM integral is densely defined, closed and normal Unbounded Borel functional calculus: domains, products, spectral mapping

[A3]

Weak convergence means convergence against every bounded linear functional. Under Countable Choice every such functional on a Hilbert space is xx,y; these pairings are bounded by Cauchy-Schwarz. Bessel bounds the sum of squared coefficients against an orthonormal family by the squared norm. Weak convergence of nets and sequences Riesz representation for Hilbert spaces Cauchy–Schwarz: x,yxy, with equality exactly for dependent pairs The Bessel inequality for an arbitrary orthonormal family Orthonormal families, complete orthonormal systems and Hilbert bases

[A4]

The declared AC supplies the spectral theorem and directly chooses a successor for every extendible finite orthonormal list; iterating that fixed choice function from the empty list gives the required sequence. The Axiom of Choice The axiom of dependent choice: a relation in which every element is related to something admits an N-indexed chain

Proof

technique · direct

Given: the self-adjoint A, real lambda and the hypotheses of the direction under consideration.

1.1

If H={0}, there is no unit vector and the essential spectrum is empty, so both sides are false. Otherwise use the PVM of [A2]. For any bounded interval J and x in ran E(J), the projection identity gives E_x(R\J)=0 and v2dEx(supvJv)2x2<, hence x belongs to D(A). For any x in D(A) and epsilon>0, [A2] gives (IPε)x2=Ex({vλε})ε2vλ2dEx=ε2(Aλ)x2. The closed complement includes both interval endpoints.

A1A2
1.2

A finite-dimensional range of an orthogonal projection P has a finite orthonormal basis e_1,...,e_r: from a finite linear basis, successively subtract its projections onto the previously obtained vectors and normalize the nonzero residuals (nonzero follows from linear independence). Then Px=j=1rx,ejej, since its difference from that sum lies in the range and is orthogonal to its basis. If x_n is weakly null, [A3] gives each coefficient tending to zero and Pxn2=j=1rxn,ej20. For rank zero this is the empty sum and Px=0.

A2A3
1.3

Every orthonormal sequence is weakly null: Bessel gives n0y,xn2y2 for each y. If infinitely many coefficients had modulus at least epsilon>0, finite partial sums of arbitrarily many such terms would exceed this bound. Thus their moduli tend to zero; conjugate symmetry gives xn,y0, and Riesz gives convergence against every bounded linear functional.

A3A4
2.1

Suppose a singular Weyl sequence exists. If any P_epsilon had finite rank, step 1.2 would give P_epsilon x_n to zero, while step 1.1 and the residual hypothesis would give (I-P_epsilon)x_n to zero. The triangle inequality would contradict norm x_n=1. Thus every P_epsilon has infinite rank, and [A1] proves lambda belongs to the essential spectrum. This proves the implication for all real lambda, including exclusion of real resolvent points, rather than merely excluding the discrete spectrum.

A1step 1.1step 1.2
2.2

Suppose lambda is in the essential spectrum. For n>=0 write V_n=ran E((lambda-1/(n+1),lambda+1/(n+1))), infinite dimensional by [A1]. Given a finite list of n previously chosen orthonormal vectors x_0,...,x_(n-1), there is a nonzero vector in V_n orthogonal to them: choose n+1 linearly independent vectors in V_n and solve the n homogeneous linear equations for their pairings with the preceding vectors; a nonzero coefficient solution exists by finite-dimensional elimination, and independence makes its vector nonzero. Normalize it. For n=0 choose any nonzero vector in V_0 and normalize; there are no orthogonality equations. On the set of finite lists meeting these conditions, the relation of adjoining such a vector is entire. Apply DC from [A4] with the empty list as initial point; the compatible lists define x_n for every n>=0. The sequence is orthonormal and lies in D(A) by step 1.1. Its scalar measure is carried by the stated interval, so (Aλ)xn2=vλ2dExn(n+1)2xn2=(n+1)2. Step 1.3 gives weak nullity. This constructs the required sequence and proves the converse.

A1A2A4step 1.1step 1.3
3.1

The two implications are steps 2.1 and 2.2. Finite-dimensional H (including dimension one) has only finite-rank interval projections, so neither side holds there. Infinite multiplicity at an isolated point and spectral accumulation points are both covered by the same infinite-rank construction. Lambda=0 is allowed, since only the positive radii n+1 and epsilon are inverted. The empty initial list and the index-zero vector are included in step 2.2; the Choice use is [A4] and the spectral/Riesz interfaces, with no separability assumption.

A1A4step 2.1step 2.2

Source notes

Teschl, Lemma 6.17, printed pp.170–171 (PDF pp.181–182), gives the singular Weyl criterion and its complete projection-estimate proof. The present proof uses the infinite-rank characterization directly in both directions and a DC construction on shrinking interval ranges, with the library's zero-based indexing.

DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Relative compactness with respect to an operator

Definition

Assume the Axiom of Choice. Let A be a self-adjoint operator on a complex Hilbert space H, and let B:D(A)H be linear and bounded for the graph norm of A (Relative boundedness with respect to an operator). Then B is A-compact, or relatively compact with respect to A, when BRA(z)K(H) is a compact operator (Compact linear operator) for one, equivalently for every, zρ(A).

