Abstract C*-calculus and order for finite weight domains

Classical mathematical exposition by GPT-6 Astra (OpenAI), Ultra, October 2026. New prose: CC0.

The finite-domain construction applies to abstract C*-algebras before a faithful Hilbert-space representation has been constructed. This proof therefore starts with a complete complex normed *-algebra satisfying the C*-identity. It supplies the self-adjoint continuous calculus, positive square roots and order inequalities used in the weight lessons, without assuming a representation theorem.

The free comparison is Bruce Blackadar's corrected author edition of Operator Algebras, II.1.6.1–3, II.2.3.1–2 and II.3.1.1–5 63–66. All necessary arguments are written below. The exact earlier analytic inputs are the real Hahn–Banach extension proof NP1 and scalar Cauchy estimates and Liouville's theorem, Theorem 3.2 and Corollary 3.3. The Banach and C*-algebra axioms and scalar completeness are entry assumptions. The zero algebra is treated separately whenever a nonempty spectrum is asserted.

AC1 — Banach-algebra spectral facts from scalar analysis

Let $B\ne\{0\}$ be a unital complex Banach algebra with $\|1\|=1$. Absolute convergence implies convergence of a series in $B$: its partial sums are Cauchy because each tail norm is bounded by the scalar sum of the tail norms. Thus, when $\|v\|<1$, multiplying finite geometric sums and taking limits proves that $\sum_{n\geq0}v^n$ is a two-sided inverse of $1-v$. Factoring a perturbation of an invertible element through it proves openness of the invertible set. The same geometric series gives the local power series of the resolvent $R(z)=(z1-b)^{-1}.$ In particular it is norm-continuous and locally analytic on its domain. For $|z|>\|b\|$, the geometric series also gives $\|R(z)\|\leq\frac1{|z|-\|b\|}.$ Consequently $\sigma_B(b)$, the complement of its domain, is compact and contained in the closed disc of radius $\|b\|$.

Here is the separation fact needed to use scalar complex analysis. On any complex normed space and for any vector $v\ne0$, real Hahn–Banach extends the functional $tv\mapsto t\|v\|$, dominated by the norm, to a real-linear $f$. Set $L(x)=f(x)-if(ix)$. This is complex-linear, and $\operatorname{Re}L=f$. Rotating $x$ by a scalar of modulus one gives $|L(x)|\leq\|x\|$. Since $f(v)=\|v\|$, this also forces $L(v)=\|v\|$. Thus bounded complex functionals separate points and $\|v\|=\sup_{\|L\|\leq1}|L(v)|.$ If the spectrum of $b$ were empty, every $L(R(z))$ would be entire. It is bounded outside a disc by the resolvent bound, and inside that disc by continuity and compactness. Liouville's theorem makes it constant; the bound at infinity makes that constant zero. Separation would give $R(z)=0$, impossible for an inverse of the nonzero identity. Hence the spectrum is nonempty.

We also need polynomial spectral mapping, including both directions. Every nonconstant complex polynomial has a root: otherwise its reciprocal would be entire and bounded, since the leading term gives growth at infinity and its reciprocal is bounded on compact sets. Liouville would make that reciprocal constant. Repeated division by linear factors proves complete factorization. For a polynomial $p$, factor $p(t)-\mu$. Its factors evaluated at $b$ commute. A product of commuting elements is invertible exactly when each factor is invertible: if $uv=vu$ is invertible, its inverse commutes with $u,v$, and $v(uv)^{-1}$ is a two-sided inverse of $u$. Induct on the number of factors. Constant polynomials are immediate. This proves $\sigma_B(p(b))=p(\sigma_B(b)).$

Put $r=\max_{\lambda\in\sigma_B(b)}|\lambda|$. Polynomial mapping and the spectral norm bound give $r\leq\|b^n\|^{1/n}$. For the other inequality, the function $G(z)=(1-zb)^{-1}$ is analytic on $|z|<1/r$, with all of $\mathbb C$ understood when $r=0$. Near zero its geometric expansion has coefficients $b^n$. On a circle of radius $s<1/r$, let $M_s=\max_{|z|=s}\|G(z)\|$. The scalar Cauchy estimate, applied to every norm-one functional and then to the preceding norming identity, gives $\|b^n\|\leq M_s s^{-n}.$ Taking upper limits of the $n$-th roots and then increasing $s$ to $1/r$ proves $\lim_{n\to\infty}\|b^n\|^{1/n}=r.$ When $r=0$, let $s$ be arbitrarily large. This proof uses only scalar Cauchy estimates, not an unproved Banach-valued integration theorem.

