C*-Algebras: Axioms, Spectrum, and the Continuous Functional Calculus

Intuition Beginner

A C*-algebra is a collection of operators that you can add, multiply, scale by complex numbers, and reflect through one extra operation called the star. On the algebra of operators on a Hilbert space 02.11.08, the star sends an operator to its adjoint — the operator that moves to the other side of an inner product. The single rule that makes the whole theory rigid is that the size of an operator and the size of its star-times-itself are locked together.

There is a second picture that turns out to be the same picture. Take a shape, and look at all the continuous complex-valued functions on it. You can add and multiply functions pointwise, and the star is complex conjugation. This family of functions is also a C*-algebra, and a remarkable fact is that every commutative C*-algebra is of exactly this form. So a commutative C*-algebra is a space of functions on a shape in disguise.

The payoff is the functional calculus: once you fix a single well-behaved operator, you may apply any continuous function to it and get a new operator, with all the algebra working as if the operator were just a number.

Visual Beginner

A normal operator $a$ has a spectrum — a compact set of complex numbers attached to it. The functional calculus matches continuous functions on that set with operators built from $a$, and the match preserves every operation.

The dictionary reads: the function equal to its input becomes the operator $a$; the constant function $1$ becomes the identity; multiplying functions matches multiplying operators; and conjugating a function matches starring its operator.

Worked example Beginner

Take the simplest operators: two-by-two complex matrices, with the star equal to the conjugate transpose. Consider the matrix that scales the first axis by $3$ and the second axis by $5$:

$$a=\left(\begin{array}{cc}3& 0\\ 0& 5\end{array}\right).$$

Its star is the conjugate transpose, which is the same matrix since the entries are real. The spectrum of $a$ is the set of its diagonal values, $\{3,5\}$.

Now apply a continuous function, say the square-root function $f(x)=\sqrt{x}$, to the two spectral values: $f(3)=\sqrt{3}$ and $f(5)=\sqrt{5}$. The functional calculus says $f(a)$ is the matrix that scales the axes by these new numbers:

$$f(a)=\left(\begin{array}{cc}\sqrt{3}& 0\\ 0& \sqrt{5}\end{array}\right).$$

Check: $f(a)$ times $f(a)$ gives the matrix scaling the axes by $3$ and $5$, which is $a$ again. So the square root of $a$ squares back to $a$, exactly as a number's square root would.

What this tells us: applying a function to an operator means applying it to each spectral value, and every numerical identity the function satisfies carries over to the operators.

Check your understanding Beginner

Formal definition Intermediate+

A Banach algebra is a complex Banach space 02.11.04 $A$ equipped with an associative bilinear multiplication satisfying $\Vert ab\Vert \le \Vert a\Vert \Vert b\Vert$. It is unital if it has a multiplicative identity $1$ with $\Vert 1\Vert =1$. An involution on $A$ is a conjugate-linear map $a\mapsto a^{\ast }$ with $(ab{)}^{\ast }=b^{\ast }{a}^{\ast }$ and $a^{\ast \ast }=a$.

A C-algebra* is a Banach algebra with an involution satisfying the C-axiom* $$ |a^* a| = |a|^2 \qquad (a \in A). $$ The axiom forces $\Vert a^{\ast }\Vert =\Vert a\Vert$ and is far stronger than the inequality $\Vert a^{\ast }a\Vert \le \Vert a{\Vert }^2$ available in any Banach $\ast$-algebra; it pins the norm to the algebra uniquely ^{[Murphy Ch. 2]}. Canonical examples: the bounded operators $B(H)$ on a Hilbert space with the adjoint as involution 02.11.01; the continuous functions $C_0(X)$ vanishing at infinity on a locally compact Hausdorff space $X$, with pointwise operations and complex conjugation; the compact operators $K(H)\subseteq B(H)$; and the matrix algebras $M_n(\mathbb{C})$.

For an element $a$ of a unital algebra $A$, the spectrum is $$ \sigma(a) = {\lambda \in \mathbb{C} : a - \lambda 1 \text{ is not invertible in } A}, $$ a non-empty compact subset of $\mathbb{C}$. The spectral radius is $r(a)=\sup \{|\lambda |:\lambda \in \sigma (a)\}$, and the spectral radius formula gives $r(a)={\lim }_{n\to \infty }\Vert a^n{\Vert }^{1/n}$. An element is self-adjoint if $a=a^{\ast }$, normal if $a^{\ast }a=a{a}^{\ast }$, unitary if $a^{\ast }a=a{a}^{\ast }=1$, and positive (written $a\ge 0$) if $a=a^{\ast }$ and $\sigma (a)\subseteq [0,\infty )$.