The resolvent set is nonempty: iρ(A) by Resolvent of a self-adjoint operator: nonreal resolvents and the estimate. If H={0}, all operators here are the unique operator, compact with bound zero; the assertions hold directly. Below suppose H{0}.

Well-definedness, with proofs.

  1. BRA(z) is everywhere defined and bounded. RA(z) maps H into D(A) and B is graph-norm bounded there, so BRA(z)yaARA(z)y+bRA(z)y(a(1+zRA(z))+bRA(z))y using ARA(z)=zRA(z)I, so that ARA(z)1+zRA(z) (Resolvent and spectrum of an unbounded operator).

  2. Independence of z. Write Rz=(zIA)1. For yH, RwyD(A) and (zIA)Rwy=y+(zw)Rwy. Applying Rz gives Rwy=Rzy+(zw)RzRwy, since Rz(zIA) is the identity on D(A). Thus RzRw=(wz)RzRw. Interchanging z,w also gives RzRw=(wz)RwRz. This derives both orders without an unproved resolvent identity; the two sides of the latter identity have values in D(A), so applying the linear map B gives BRz=BRw+(wz)(BRw)Rz. Compactness at w implies compactness at z by composition with the bounded Rz and finite linear combinations Compositions with a compact operator are compact Linear combinations of compact operators are compact. Exchanging z,w proves the converse, including the trivial case z=w.

  3. Vector space. If B1,B2 are graph-norm bounded and A-compact, then αB1+βB2 is graph-norm bounded and (αB1+βB2)RA(z)=αB1RA(z)+βB2RA(z) is compact, being a linear combination of compact operators (Linear combinations of compact operators are compact).

  4. An A-compact B has A-bound zero. For zρ(A) and ψD(A), the inverse identity gives Bψ=BRA(z)(zIA)ψ. Hence, with az=BRA(z), BψazAψ+azzψ. It suffices to prove ain0 along positive integers n. The spectral theorem Spectral theorem for unbounded self-adjoint operators (PVM form) and product/domain rule Unbounded Borel functional calculus: domains, products, spectral mapping identify RA(in) with the bounded function (inμ)1 of A: multiplication by inμ gives the two inverse identities, with range in D(A) since both (inμ)1 and μ(inμ)1 are bounded. Consequently Fn:=(iIA)RA(in) is the bounded function hn(μ)=(iμ)/(inμ) of A. For real μ and n1, hn(μ)2=1+μ2n2+μ21,hn(μ)0. The bounded PVM calculus and its adjoint rule Bounded borel pvm integral Pvm integral is a star homomorphism give Fn=hn(A). For every vH both squared norms Fnv2 and Fnv2 equal hn2dEv, which tends to zero by Dominated convergence, dominated by 1 in the finite measure of mass v2. In particular both families converge strongly to zero.

    Put C=BRA(i), compact by item 2. The inverse identity on D(A) gives CFn=BRA(in). Suppose its norm does not tend to zero. There exist δ>0, a strictly increasing integer subsequence nk, and, using the declared AC, vectors yk with yk1 and CFnkyk>δ. For any vH, Fnkyk,v=yk,FnkvFnkv0. Riesz representation Riesz representation for Hilbert spaces therefore proves Fnkyk0. A compact operator sends a weakly null sequence to a norm-null sequence under AC Compact operator sends weakly convergent sequences to norm convergent sequences, contradicting the displayed lower bound. Thus ain0. Given any ε>0, choose n with ain<ε in the first estimate: b=nain is finite and the A coefficient is below ε. Its infimum is therefore zero. This does not assert that the zero coefficient itself is attained. The full AC assumption covers the spectral theorem and the compactness/sequence argument.

TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Weyl's theorem: invariance of the essential spectrum

Statement

Assume the Axiom of Choice. Let A,C be self-adjoint operators such that RA(z)RC(z) is compact for one nonreal z (equivalently, for every zρ(A)ρ(C)). Then σess(A)=σess(C). In particular a bounded self-adjoint compact perturbation preserves the essential spectrum, and if K is symmetric and A-compact, then σess(A+K)=σess(A) with D(A+K)=D(A).

Facts & Assumptions

[A1]

λσess(S) exactly when S has an orthonormal singular Weyl sequence at λ (Weyl criterion for the essential spectrum, Discrete and essential spectrum of a self-adjoint operator).

[A2]

For zρ(S) and λR one has RS(z)+1λzI=1λzRS(z)(Sλ) on D(S), equivalently RS(z)(Sλ)=(λz)(RS(z)+1λzI) (Resolvent and spectrum of an unbounded operator).

[A3]

A compact operator maps weakly convergent sequences to norm convergent sequences, and RS(z) is bounded (Compact operator sends weakly convergent sequences to norm convergent sequences, Compact linear operator).

[A4]

For self-adjoint A,C and nonreal z,w, the bounded resolvent identity gives Dw=[I+(wz)RA(z)]1Dz[I(wz)RC(w)], where Dz:=RA(z)RC(z) and the inverse first factor is I+(zw)RA(w). Hence compactness of Dz transfers to Dw, and conversely by exchanging z,w (The resolvent star algebra is dense in C_0(R), Compositions with a compact operator are compact).