Finally, $ab$ and $ba$ have the same nonzero spectrum. For example, multiplication verifies that $(1-ab)^{-1}=1+a(1-ba)^{-1}b$ whenever the inverse on the right exists. Interchanging $a,b$ gives the converse, and scaling treats every nonzero spectral parameter.

AC2 — Self-adjoint spectra and the polynomial norm

Now let $B\ne\{0\}$ be a unital C*-algebra. Its involution is isometric: the C*-identity and submultiplicativity give $\|x\|\leq\|x^*\|$ when $x\ne0$, and applying the same argument to $x^*$ gives equality. The identity has norm one, since $\|1\|^2=\|1\|$, and a unitary also has norm one.

For self-adjoint $h$, the absolutely convergent exponential series defines $e^{ith}$ for real $t$. Multiplication of the two absolutely convergent series for $e^{ith}$ and $e^{-ith}$, using the finite binomial identity at each total degree, gives the identity. Taking adjoints term by term proves that these elements are unitary.

If $\lambda\in\sigma_B(h)$, then $e^{it\lambda}\in\sigma_B(e^{ith})$. To see the inclusion without invoking holomorphic spectral mapping, subtract the scalar exponential series and factor every power difference by $h-\lambda$. The resulting quotient series converges absolutely, because the quotient of the $n$-th power has norm at most $n\max(\|h\|,|\lambda|)^{n-1}$. Thus $e^{ith}-e^{it\lambda}1=(h-\lambda1)g(h),$ where the factors commute. Invertibility of the left side would force invertibility of $h-\lambda1$, a contradiction. The norm-one bound on the unitary now gives $|e^{it\lambda}|\leq1$ for every real $t$; taking positive and negative $t$ forces $\operatorname{Im}\lambda=0$.

Repeated use of the C*-identity gives $\|h^{2^k}\|=\|h\|^{2^k}$. AC1 therefore implies $r(h)=\|h\|$. If $x$ is normal, then $\begin{aligned} \|x^{2^k}\|^2 &=\|(x^*x)^{2^k}\|\\ &=\|x^*x\|^{2^k} =\|x\|^{2^{k+1}}. \end{aligned}$ So AC1 also gives $r(x)=\|x\|$. Every polynomial in a self-adjoint $h$ is normal. Combining this fact with polynomial spectral mapping proves $\|p(h)\|=\max_{t\in\sigma_B(h)}|p(t)|.$ In particular the spectrum of $h$ is a nonempty compact subset of $[-\|h\|,\|h\|]$.

AC3 — Continuous calculus on the actual spectrum

Write $S=\sigma_B(h)$. Continuous functions on a compact real interval admit uniform polynomial approximation. For clarity, after rescaling to $[0,1]$, the Bernstein polynomial of a continuous $f$ is $\begin{aligned} &B_nf(t)\\ &\quad=\sum_{k=0}^n f(k/n){n\choose k}t^k(1-t)^{n-k}. \end{aligned}$ The weights sum to one, their mean is $t$, and their variance is $t(1-t)/n$, by differentiating the finite binomial formula. The sum of weights at distance at least $\delta$ from $t$ is at most $1/(4n\delta^2)$. Uniform continuity consequently gives $\|B_nf-f\|_\infty \leq \omega_f(\delta) +\frac{\|f\|_\infty}{2n\delta^2},$ where $\omega_f(\delta)$ is the supremum of $|f(s)-f(t)|$ over distances at most $\delta$. First let $n$ increase, then let $\delta$ decrease. Complex functions obey the same estimate.

A continuous function on $S$ extends to its containing interval with the same norm: interpolate linearly on each complementary interval and keep the endpoint value constant outside the extreme points of $S$. Continuity at points of $S$ follows from uniform continuity on $S$; small gaps have close endpoints, and only finitely many gaps exceed any fixed positive length. A singleton $S$ admits the constant extension. Thus polynomials restricted to $S$ are dense in $C(S)$.

For polynomials $p_n\to f$ uniformly on $S$, AC2 makes $p_n(h)$ Cauchy. Define $f(h)$ as its norm limit. AC2 proves independence of the approximants and $\|f(h)\|=\|f\|_{C(S)}.$ Passing polynomial sums, products and adjoints to the limit shows that this is a unital isometric *-homomorphism onto the norm-closed algebra generated by $1,h$. The range is closed by completeness; it is exactly that algebra because it contains the polynomials and consists of their limits. It commutes with every element commuting with $h$.

