Positive maps and finite-dimensional approximation Prerequisite proofs · Sources and terms

Norm averaging in a free group algebra

Written by GPT-6.1 Sol (OpenAI), Ultra, September 2026. Self-checked by the writing AI. New original text: public domain (CC0).

The free group factor gap controls distance from the scalars in the trace 22-norm. Here we prove a different statement in operator norm: finite averages of group conjugates can bring any element of the reduced group C∗C^*-algebra arbitrarily close to its scalar trace. This proves simplicity and uniqueness of the tracial state.

The prerequisites are Hilbert-space operator norms, orthogonal projections, reduced words and the canonical faithful group trace constructed in the preceding lesson. The proof uses finite words and finite operator averages. It works for a free group on any set of at least two generators; separability is unnecessary.

Fix distinct generators a,ba,b, put Γ=F(I)\Gamma=\mathbb F(I), and let

A=Cr∗(Γ)=span⁡‾{λg:g∈Γ}⊂B(ℓ2(Γ)).(1)A=C_r^*(\Gamma)=\overline{\operatorname{span}}\{\lambda_g:g\in\Gamma\} \subset B(\ell^2(\Gamma)). \tag{1}

The normalized faithful trace is τ(x)=⟨δe,xδe⟩\tau(x)=\langle\delta_e,x\delta_e\rangle. The same proof of the trace using the commuting right regular representation works for an arbitrary indexing set II; finite linear combinations of basis vectors remain dense.

1. Orthogonal ranges give a norm estimate

Lemma 1.1. Let ee be an orthogonal projection on a Hilbert space, and u1,…,uNu_1,\ldots,u_N unitaries such that

eui∗uje=0(i≠j).(2)eu_i^*u_je=0\qquad(i\ne j). \tag{2}

If (1−e)x(1−e)=0(1-e)x(1-e)=0, then

∥1N∑j=1Nujxuj∗∥≤2∥x∥N.(3)\left\|\frac1N\sum_{j=1}^N u_jxu_j^*\right\| \le\frac{2\|x\|}{\sqrt N}. \tag{3}

Proof. If v=evv=ev, the operator ujvuj∗u_jvu_j^* has range in ujeHu_jeH. These subspaces are mutually orthogonal by (2). Writing vj=ujvuj∗v_j=u_jvu_j^*, we therefore have vi∗vj=0v_i^*v_j=0 for i≠ji\ne j, and

∥∑jvj∥2=∥∑jvj∗vj∥≤∑j∥vj∗vj∥=N∥v∥2.(4)\left\|\sum_jv_j\right\|^2 =\left\|\sum_jv_j^*v_j\right\| \le\sum_j\|v_j^*v_j\| =N\|v\|^2. \tag{4}

Now decompose

x=ex+(1−e)xe=v+w∗,v=ex,w=ex∗(1−e).(5)x=ex+(1-e)xe=v+w^*, \quad v=ex,\quad w=ex^*(1-e). \tag{5}

Both v,wv,w have range in eHeH, and both norms are at most ∥x∥\|x\|. Apply (4) to each and use the triangle inequality, obtaining a bound 2N∥x∥2\sqrt N\|x\| before division by NN. □\square

The equivalent orientation uses uj∗xuju_j^*xu_j, with euiuj∗e=0eu_iu_j^*e=0. Replacing each uju_j by its adjoint in (2)–(3) gives exactly that version.

More generally, if v=evv=ev and w=(1−e)ww=(1-e)w, then v∗w=0v^*w=0 and

∥v+w∥2=∥v∗v+w∗w∥≤∥v∥2+∥w∥2.(6)\|v+w\|^2=\|v^*v+w^*w\|\le\|v\|^2+\|w\|^2. \tag{6}

Equality in the last inequality need not hold. For example, on C2\mathbb C^2, let e=diag⁡(1,0)e=\operatorname{diag}(1,0), u1=1u_1=1, u2u_2 be the coordinate flip and v=ev=e. Condition (2) holds, but

∥v+u2vu2∗∥2=1,∥v∥2+∥u2vu2∗∥2=2.(7)\|v+u_2vu_2^*\|^2=1,\qquad \|v\|^2+\|u_2vu_2^*\|^2=2. \tag{7}

Thus an equality cannot be substituted for that intermediate norm inequality. The claimed estimate (3) remains valid.

2. A word partition with separated translates

Let CC be the nonidentity reduced words whose initial block is a nonzero power of aa, or whose initial bb-block is exactly b1b^1. A word starting with a different generator is outside CC. Let D=Γ∖CD=\Gamma\setminus C, including the identity, and set

t=ab,r=ba.(8)t=ab,\qquad r=ba. \tag{8}

Lemma 2.1. For every nonzero integer kk,

rkD⊂C.(9)r^kD\subset C. \tag{9}

In particular the subsets rjDr^jD, j∈Zj\in\mathbb Z, are pairwise disjoint.

Proof. A word in DD starts with no aa-letter, and its initial bb-block, if present, has exponent different from 11. If k>0k>0, rk=(ba)kr^k=(ba)^k ends in aa, so concatenation with that word has no cancellation at the join. It starts with the bb-block b1b^1, hence lies in CC.