[A5]

An A-compact symmetric K has A-bound zero, so Kato-Rellich makes A+K self-adjoint on D(A); the second resolvent identity RA+K(z)RA(z)=RA+K(z)KRA(z) holds for z in the common resolvent set (Relative compactness with respect to an operator, Kato-Rellich theorem, Second resolvent identity for a closed perturbation).

Proof

technique · direct

Given: Self-adjoint A,C with compact resolvent difference at a nonreal z.

1.1

Let λσess(A) and let (xn) be the orthonormal Weyl sequence of [A1]. By [A2] and (Aλ)xn0 one has (RA(z)+1λz)xn0; since RA(z)RC(z) is compact and xn0, [A3] gives (RC(z)+1λz)xn0 as well.

A1A2A3
1.2

Parameter independence is [A4].

A4
2.1

Then (Cλ)RC(z)xn=(zλ)RC(z)xnxn0 by [A2] and step 1.1, and RC(z)xnλz1>0; the normalized vectors yn:=RC(z)xn1RC(z)xn lie in D(C), have unit norm, converge weakly to 0 and satisfy (Cλ)yn0, so they form a singular Weyl sequence and λσess(C) by [A1]. Interchanging the roles of A and C gives equality.

A1A2step 1.1
3.1

Bounded compact perturbations: if K is bounded, symmetric and compact, then A+K is self-adjoint with D(A+K)=D(A) by Kato-Rellich applied with the admissible pair (0,K), and the second resolvent identity gives RA+K(z)RA(z)=RA+K(z)KRA(z), compact as a product of the compact K with bounded factors; so σess(A+K)=σess(A) by step 2.1.

A4A5step 2.1
3.2

A-compact perturbations: for symmetric K that is A-compact, [A5] makes A+K self-adjoint with D(A+K)=D(A) and gives RA+K(z)RA(z)=RA+K(z)KRA(z), a product of the bounded operator RA+K(z) with the compact operator KRA(z), hence compact; then step 2.1 applies.

A5step 2.1
4.1

The claims are steps 1.1, 1.2 and 2.1 (compact resolvent difference), 3.1 (bounded compact perturbations) and 3.2 (A-compact perturbations). ∎

DefinitionDefinition: AI-adaptedProof: Not applicableaudited 2026-09-22Open item page →

Norm and strong resolvent convergence

Definition

Assume Countable Choice (The Axiom of Countable Choice (ACω)). Let An (nN) and A be self-adjoint operators on the same Hilbert space H and fix a nonreal z0. One writes AnA in the norm resolvent sense when RAn(z0)RA(z0)0 in operator norm, and in the strong resolvent sense when RAn(z0)xRA(z0)xfor every xH, that is, strong operator convergence (The operator norm as the least bound and as the unit-sphere or unit-ball supremum, A bounded linear operator between normed spaces, Weak convergence of nets and sequences).

Both notions are well posed for every choice of nonreal z0: by Resolvent of a self-adjoint operator: nonreal resolvents and the estimate each nonreal number belongs to ρ(An)ρ(A), so all resolvents occurring are bounded with RAn(z0)1/Imz0. The definition deliberately does not assert independence of the parameter z0: that independence is a theorem, proved for the norm case by the resolvent-star-algebra density lemma below and used in the continuous-calculus-under-resolvent-convergence theorem below. Norm resolvent convergence implies strong resolvent convergence, and both are notions about the resolvents rather than about the operators: no convergence of the operators themselves is asserted or implied.

LemmaStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

The resolvent star algebra is dense in C_0(R)

Statement

Fix zR and let r(x)=(xz)1. Let A be the -algebra generated by r in C0(R), that is, the uniform closure of the linear span of the products rkrj with k+j1. Then A=C0(R). Under Countable Choice (The Axiom of Countable Choice (ACω)), consequently, if self-adjoint An,A satisfy RAn(z)RA(z) strongly (respectively in norm) at one nonreal z, then the same convergence holds at every nonreal z, in particular at z.

Facts & Assumptions

[A2]

Let X be a compact Hausdorff space and let BC(X,C) be a self-adjoint complex function algebra containing the constants, separating points, with no common zero. Then B is uniformly dense in C(X,C) (Complex Stone–Weierstrass dichotomy for separating self-adjoint algebras; the unital case is dense).

[A3]

Under Countable Choice, with Dz:=RA(z)RA(z) for self-adjoint A,A and nonreal z, the resolvent identity gives Dz=[I+(zz)RA(z)]1Dz[I(zz)RA(z)] for nonreal z, the first factor being I+(zz)RA(z) and the second I+(zz)RA(z); these affine transforms of resolvents are bounded with norms at most 1+zz/Imz (Resolvent and spectrum of an unbounded operator, Resolvent of a self-adjoint operator: nonreal resolvents and the estimate, Norm and strong resolvent convergence, The Axiom of Countable Choice (ACω)).

Proof

technique · direct

Given: zR, r(x)=(xz)1 and the -algebra A generated by r.

1.1

Extend r to R by r()=0; then rC0(R)C(R) and r separates the points of R: it is injective on R, and r(x)0=r().

A1
1.2

Parameter independence: for nonreal z the difference Dz=RA(z)RA(z) factors as in [A3] through Dz, with both outer factors of norm at most 1+zz/Imz, since they are I+(zz)RA(z) and I+(zz)RA(z) and RA(z),RA(z)1/Imz; hence Dz0 in norm whenever Dz0 in norm, and Dzx0 for every x whenever Dzy0 for every y, because bounded operators preserve both modes of convergence.