Counterexamples to common slips

• The C*-axiom is not the submultiplicative bound. The disc algebra $A(\mathbb{D})$ of functions holomorphic on the open disc and continuous on its closure, with the supremum norm and complex conjugation as involution, is a commutative Banach $\ast$-algebra but is not a C*-algebra: $z\mapsto \bar{z}$ does not map holomorphic functions to holomorphic functions, so the involution is not even well-defined inside the algebra, and any attempted norm fails $\Vert a^{\ast }a\Vert =\Vert a{\Vert }^2$.
• The spectrum is algebra-dependent in general but stable for C*-algebras. If $B\subseteq A$ is a unital C*-subalgebra and $a\in B$, then ${\sigma }_B(a)={\sigma }_A(a)$; the spectrum does not shrink on passing to a larger C*-algebra. This spectral permanence fails for general Banach subalgebras.
• Positivity needs self-adjointness. The matrix $\left(\begin{array}{cc}0& 1\\ 0& 0\end{array}\right)$ has spectrum $\{0\}\subseteq [0,\infty )$ but is not positive, because it is not self-adjoint. The spectral condition alone is insufficient.

Key theorem with proof Intermediate+

Theorem (continuous functional calculus for a normal element). Let $A$ be a unital C-algebra and $a\in A$ normal. There is a unique unital $\ast$-homomorphism* $$ \Phi : C(\sigma(a)) \longrightarrow A, \qquad \Phi(f) = f(a), $$ sending the identity function ${\mathrm{id}}_{\sigma (a)}$ to $a$ and the constant $1$ to $1$. The map $\Phi$ is an isometric $$-isomorphismontotheunitalC\ast -subalgebra$C^(a,1)$generatedby$a$,anditsatisfiesthespectralmappingtheorem$\sigma(f(a)) = f(\sigma(a))$. ^{[Davidson Ch. I; Murphy Ch. 2]}

Proof. Let $B=C^{\ast }(a,1)$, the closed unital $\ast$-subalgebra generated by $a$. Since $a$ is normal and $a,a^{\ast },1$ commute, $B$ is commutative. Apply the Gelfand theory of commutative unital C*-algebras: the character space $\Omega (B)$ of nonzero $\ast$-homomorphisms $\tau :B\to \mathbb{C}$ is a compact Hausdorff space in the weak-$\ast$ topology, and the Gelfand transform $$ \Gamma : B \to C(\Omega(B)), \qquad \Gamma(b)(\tau) = \tau(b), $$ is an isometric $\ast$-isomorphism. (Isometry uses the C*-axiom: for self-adjoint $b$, $\Vert \Gamma (b){\Vert }_{\infty }=r(b)=\Vert b\Vert$ by the spectral radius formula, and the general case follows from $\Vert b{\Vert }^2=\Vert b^{\ast }b\Vert =\Vert \Gamma (b^{\ast }b){\Vert }_{\infty }=\Vert \Gamma (b){\Vert }_{\infty }^2$.)

The evaluation $\tau \mapsto \tau (a)$ is a continuous map $\hat{a}:\Omega (B)\to \mathbb{C}$ with image ${\sigma }_B(a)={\sigma }_A(a)=\sigma (a)$ (spectral permanence). Because $a$ and $1$ generate $B$, a character is determined by its value on $a$, so $\hat{a}$ is injective; being a continuous injection from a compact space to a Hausdorff space, it is a homeomorphism $\Omega (B)\cong \sigma (a)$. Composition with ${\Gamma }^{-1}$ yields $$ \Phi : C(\sigma(a)) \xrightarrow{\ \hat a^\ } C(\Omega(B)) \xrightarrow{\ \Gamma^{-1}\ } B, $$ where $\hat a^(f) = f \circ \hat a$.This$\Phi$isanisometric$$-isomorphismonto$B = C^(a,1)$,sends$\mathrm{id}_{\sigma(a)}$to$a$and$1$to$1$.