Its spectral mapping holds in the full algebra $B$. If $\mu\notin f(S)$, apply the calculus to $1/(f-\mu)$ to construct an inverse of $f(h)-\mu1$. Conversely, let $f(t_0)=\mu$. If that element had an inverse of norm $C$, choose a continuous function $g$ on $S$, of norm one and equal to one at $t_0$, supported where $|f-\mu|<\varepsilon$. A small triangular distance bump gives such a function. Then $\begin{aligned} 1=\|g(h)\| &\leq C\|(f(h)-\mu1)g(h)\|\\ &\leq C\varepsilon, \end{aligned}$ which is impossible for $\varepsilon<1/C$. Therefore $\sigma_B(f(h))=f(S)$.

For real-valued $f$, the element $f(h)$ is self-adjoint. If $g$ is continuous on $f(S)$, polynomial approximation to $g$ and multiplicativity give $g(f(h))=(g\circ f)(h)$. This proves the composition rule used for square roots and positive parts.

AC4 — Squares, the positive cone and its norm bounds

Temporarily call a self-adjoint element spectrally positive when its spectrum is nonnegative. For self-adjoint $h$ and $t\geq\|h\|$, the norm identity in AC2 shows that spectral positivity is equivalent to $\|t1-h\|\leq t$. This criterion proves closure under addition: use $t=\|a\|+\|b\|$ and the triangle inequality for $t1-a-b$. Positive scalar multiplication follows by spectral mapping. For a convergent sequence of spectrally positive elements, choose one common norm bound $t$ and pass the criterion to the limit. This proves norm closure. The cone is proper because a self-adjoint element whose spectrum is $\{0\}$ has norm zero. Squares of self-adjoint elements belong to the cone by polynomial spectral mapping.

We now prove that $y^*y$ is spectrally positive for arbitrary $y$; it must not be assumed from a Hilbert-space representation. Put $c=y^*y$, and use AC3 to write $c=c_+-c_-$, with $c_\pm$ spectrally positive, commuting, and $c_+c_-=0$. Set $w=yc_-$. Then $w^*w=-c_-^3.$ Writing $w=k+il$ with $k,l$ self-adjoint gives $ww^*=2k^2+2l^2+c_-^3.$ The right side is spectrally positive by the cone properties just proved. AC1 gives equality of the nonzero spectra of $ww^*$ and $w^*w$. Thus $w^*w$ is spectrally positive as well, while its negative $c_-^3$ is spectrally positive. Properness gives $c_-^3=0$; spectral mapping and the norm formula give $c_-=0$. Hence $y^*y=c_+$ is spectrally positive.

Conversely, if $a$ is spectrally positive, AC3 defines the positive element $b=\sqrt a$ and gives $b^2=a=b^*b$. Spectral positivity therefore agrees exactly with being of the form $y^*y$. We denote this cone by $B_+$. Its square root is unique: if $d\geq0$ and $d^2=a$, the composition rule for the self-adjoint calculus of $d$ gives $\sqrt a=\sqrt{d^2}=d$. Polynomial approximation also proves that the square root commutes with every element commuting with $a$.

Conjugation preserves positivity, since $z^*az=(a^{1/2}z)^*(a^{1/2}z).$ For self-adjoint $h$, spectral mapping gives $h\leq t1$ exactly when $\sigma_B(h)\subseteq(-\infty,t]$. In particular $\begin{gathered} -\|h\|1\leq h\leq\|h\|1,\\ x^*x\leq\|x\|^2 1. \end{gathered}$ If $0\leq a\leq b$, then $a\leq\|b\|1$, and the spectral norm formula yields $\|a\|\leq\|b\|$. The continuous functions $t\mapsto\max(t,0)$ and $t\mapsto\max(-t,0)$ give the positive and negative parts of any self-adjoint element. Taking real and imaginary parts shows that every element is a linear combination of four positive elements.

All the preceding order statements concern an arbitrary unital C*-algebra, not merely an operator algebra. For a nonunital algebra, UZ02–05 construct its forced C*-unitization directly from the axioms, without using AC1–4. UZ06–07 then apply this calculus and the scalar quotient to prove the inherited cone and square roots in the original ideal. This order of proof avoids a representation or unitization circle. For the zero algebra all order and square-root assertions reduce to zero; the nonempty-spectrum argument is not applied to it.