A3
2.1

The algebra B:={f+c1:fA,cC}C(R) contains the constants and r,r, is self-adjoint, separates points by step 1.1, and has no common zero because of the constant function 1; hence B is uniformly dense in C(R) by [A2].

A1A2step 1.1
3.1

Therefore A=C0(R): given fC0(R) and ε>0, density of B gives f=c1+g with gA and fc1g<ε; evaluating at , where f()=0, gives c<ε, so fg<2ε and gA.

A1step 2.1
4.1

The density claim is step 3.1 and the consequence is step 1.2, which also covers z=z. ∎

TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Continuous functional calculus under resolvent convergence

Statement

Assume the Axiom of Choice. Let An,A be self-adjoint operators on a complex Hilbert space H and suppose AnA in the strong resolvent sense. Then f(An)xf(A)x for every bounded continuous f:RC and every xH. If AnA in the norm resolvent sense, then f(An)f(A)0 for every bounded continuous f with limt+f(t)=limtf(t). In both cases the conclusion does not depend on the nonreal parameter used in the definition of convergence.

Facts & Assumptions

[A1]

On nonzero complex H, AC supplies the spectral PVM of each self-adjoint S, representing S as the integral of the identity on its squared-integrability domain. The unbounded calculus has the exact product domain D(g(S))D((fg)(S)), and sums and products agree with their pointwise counterparts on their domains (Spectral theorem for unbounded self-adjoint operators (PVM form), Unbounded Borel functional calculus: domains, products, spectral mapping).

[A2]

The bounded PVM calculus is linear, unital, multiplicative and conjugation preserving, satisfies h(S)h and h(S)x2=h2dExS, and ExS(R)=x2. Its Countable Choice assumptions are supplied by AC (Bounded borel pvm integral, Pvm integral is a star homomorphism, Scalar and complex measures from a pvm, The Axiom of Choice).

[A3]

RS(z)=(zIS)1 exists for nonreal z with RS(z)1/Imz. Convergence is initially assumed at one fixed nonreal parameter only (Resolvent of a self-adjoint operator: nonreal resolvents and the estimate, Resolvent and spectrum of an unbounded operator, Norm and strong resolvent convergence).

[A4]

For fixed nonreal z, the linear span of rkrj, k+j1, r(λ)=(λz)1, is uniformly dense in C0(R) (The resolvent star algebra is dense in C_0(R)). Only its function-algebra density assertion is used; parameter independence is proved below.

[A5]

Scalar dominated convergence holds with an integrable majorant (Dominated convergence). Operator norm bounds give BxBx and, by applying this twice, BCBC (A bounded linear operator between normed spaces, The operator norm as the least bound and as the unit-sphere or unit-ball supremum).

Proof

technique · direct

Given: AC, the self-adjoint operators and convergence at a fixed nonreal z in the indicated mode.

1.1

If H={0}, all operators, resolvents and bounded functions of them are the unique full-domain operator, so all assertions hold. Suppose H{0}. For nonreal w put qw(λ)=(wλ)1. Both qw and λqw are bounded on the real line. The product-domain rule in [A1] shows that qw(S) maps H into D(S) and (wIS)qw(S)=I on H, while qw(S)(wIS)=I on D(S). Thus qw(S)=RS(w) with the convention of [A3]. The bounded calculus agrees with its unbounded truncation definition because the truncations are eventually the same bounded function.

A1A2A3given
2.1

Put d=zw. The scalar identity (1+dqw)qz=qw=qz(1+dqw) and [A2] give these identities for each single operator S. Define Dn(v)=RAn(v)RA(v) and F=I+dRA(w), which is independent of n. Expanding the product using the preceding identities yields Dn(w)=(I+dRAn(w))Dn(z)F: its two terms are RAn(w)F and (I+dRAn(w))RA(w), whose mixed terms cancel. The left factor has norm at most C=1+d/Imw. Consequently Dn(w)xCDn(z)Fx0 in the strong case, since Fx is fixed, and Dn(w)CDn(z)F0 in the norm case. This proves parameter independence without taking a strong limit on a varying vector.

A2A3A5step 1.1
3.1

In particular convergence holds at z and z, and r(An)=RAn(z), r(An)=RAn(z), with the analogous formulas for A. Both sequences are uniformly bounded. If BnB and CnC strongly and supnBnK, then (BnCnBC)xK(CnC)x+(BnB)Cx0. In the norm case the same inequality with operator norms proves convergence of products when the factors are uniformly bounded. Iteration and finite linear combinations, using [A2], therefore prove convergence for every polynomial in r,r with zero constant term.

A2A3A5step 1.1step 2.1
4.1

Given fC0(R) and δ>0, choose one such polynomial p with fp<δ by [A4]. The bounded calculus gives (f(An)f(A))x2δx+(p(An)p(A))x. For fixed p the last term tends to zero by step 3.1; since δ is arbitrary, strong convergence follows for each x (including x=0 directly). The operator-norm inequality is f(An)f(A)2δ+p(An)p(A), proving the norm version as well.