Uniqueness: any unital $\ast$-homomorphism agreeing with $\Phi$ on $\mathrm{id}$ and $1$ agrees on all polynomials in $z$ and $\bar{z}$; these are dense in $C(\sigma (a))$ by Stone-Weierstrass, and a $\ast$-homomorphism between C*-algebras is norm-decreasing, hence continuous, so the two maps coincide. The spectral mapping theorem follows because $\Phi$ is an isometric isomorphism: $\sigma (f(a))={\sigma }_{C(\sigma (a))}(f)=f(\sigma (a))$, the range of $f$. $\square$

Bridge. The functional calculus builds toward the entire structure theory of C*-algebras, and it appears again in 39.03.01 (von Neumann algebras and the double commutant) where the continuous calculus is widened to a Borel calculus against a spectral measure. The foundational reason it works is exactly the C*-axiom: it forces the Gelfand transform to be isometric, so the commutative subalgebra is genuinely an algebra of functions and nothing is lost. This is exactly the abstract face of the finite-dimensional spectral decomposition $T={\sum }_i{\lambda }_i{P}_i$ from 01.01.13, where the calculus $f(T)={\sum }_{i}f({\lambda }_i)P_i$ is the same dictionary with a finite spectrum. Putting these together, the single normal element generates a copy of $C(\sigma (a))$, and the calculus generalises the Hilbert-space Borel calculus of 02.11.03 downward to the purely algebraic level, where the bridge is that "operator" and "continuous function on the spectrum" are two names for one object.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: none — Mathlib carries the constituent pieces (the CStarRing class, the spectrum, the spectral-radius formula, and cfc/cfcHom for the continuous functional calculus), but the Gelfand-theory derivation packaged as a single named theorem — character space, Gelfand transform isometry, and the resulting $C(\sigma (a))\cong C^{\ast }(a,1)$ with spectral mapping — is not consolidated.

The intended formalisation reads schematically:

import Mathlib.Analysis.CStarAlgebra.ContinuousFunctionalCalculus.Basic
import Mathlib.Analysis.CStarAlgebra.GelfandDuality

variable {A : Type*} [CStarAlgebra A] (a : A)

/-- Continuous functional calculus for a normal element: a unique unital
isometric *-isomorphism from C(σ a) onto the generated C*-subalgebra. -/
theorem cfc_normal (ha : IsStarNormal a) :
    ∃ Φ : C(spectrum ℂ a, ℂ) →⋆ₐ[ℂ] A,
      Function.Injective Φ ∧
      (∀ f, ‖Φ f‖ = ‖f‖) ∧
      Φ (ContinuousMap.restrict _ (.id ℂ)) = a :=
  sorry  -- Gelfand transform isometry + character-space homeomorphism

Advanced results Master

The structure theorems below organise the basic theory beyond a single element.

Positivity and order. In a C*-algebra the positive elements form a closed cone $A_{+}=\{a^{\ast }a:a\in A\}$; equivalently $a\ge 0$ iff $a=a^{\ast }$ and $\sigma (a)\subseteq [0,\infty )$, and the two descriptions agree. The cone induces a partial order $a\le b⟺b-a\ge 0$ compatible with the algebra: $0\le a\le b$ implies $\Vert a\Vert \le \Vert b\Vert$, and $c^{\ast }ac\ge 0$ whenever $a\ge 0$. The identity $\sigma (a^{\ast }a)\setminus \{0\}=\sigma (a{a}^{\ast })\setminus \{0\}$ shows that the cone is independent of left versus right.

Approximate units. A C*-algebra $A$ need not be unital ($K(H)$ is not, nor is $C_0(X)$ for non-compact $X$), but it always has an approximate unit: an increasing net $(u_{\lambda })$ in $A_{+}$ with $\Vert u_{\lambda }\Vert \le 1$ and $\Vert a{u}_{\lambda }-a\Vert \to 0$, $\Vert u_{\lambda }a-a\Vert \to 0$ for every $a$. One construction takes $u_{\lambda }$ to range over the positive elements of norm $<1$ in the order, using the functional calculus to verify the limit on a dense set. Approximate units are the technical engine behind every result that "reduces to the unital case".

Ideals and quotients. A closed two-sided ideal $I$ of a C*-algebra is automatically self-adjoint ($I=I^{\ast }$), so $I$ is itself a C*-algebra, and the quotient $A/I$ with the quotient norm $\Vert a+I\Vert ={\inf }_{x\in I}\Vert a+x\Vert$ is again a C*-algebra. The proof uses an approximate unit of the ideal to show the quotient norm satisfies the C*-axiom. In $B(H)$ the compact operators $K(H)$ form a closed ideal; the quotient $B(H)/K(H)$ is the Calkin algebra, the natural home of Fredholm theory.

Gelfand-Naimark, commutative form. The functor $X\mapsto C_0(X)$ is an equivalence between locally compact Hausdorff spaces (with proper maps) and commutative C*-algebras (with $\ast$-homomorphisms), the unital case corresponding to compact $X$. Every commutative C*-algebra is $C_0$ of its character space. This is the precise sense in which a commutative C*-algebra is a topological space, and it is the launching point for noncommutative geometry, where a noncommutative C*-algebra is treated as functions on a "noncommutative space".

