Contribution
A reference article. It states the correspondence between states and cyclic -representations of a C-algebra, gives the construction (after Kadison, Thm 4.5.2) and the uniqueness (Kadison, Prop. 4.5.3), sketches the proof that the GNS representation is irreducible iff the state is pure, and names the construction’s role in the Gelfand–Naimark theorem, the universal representation and the enveloping von Neumann algebra.
Key insight
A state is enough to rebuild a Hilbert space and a representation in which it is a vector expectation, and the result is fixed up to a unitary that is forced on you: it must send π(a)ξ to π′(a)ξ′. The inner products ⟨π(a)ξ, π(b)ξ⟩ = ρ(b*a) are all the data there is, so any two realizations of them are isometric on a dense set.
Assumptions
- A is a C-algebra.* Unital or not; in the non-unital case the cyclic vector is a limit of an approximate identity. The closing remark extends the theorems to B*-algebras with approximate identity.
- ρ is a state: a positive linear functional of norm 1, equivalently ρ(1) = 1 when A is unital.
- *Representations are nondegenerate -homomorphisms into the bounded operators on a Hilbert space.
- Convention. ⟨a, b⟩ = ρ(b*a) is linear in its first argument; vector states are ⟨π(a)ξ, ξ⟩. The scalars are complex.
Key results
- Vector states. For a *-representation π with unit cyclic vector ξ, a ↦ ⟨π(a)ξ, ξ⟩ is a state.
- Existence (Kadison 4.5.2). For each state ρ there are π on H and a
unit cyclic ξ with ρ(a) = ⟨π(a)ξ, ξ⟩ for all a.
- ⟨a, b⟩ = ρ(ba) is a semi-definite sesquilinear form. The null elements I = {a : ρ(aa) = 0} form a subspace and a left ideal, “the largest left ideal in the null space of ρ”. A/I is an inner-product space, and H is its completion.
- π(a)(b + I) = ab + I is bounded and extends to H, and π is *-preserving.
- ξ is the class of 1, or the limit of the classes of an approximate identity.
- Uniqueness (Kadison 4.5.3). If (π, ξ) and (π′, ξ′) both reproduce ρ with unit cyclic vectors, there is a unitary U : H → H′ with π′(a) = Uπ(a)U* for all a, and U maps π(a)ξ to π′(a)ξ′.
- Significance. Enough pure states exist that the direct sum of their irreducible GNS representations is faithful, which is the heart of the Gelfand–Naimark theorem. The sum over all states is the universal representation, which contains every cyclic representation; its weak-operator closure is the enveloping von Neumann algebra, identified with A**.
- Irreducibility. The states of a unital A form a weak-* compact convex set (Banach–Alaoglu). For C(X), Riesz–Markov–Kakutani makes the positive functionals of norm ≤ 1 the positive Borel measures of mass ≤ 1, and the extreme states the point masses (Krein–Milman); a representation of C(X) is irreducible iff one-dimensional. In general, π with unit cyclic ξ is irreducible iff its state is an extreme point of the positive functionals of norm ≤ 1. The proof sketch uses a Radon–Nikodym-type lemma: a positive g ≤ f has the form g(x*x) = ⟨π(x)ξ, π(x)T_g ξ⟩ with 0 ≤ T_g ≤ 1 in the commutant π(A)′.
- History. Gelfand and Naimark (1943); Segal saw the construction implicit in their work and sharpened it, and in 1947 showed that irreducible representations suffice for a physical system described by an operator algebra, a point von Neumann had shown only for Schrödinger–Heisenberg theory. Stinespring’s theorem on completely positive maps generalizes the construction.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Every state on a C*-algebra is a vector state of a cyclic *-representation | strong | standard theorem, proof sketched (Kadison 4.5.2) |
| C2 | That representation is unique up to unitary equivalence, by the unitary π(a)ξ ↦ π′(a)ξ′ | strong | stated with reference (Kadison 4.5.3), not proved |
| C3 | The GNS representation is irreducible iff the state is pure | strong | standard theorem, proof sketched |
Concepts
- State. A positive linear functional of norm 1 on a C*-algebra.
- Cyclic vector. ξ with {π(x)ξ : x ∈ A} norm-dense in H.
- Left kernel of ρ. I = {a : ρ(a*a) = 0}.
- Pure state. An extreme point of the convex set of states.
- Universal representation. The direct sum of the GNS representations of all states.
Connections
The article’s sources are Kadison and Ringrose, Arveson, Dixmier and Sakai. Murray and von Neumann’s construction of R/M with (ξ, η) = f(y*x) is the first form of it (NOTE-303). The Riesz–Markov– Kakutani theorem it uses for C(X) is the measure-theoretic Riesz theorem, not the Hilbert-space one of LIT-230.
Bearing on the record
- THEORY-004 cites this article as “the GNS statement”. The
article supports the statement used: a state fixes its cyclic
representation up to a unitary, and the unitary is the one sending
π(a)ξ to π′(a)ξ′. Two cautions for that citation.
- The article is about states on C*-algebras. It never mentions positive-definite kernels on a set, RKHS or Moore–Aronszajn, and is over the complex numbers. The kernel form, k(x, y) = ⟨Φ(x), Φ(y)⟩ with minimal realizations unique up to a unitary, comes from Jorgensen & Tian (LIT-259, Remark 1.34, Cor. 1.35), as the theory already says. This article supports only the analogy at the level of algebras.
- The uniqueness’s proof idea is the same as the kernel case’s: the data fix all inner products ⟨π(a)ξ, π(b)ξ⟩, so the map between the dense sets is isometric. That is consistent with the theory’s sentence “the kernel is the invariant and the representation is one realization of it”, but the passage to real representation matrices and orthogonal maps is not here.
- THEORY-017 names this article for the setting where an algebra of observables is supplied from the start, beyond discrete bases. The article fits that reading: GNS takes an algebra and a state as given and produces the Hilbert space, and its history section says Segal took the C*-algebra to be “generated by the observables”. It says nothing about continuous observables or the absence of eigenbases, so the citation marks a setting, as THEORY-017 uses it, and should not be read as support for the claims about continuous observables. The directive in THEORY-017 that calls LIT-241 Deferred lapses once the LIT is Active.
- The skim’s kernel analogy. The skim’s open questions (kernels as states, L²(data) as the commutative case, the Platonic representation hypothesis) are the reader’s. The commutative example the article gives is C(X) with states as probability measures, and it uses it only for irreducibility.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- The uniqueness theorem is stated with a reference, not proved here.
- Small slips: the left kernel is said to be a subspace “by the triangle inequality”, where the usual argument is the Cauchy–Schwarz inequality for positive functionals; the history’s “the C*-algebra is generated by the observables” is loose.
- Nothing on real C*-algebras, on unbounded observables, or on non-cyclic representations beyond the direct-sum remark.