A2A4A5step 3.1
5.1

In the norm case, if f has a common finite limit L at both ends, then g=fLC0(R). Unital linearity gives f(An)f(A)=g(An)g(A), so step 4.1 proves the assertion.

A2step 4.1
5.2

For the strong case let f be any bounded continuous function and set M=f. For integers m1 define χm(λ)=min(1,max(0,m+1λ)). These are continuous, compactly supported, between zero and one, equal to one on [m,m], and converge pointwise to one. Thus (Iχm(A))x2=1χm2dExA0 by dominated convergence with majorant 1, integrable against the finite measure of mass x2. For fixed m, both χm and fχm belong to C0(R), so their calculi converge strongly by step 4.1.

A2A5step 4.1
6.1

Bounded multiplicativity and linearity give the exact four-term decomposition f(An)f(A)=f(An)(Iχm(A))+f(An)(χm(A)χm(An))+(fχm)(An)(fχm)(A)+f(A)(χm(A)I). Applying it to x, its norm is at most 2M(Iχm(A))x+M(χm(A)χm(An))x+((fχm)(An)(fχm)(A))x. If M=0 the claim is immediate. Otherwise choose m so that the first term is less than half a prescribed positive error, using step 5.2, and then n so that the other two together are less than its other half. This proves strong convergence for bounded continuous f.

A2A5step 5.2
7.1

Step 2.1 proves independence of every nonreal parameter, step 5.1 the norm conclusion and step 6.1 the strong conclusion. AC supplies the spectral and countable-choice calculus hypotheses; the cutoffs are explicit. The zero Hilbert space and zero function have been treated in steps 1.1 and 6.1; no limit at either infinity is required in the strong case.

A1A2A3step 1.1step 2.1step 5.1step 6.1

Source notes

Teschl, Theorem 6.31 and its proof, pp.179-180, gives the polynomial approximation and four-term cutoff route, with Corollary 6.32 giving parameter independence. Here the latter is proved first using a resolvent-difference factorization with a fixed right factor. The library convention is (zIS)1, so r(S)=RS(z). No general continuity of adjoints for strongly convergent bounded operators is assumed.

CorollaryStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Unitary groups converge under strong resolvent convergence

Statement

Assume the Axiom of Choice. Let An,A be self-adjoint operators on a complex Hilbert space H with AnA in the strong resolvent sense. Then eitAnxeitAxfor every tR and every xH. If, in addition, A and all An are bounded below by a common real constant γ (meaning Sx,xγx2 for xD(S) and S=A,An), then etAnetA strongly for every t0.

Facts & Assumptions

[A1]

For a self-adjoint S with spectral PVM, eitS is the Borel calculus of λeitλ and is a strongly continuous unitary group, with infinitesimal generator iS (A self-adjoint operator generates a strongly continuous unitary group, Infinitesimal generator of a unitary group, Strongly continuous one-parameter unitary group).

[A2]

Under AC, strong resolvent convergence gives g(An)xg(A)x for every bounded continuous g:RC and every x (Continuous functional calculus under resolvent convergence, The Axiom of Choice).

[A3]

On nonzero H, each self-adjoint S has a spectral PVM ES with D(S)={x:λ2dExS<} and S=λdES. For domain vectors, Sx,x=λdExS, using the quadratic pairing identity and the reality of λ in the first-variable-linear convention. Functions agreeing off a measurable ES-null set have the same integral operator and domain; the zero-space calculus is defined directly (Spectral theorem for unbounded self-adjoint operators (PVM form), The unbounded PVM integral is densely defined, closed and normal, Integral of a measurable function against a projection-valued measure).

[A4]

PVM projections satisfy E(B)E(C)=E(BC) and E(R)=I; the scalar measures are positive of mass x2, and scalar monotone convergence holds (Projection valued measure, Scalar and complex measures from a pvm, Monotone convergence for the integral).

Proof

technique · direct

Given: AC and the self-adjoint operators with strong resolvent convergence in the statement.

1.1

If H={0}, all operators in either conclusion are its unique operator, so both conclusions hold. Otherwise the spectral PVMs exist by [A3]. Fix any tR, including negative times. The function λeitλ is continuous and has absolute value one everywhere. Thus [A2] gives eitAnxeitAx for every x; [A1] identifies these as the stated unitary groups. At t=0 each operator is I.

A1A2A3given
1.2

For the second claim suppose S is one of A,An and satisfies the common lower bound. Let Bm=[m,γ1/m] for integers m1, with an empty interval interpreted as empty. If ES(Bm)0, choose x=ES(Bm)y0 for some y. The projection identities give ES(RBm)x=0, so the scalar measure of x is supported on Bm. As Bm is bounded, [A3] gives xD(S) and Sx,x=λdExS(γ1/m)x2<γx2, a contradiction. Hence each ES(Bm)=0. The sets Bm increase to (,γ); monotone convergence gives ExS((,γ))=0 for every x. Since ES(B)x2=ES(B)x,x, the projection ES((,γ)) itself is zero.

A3A4given
2.1

Now fix t0 and define gt(λ)=etmax(λ,γ). It is continuous and bounded by etγ, and it agrees with etλ on [γ,). Step 1.2 and null-set invariance in [A3] give equality of the integral operators gt(S)=etS, including domains, for S=A,An. In particular each exponential here has full domain and is bounded: its defining squared integral is at most e2tγx2, and its quadratic norm identity gives the same operator bound. Applying [A2] to gt proves etAnxetAx for every x.