Synthesis. The C*-axiom is the foundational reason every structural statement above holds: it makes the Gelfand transform isometric, and this is exactly what converts the abstract algebra of a single normal element into the concrete algebra $C(\sigma (a))$. The continuous functional calculus is dual to the Gelfand transform — one builds operators from functions, the other reads functions off operators — and putting these together gives the central insight that spectrum, character space, and the generated commutative subalgebra are three names for one compact Hausdorff space. This generalises in two directions that appear again throughout operator algebras: downward to the order theory, where positivity $a^{\ast }a\ge 0$ and approximate units let every non-unital question reduce to the unital functional calculus, and upward to von Neumann algebras 39.03.01, where the continuous calculus is widened to a Borel calculus and the commutative model $C(X)$ is replaced by $L^{\infty }(X,\mu )$. The bridge to the rest of the subject is that the commutative Gelfand-Naimark equivalence reframes a noncommutative C*-algebra as functions on a noncommutative space, and the whole of noncommutative geometry 39.06.01 is the systematic exploitation of that reframing.

Full proof set Master

Proposition (the C-axiom forces $\Vert a^{\ast }\Vert =\Vert a\Vert$ and $\Vert 1\Vert =1$ in a unital C-algebra).** From $\Vert a{\Vert }^2=\Vert a^{\ast }a\Vert \le \Vert a^{\ast }\Vert \Vert a\Vert$ one gets $\Vert a\Vert \le \Vert a^{\ast }\Vert$; applying the same to $a^{\ast }$ and using $a^{\ast \ast }=a$ gives $\Vert a^{\ast }\Vert \le \Vert a\Vert$, hence equality. For the unit, $\Vert 1\Vert =\Vert 1^{\ast }1\Vert /\Vert 1\Vert$ is not directly available, so argue: $\Vert 1{\Vert }^2=\Vert 1^{\ast }1\Vert =\Vert 1\Vert$, so $\Vert 1\Vert \in \{0,1\}$, and since $1\ne 0$ in a unital algebra acting nontrivially, $\Vert 1\Vert =1$.

Proposition (spectral radius formula in a Banach algebra). For $a$ in a unital Banach algebra, $r(a)={\lim }_{n\to \infty }\Vert a^n{\Vert }^{1/n}$. The limit exists by submultiplicativity (Fekete's lemma applied to $\log \Vert a^n\Vert$). The resolvent $\lambda \mapsto (a-\lambda 1{)}^{-1}$ is analytic on $\{|\lambda |>r(a)\}$ with Laurent expansion $-{\sum }_{n\ge 0}{a}^n{\lambda }^{-n-1}$, convergent precisely for $|\lambda |>{\mathrm{limsup}}_n\Vert a^n{\Vert }^{1/n}$; comparing the two radii of convergence forces $r(a)={\mathrm{limsup}}_n\Vert a^n{\Vert }^{1/n}$, and Fekete upgrades the $\mathrm{limsup}$ to a limit.

Proposition (Gelfand-Mazur). A unital Banach algebra in which every nonzero element is invertible (a division algebra) is isometrically isomorphic to $\mathbb{C}$. For $a\in A$, the spectrum $\sigma (a)$ is non-empty (else the entire-bounded resolvent would be constant zero by Liouville, contradicting $(a-\lambda 1{)}^{-1}\to 0$), so pick $\lambda \in \sigma (a)$. Then $a-\lambda 1$ is not invertible, and by the division hypothesis $a-\lambda 1=0$, i.e. $a=\lambda 1$. The map $\lambda 1\mapsto \lambda$ is the required isomorphism, isometric because $\Vert \lambda 1\Vert =|\lambda |$.

Proposition (a $\ast$-homomorphism between C*-algebras is norm-decreasing). Let $\pi :A\to B$ be a $\ast$-homomorphism of C*-algebras, $A$ unital. For self-adjoint $a$, ${\sigma }_B(\pi (a))\subseteq {\sigma }_A(a)$ since $\pi$ preserves invertibility relations, so $\Vert \pi (a)\Vert =r(\pi (a))\le r(a)=\Vert a\Vert$ using the C*-norm-equals-spectral-radius identity for self-adjoint elements. For general $b$, $\Vert \pi (b){\Vert }^2=\Vert \pi (b{)}^{\ast }\pi (b)\Vert =\Vert \pi (b^{\ast }b)\Vert \le \Vert b^{\ast }b\Vert =\Vert b{\Vert }^2$. Hence $\Vert \pi (b)\Vert \le \Vert b\Vert$; equality holds when $\pi$ is injective, making an injective $\ast$-homomorphism isometric.