A2A3A4step 1.2
3.1

The first conclusion holds for every real time by step 1.1, and the second for every nonnegative time by step 2.1. At time zero both reduce to the identity. The lower bound is assumed for the limit and every approximant; no preservation-of-lower-bound theorem is assumed. AC supplies the spectral and calculus hypotheses, including their Countable Choice assumptions.

A1A2A3step 1.1step 2.1
LemmaStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Spectral form domain and core of a semibounded operator

Statement

Assume the Axiom of Choice. Let A be self-adjoint on a complex Hilbert space H, with AcI meaning Ax,xcx2 for every xD(A), for some real c. Put Q(A):=D((AcI)1/2),qA[x]:=cx2+(AcI)1/2x2(xQ(A)). Then Q(A) and qA do not depend on the choice of the constant cinfσ(A) (the form qA[x]=λdEx is itself unchanged, while the summand (AcI)1/2x2 changes by the constant (cc)x2 when c is replaced by cc), qA[x]=λdEx(λ)=Ax,x for xD(A), the domain D(A) is dense in Q(A) for the norm xQ=(x2+(AcI)1/2x2)1/2, and qA is a closed quadratic form with qA[x]cx2.

Here the square root is the Borel calculus of max(λc,0); the proof shows that E is carried on [c,). A closed semibounded quadratic form means the diagonal of a Hermitian sesquilinear form on a dense linear domain, complete in the displayed shifted form norm. If H={0}, use the unique PVM and operator and the convention inf(empty spectrum)=+infinity.

Facts & Assumptions

[A1]

The spectral theorem gives D(A)={x:λ2dEx<} and A as the coordinate integral on nonzero H. The unbounded integral is closed, has linear domain, squared norm integral, and real-function pairing f(E)x,x=fdEx. Real f gives a self-adjoint operator. The zero-space integral is defined directly. Spectral theorem for unbounded self-adjoint operators (PVM form) The unbounded PVM integral is densely defined, closed and normal Integral of a measurable function against a projection-valued measure

[A2]

Projections multiply by intersection, and Ex(B)=E(B)x2 is a finite measure of mass x2. Consequently EE(J)x(B)=Ex(BJ), using E(B)E(J)=E(BJ) and the norm formula; complementary projections give the analogous complementary restriction. The spectrum of A is the essential range of the coordinate function. Projection valued measure Unbounded Borel functional calculus: domains, products, spectral mapping

[A3]

Scalar dominated convergence applies to the finite measures E_x. Dominated convergence

[A4]

H is complete and its inner product is first-linear. Self-adjoint operators have dense linear domains. The assumed AC directly supplies every choice function required by the PVM, closed-integral and spectral-theorem interfaces. Hilbert space Symmetric, self-adjoint and essentially self-adjoint operators The Axiom of Choice

Proof

technique · direct

Given: AC, self-adjoint A and its lower bound c.

1.1

If H={0}, every domain and form consists of zero, all norms vanish, and every assertion follows directly from the zero-space convention in [A1]. Otherwise obtain E from [A1], under the choice assumption in [A4]. For Jm=[m,c1/m] (empty intervals allowed), a vector v=E(Jm)x belongs to D(A) by [A2] and boundedness of lambda on J_m. If v were nonzero then Av,v=JmλdEv(c1/m)v2, contradicting the lower bound. Thus E(J_m)=0 for all positive integers m. Their union is (,c), whose scalar measures therefore vanish by countable subadditivity. The projection norm formula gives E((-infinity,c))=0. The essential-range description in [A2] implies σ(A)[c,).

A1A2A4given
2.1

Put B=max(λc,0)(E). It is closed and self-adjoint by [A1], and step 1.1 gives D(B)={x:(λc)dEx<} and Bx2=(λc)dEx. Integrals here and below can be restricted to [c,infinity). Since λ(λc)+c there, lambda is absolutely integrable for x in Q(A). Hence qA[x]=cx2+Bx2=λdEx and qA[x]cx2.

A1A2step 1.1
3.1

For any other lower spectral bound c'<=c, λc=(λc)+(cc) on the carrier. Since E_x has finite mass, the two domain integrals are finite simultaneously. Adding the appropriate constant times the mass gives the same q_A, while the square-root squared norm increases by (cc)x2. Two arbitrary admissible lower bounds can be compared in their numerical order, so this proves full independence. Their squared form norms differ by that same multiple of x2, hence are equivalent since each dominates x2.

A2step 2.1
3.2

If x belongs to D(A), then λc1+λ2+c on the carrier, so x belongs to Q(A). By [A1] and step 2.1, qA[x]=λdEx=Ax,x.

A1A2step 2.1
3.3

For x in Q(A), set xn=E([n,n])x, n>=1. By [A2], λ2dExnn2x2, so x_n belongs to D(A). The same restriction identity gives xxnQ2=λ>n(1+λc)dEx0 by dominated convergence, with nonnegative integrable majorant 1+λc on the carrier. This is the asserted form-norm density, with the exact identity zQ2=qA[z]+(1c)z2.

A1A2A3step 2.1
4.1

The form is the diagonal of a(x,y)=cx,y+Bx,By on the linear domain D(B); this is Hermitian and sesquilinear by [A4]. Its domain is dense in H because it contains D(A) by step 3.2. For a Cauchy sequence in the form norm, both x_n and Bx_n are Cauchy in H. Completeness gives limits x and y. Closedness of B implies x in D(B) and Bx=y. Therefore xnxQ2=xnx2+BxnBx20, proving completeness and closedness in the stated sense.

A1A4step 2.1step 3.2
5.1

Steps 2.1 and 3.1 establish the domain, integral identity, lower bound and independence; steps 3.2 and 3.3 give the operator-domain identity and core, and step 4.1 gives the closed quadratic form. Positive integer cutoffs are specified without choices. AC is inherited through [A4]; negative and zero lower bounds are allowed without taking a square root of q_A itself. The zero Hilbert space was handled in step 1.1.

A4step 1.1step 2.1step 3.1step 3.2step 3.3step 4.1
TheoremStatement: Literature-sourcedProof: AI-adaptedaudited 2026-09-22Open item page →

Min-max principle below the essential spectrum

Statement

Assume the Axiom of Choice. Let A be a self-adjoint operator on a complex Hilbert space H, bounded below by c, with form domain Q(A) and form q_A of Spectral form domain and core of a semibounded operator. Put Λ=infσess(A), with inf(empty)=+infinity. List the eigenvalues below Lambda in nondecreasing order with multiplicity as E_1,E_2,..., setting E_n=Lambda after the list is exhausted. Then for every integer n>=1, En=infLQ(A) lineardimL=n supxLx=1qA[x]=supFQ(A) lineardimFn1 infxQ(A)Fx=1qA[x]. All infima over empty sets are +infinity. Finite dimensions here are ordinary linear dimensions, equivalently Hilbert dimensions for these finite-dimensional subspaces. The second outer family is always nonempty, since it contains {0}. Whenever Q(A) contains an (n-1)-dimensional subspace, the same value is obtained by requiring dim F=n-1. The at-most convention includes the exhausted finite-dimensional case for all n without taking a supremum over an empty outer family.

The first infimum is unchanged if its trial spaces L are restricted to D(A). If E_n<Lambda, both outer values are attained, respectively by spans of the first n and the first n-1 orthonormal eigenvectors (the latter span is {0} when n=1).

Facts & Assumptions

[A1]

The essential spectrum is closed, and a real v is in it exactly when every interval about v has infinite-rank spectral projection. A finite-rank interval contains only finitely many spectral points, and each such point is an isolated finite-multiplicity eigenvalue. The singleton projection is the eigenspace projection. The spectrum carries E. Discrete and essential spectrum of a self-adjoint operator

[A2]

The spectral PVM has intersection products, orthogonal disjoint ranges and strong countable additivity. The spectral domain is D(A)={x:v2dEx<}. For a vector in the range of a bounded interval projection its scalar measure is carried by that interval, so it belongs to D(A). The calculus identifies spectral support with spectrum. Projection valued measure Spectral theorem for unbounded self-adjoint operators (PVM form) Unbounded Borel functional calculus: domains, products, spectral mapping

[A3]

qA[x]=vdEx is finite on Q(A), qA[x]cx2, and E is carried on [c,infinity). The form domain is a dense linear subspace containing D(A); its form is Hermitian with qA[x]=Ax,x on D(A). Self-adjoint operators are symmetric. Spectral form domain and core of a semibounded operator Symmetric, self-adjoint and essentially self-adjoint operators

[A4]

Orthogonality uses the first-variable-linear inner product; a finite orthonormal family is linearly independent and the squared norm of its linear combination is the sum of squared coefficient moduli. Pairings are continuous by Cauchy-Schwarz. A closed bounded real interval has the finite-open-subcover property. Orthogonality and the orthogonal complement Orthonormal families, complete orthonormal systems and Hilbert bases Cauchy–Schwarz: x,yxy, with equality exactly for dependent pairs Heine-Borel by bisection: every closed bounded interval [a,b] is compact

[A5]

AC is declared for the spectral and form interfaces and, if the eigenvalue list is infinite, for choosing orthonormal bases in its countably many finite-dimensional eigenspaces. All variational subspace and kernel arguments below are finite-dimensional. The Axiom of Choice

Proof

technique · direct

Given: A>=c, its spectral measure E, and the variational quantities in the statement.

1.1

Suppose first H is nonzero. For any real r<Lambda with r>=c, each point of [c,r] has a finite-rank interval neighborhood by [A1]. Use the family of all such intervals and the finite subcover property [A4]. Each chosen interval contains finitely many spectral points, so sigma(A) intersect [c,r] is finite. These points are discrete eigenvalues of finite multiplicity by [A1]. Since E is carried by sigma(A) intersect [c,infinity), E((,r]) is the finite sum of their singleton projections. For r<c it is zero. In particular the total multiplicity below each r<Lambda is finite. The eigenvalues below Lambda can therefore be ordered from below: if any remain, choose one such u; the nonempty finite set of remaining values <=u has a smallest member, which is the smallest remaining value altogether. Repeat with its finite multiplicity. No value can be omitted forever, since only finitely many terms lie below any fixed u<Lambda. A countable increasing sequence of bounds r approaching Lambda (or infinity) covers all values; [A5] licenses the associated orthonormal eigenvector choices.