Proposition (the quotient norm satisfies the C-axiom).* Let $I$ be a closed two-sided ideal of a C*-algebra $A$ with approximate unit $(u_{\lambda })$ for $I$. For $a\in A$ the quotient norm is $\Vert a+I\Vert ={\lim }_{\lambda }\Vert a-a{u}_{\lambda }\Vert$. A computation with this representation yields $\Vert a+I{\Vert }^2={\lim }_{\lambda }\Vert (a-a{u}_{\lambda }{)}^{\ast }(a-a{u}_{\lambda })\Vert =\Vert a^{\ast }a+I\Vert \le \Vert a+I\Vert \Vert a^{\ast }+I\Vert$, and the reverse inequality $\Vert a^{\ast }a+I\Vert \le \Vert a+I{\Vert }^2$ is automatic, giving $\Vert a^{\ast }a+I\Vert =\Vert a+I{\Vert }^2$. So $A/I$ is a C*-algebra.

Connections Master

• Spectral theorem for normal operators 01.01.13 — the finite-dimensional decomposition $T={\sum }_i{\lambda }_i{P}_i$ with calculus $f(T)={\sum }_{i}f({\lambda }_i)P_i$ is the continuous functional calculus over a finite spectrum; this unit is the infinite-dimensional, coordinate-free abstraction.

• Unbounded self-adjoint operators 02.11.03 — the Hilbert-space spectral measure and Borel calculus extend the continuous calculus from $C(\sigma (a))$ to bounded Borel functions; this unit supplies the bounded-element foundation that the unbounded theory completes through resolvents.

• Bounded linear operators 02.11.01 — $B(H)$ is the universal example of a C*-algebra, and every C*-algebra embeds into some $B(H)$ by the Gelfand-Naimark-Segal construction; the operator-norm and adjoint structure of this unit live there first.

• Von Neumann algebras 39.03.01 — strengthening norm-closure to weak-operator closure turns the continuous calculus into a Borel calculus and the commutative model $C(X)$ into $L^{\infty }(X,\mu )$; the double commutant theorem is the next structural layer.

• Spectral triples and noncommutative geometry 39.06.01 — the commutative Gelfand-Naimark equivalence $C_0(X)\leftrightarrow X$ is the entry point: a noncommutative C*-algebra is read as functions on a noncommutative space, equipped with a Dirac operator to recover metric data.

Historical & philosophical context Master

The abstract definition originates with Gelfand and Naimark, who proved in 1943 that any abstract C*-algebra (then called a $B^{\ast }$-algebra, defined by the norm axiom $\Vert a^{\ast }a\Vert =\Vert a{\Vert }^2$) is isometrically $\ast$-isomorphic to a norm-closed $\ast$-subalgebra of $B(H)$ ^{[Gelfand-Naimark 1943]}. Their commutative theorem, identifying commutative C*-algebras with algebras of continuous functions, grew out of Gelfand's earlier theory of normed rings and their maximal-ideal spaces. The original 1943 paper assumed an extra axiom ($1+a^{\ast }a$ invertible) later shown by Kaplansky and others to be redundant, leaving the single norm identity as the definition.

Segal introduced the term "C*-algebra" and, with the GNS construction, connected the abstract algebra to representations on Hilbert space, supplying the bridge from the Gelfand commutative picture to the full noncommutative theory. The subject was driven from the start by quantum mechanics: von Neumann's operator-ring program sought an algebraic substrate for observables, and the C*-axiom is the precise condition under which the spectrum of an observable and the norm of its measurement statistics are compatibly constrained. Connes later turned the commutative Gelfand-Naimark duality into a definition of noncommutative geometry, where the C*-algebra replaces the space entirely.

Bibliography Master

• Davidson, K. R., C-Algebras by Example*, Fields Institute Monographs 6, American Mathematical Society, 1996. Ch. I.
• Murphy, G. J., C-Algebras and Operator Theory*, Academic Press, 1990. Ch. 1–2.
• Gelfand, I. and Naimark, M., "On the imbedding of normed rings into the ring of operators in Hilbert space", Mat. Sbornik 12 (1943), 197–213.
• Dixmier, J., C-Algebras*, North-Holland, 1977. §1.
• Takesaki, M., Theory of Operator Algebras I, Springer, 1979. Ch. I.
• Segal, I. E., "Irreducible representations of operator algebras", Bulletin of the American Mathematical Society 53 (1947), 73–88.

---

Operator-algebras spine, foundational C-algebra unit. Produced as the abstract continuous-functional-calculus anchor distinct from the finite-dim spectral theorem (01.01.13) and the Hilbert-space Borel calculus (02.11.03).*