A1A2A3A4A5
1.2

Write alpha_n for the first outer value over Q(A), beta_n for the second (dim F<=n-1), and alpha_n^D for the first over D(A). If dim L=n and dim F=r<=n-1, choose finite bases and impose the r equations x,fj=0 on x in L. Finite-dimensional elimination gives a nonzero solution, since there are fewer equations than unknowns. Normalize it to obtain a unit vector in L intersect F-perp. Thus the supremum on L is at least the infimum on Q(A) intersect F-perp. For each F take the infimum over L, then the supremum over F, giving beta_n<=alpha_n; this remains true when there is no L because alpha_n=+infinity. The trial-space inclusion gives alpha_n<=alpha_n^D.

A3A4
2.1

If Lambda is finite, it belongs to the essential spectrum: that set is nonempty in this case, bounded below by c, and closed; for every positive integer m choose a point between its infimum and Lambda+1/m and use closedness. Hence every interval around Lambda has infinite-rank projection by [A1]. If Lambda=+infinity and only k eigenvalues with multiplicity occur, there is no other spectrum: every finite spectral point lies below Lambda and is one of these eigenvalues. Support then makes their eigenspaces sum to all of H, so dim H=k and D(A)=Q(A)=H. The same last assertion holds directly with k=0 for H={0}.

A1A2A3A5step 1.1
2.2

If E_n<Lambda, choose orthonormal eigenvectors phi_1,...,phi_n ordered with multiplicity as in step 1.1. Distinct eigenspaces are orthogonal because symmetry gives (EjEk)φj,φk=0; within each eigenspace choose an orthonormal basis by finite Gram-Schmidt. For L0=span{φ1,,φn}, contained in D(A), qA[jajφj]=jEjaj2, so its unit-sphere supremum is E_n. For F0=span{φ1,,φn1}, all eigenspaces strictly below E_n are contained in F_0. The projection E((-infinity,E_n)) therefore annihilates every x in F_0-perp; repetitions of E_n need not be removed. Its scalar measure is carried on [E_n,infinity), so q_A[x]>=E_n for unit x in Q(A) intersect F_0-perp. Equality holds at phi_n. Consequently E_n<=beta_n<=alpha_n<=alpha_n^D<=E_n by step 1.2. This proves the formulas, the D(A) version and both attainments in this case.

A1A2A3A4step 1.1step 1.2
3.1

If E_n=Lambda is finite, exactly k<n eigenvalues occur below Lambda with multiplicity. Let F_0 be the span of all their orthonormal eigenvectors (zero if k=0). It is an admissible space of dimension k<=n-1. By support, vectors perpendicular to it have no spectral mass below Lambda, so the inner infimum is at least Lambda and beta_n>=Lambda. For any epsilon>0, the range of E((Lambda-epsilon,Lambda+epsilon)) is infinite dimensional by step 2.1. Choose n independent vectors there and use finite Gram-Schmidt to get an n-dimensional subspace L_epsilon in the same range. Bounded spectral support puts it in D(A) and gives q_A[x]<=Lambda+epsilon on its unit sphere. Hence alpha_n^D<=Lambda+epsilon. Let epsilon decrease to zero in the inequalities of step 1.2 to conclude beta_n=alpha_n=alpha_n^D=Lambda. No form-cross-term estimate or Weyl-sequence approximation is needed.

A1A2A3A4step 2.1step 1.2
3.2

If E_n=Lambda=+infinity, step 2.1 gives dim H=k<n, D(A)=Q(A)=H, with k the exhausted total multiplicity. There is no n-dimensional trial space, so alpha_n=alpha_n^D=+infinity. F=H has dimension k<=n-1 and is admissible for beta_n. Its orthogonal complement contains no unit vector, so its inner infimum is +infinity and beta_n=+infinity. This includes the zero Hilbert space for every n>=1.

A4step 2.1step 1.2
4.1

Finally suppose Q(A) has a subspace of dimension n-1. Any finite-dimensional F contained in Q(A) with dim F<n-1 can be enlarged inside Q(A) to dimension n-1: as long as its dimension is smaller, choose a vector from the given (n-1)-dimensional subspace not in the current span and adjoin it. Enlarging F shrinks Q(A) intersect F-perp, so its inner infimum cannot decrease. Taking suprema shows that restricting the second outer family to exact dimension n-1 leaves beta_n unchanged; the reverse inequality is family inclusion. When n=1 the only space is F={0}. When H is finite dimensional and n exceeds dim H+1, the at-most convention remains necessary. The real lower bound c was never shifted away, so negative eigenvalues and threshold zero require no special argument. AC is used exactly in [A5] and the infimum-approaching sequence in step 2.1.

A4A5step 2.1step 2.2step 3.1step 3.2

Source notes

Teschl, Section 4.4, printed pp.139–141 (PDF pp.150–152), equations (4.37)–(4.41) and Theorem 4.12, gives the max-min argument using n-1 trial vectors, whose span can have smaller dimension. Theorem 4.14's min-max proof is assigned as Problem 4.11 and its printed trial count is inconsistent with (4.43). The full projection proof above supplies both formulas directly, with n-dimensional min-max spaces and explicit finite-dimensional exhaustion conventions.

5 · Examples, counterexamples and false statements

None yet.

Sources