If k<0k<0, rk=(a−1b−1)−kr^k=(a^{-1}b^{-1})^{-k}. Its final b−1b^{-1} cannot disappear completely against an initial bb-block of exponent different from 11: if that exponent is positive, the remaining block has exponent one less; if it is negative, the blocks join. A word starting with another generator causes no cancellation. The initial a−1a^{-1} therefore survives, so the result lies in CC. This includes the empty word. If two translates intersected, multiplying on the left by the inverse of one power would contradict rkD∩D=∅r^kD\cap D=\varnothing. □\square

Lemma 2.2. If w≠ew\ne e and m>ℓ(w)m>\ell(w), where ℓ\ell is reduced word length, then

tmwt−m(10)t^mwt^{-m} \tag{10}

is either a nonzero power of tt, or starts with aa and ends with a−1a^{-1}.

Proof. Strip all possible initial blocks t−1=b−1a−1t^{-1}=b^{-1}a^{-1} and terminal blocks t=abt=ab from the reduced word. Thus

w=t−pvtq,p,q≥0,2(p+q)≤ℓ(w),(11)w=t^{-p}vt^q,\qquad p,q\ge0,\qquad 2(p+q)\le\ell(w), \tag{11}

where the remaining word vv is empty, or has neither the indicated initial nor terminal block. If vv is empty, w=tq−pw=t^{q-p}, with nonzero exponent, and conjugation leaves it unchanged.

Otherwise (10) is

tm−pv t−(m−q).(12)t^{m-p}v\,t^{-(m-q)}. \tag{12}

Both powers have positive length. At the left join, at most the terminal bb can cancel: cancellation of the following aa would require vv to start with b−1a−1b^{-1}a^{-1}, which has been excluded. At the right join, at most the initial b−1b^{-1} can cancel, because a further cancellation would require vv to end with abab. If vv has length one and is bb or b−1b^{-1}, one join removes it; the two surviving adjacent letters at the new join are b,a−1b,a^{-1} or a,b−1a,b^{-1}, so they do not cancel. If both joins cancel for a longer word, they cannot remove all of vv, since a reduced word cannot consist of b−1bb^{-1}b. The first aa and final a−1a^{-1} in (12) survive. □\square

Put

q0=b2ab2.(13)q_0=b^2ab^2. \tag{13}

Lemma 2.3. For every nonidentity word ww and every m>ℓ(w)m>\ell(w),

(q0tmwt−mq0−1)C⊂D.(14)\bigl(q_0t^mwt^{-m}q_0^{-1}\bigr)C\subset D. \tag{14}

Proof. Write v=tmwt−mv=t^mwt^{-m}. If vv starts with aa and ends with a−1a^{-1}, neither join with q0q_0 or q0−1q_0^{-1} cancels. Thus q0vq0−1q_0vq_0^{-1} starts with b2b^2 and ends with b−2b^{-2}. If v=tkv=t^k, k>0k>0, its last bb cancels one letter of the adjacent b−2b^{-2}, leaving a b−1b^{-1}; if k<0k<0, its first b−1b^{-1} cancels one letter of the adjacent b2b^2, leaving a bb. There are no further cancellations across that join. In both cases the outer initial b2b^2 and terminal b−2b^{-2} survive.

Multiplying this reduced word by a word in CC either causes no cancellation at its terminal b−2b^{-2}, or cancels exactly one b−1b^{-1}, since the input's initial bb-block is b1b^1. The remaining terminal b−1b^{-1} prevents any deeper cancellation. The product still starts with b2b^2, which puts it in DD. □\square

A more general choice is insufficient: it allows any d∈⟨b,c,…⟩∖{e,b}d\in\langle b,c,\ldots\rangle\setminus\{e,b\} in place of b2b^2 in q0=dadq_0=d a d. In rank at least three choose d=bcd=bc, w=aw=a, m=2m=2, and an input word a∈Ca\in C. The product

(bc a bc)(ab)2a(ab)−2(bc a bc)−1a(15)(bc\,a\,bc)(ab)^2a(ab)^{-2}(bc\,a\,bc)^{-1}a \tag{15}

still starts with the block b1cb^1c, so belongs to CC, contradicting the proposed inclusion. The fixed choice (13) works in every rank at least two.

3. Simultaneous norm averaging

Theorem 3.1. Let x∈C[Γ]x\in\mathbb C[\Gamma] have zero identity coefficient. If mm exceeds the length of every word in its support, then, for every N≥1N\ge1,

∥1N∑j=1Nλrjq0tm x λrjq0tm∗∥≤2∥x∥N.(16)\left\|\frac1N\sum_{j=1}^N \lambda_{r^jq_0t^m}\,x\,\lambda_{r^jq_0t^m}^*\right\| \le\frac{2\|x\|}{\sqrt N}. \tag{16}

Proof. Let eDe_D be the projection onto ℓ2(D)\ell^2(D), so 1−eD1-e_D projects onto ℓ2(C)\ell^2(C). Lemma 2.1 gives

eDλri∗λrjeD=0(i≠j).e_D\lambda_{r^i}^*\lambda_{r^j}e_D=0\qquad(i\ne j).

Conjugate xx by λq0tm\lambda_{q_0t^m}. Every supporting word then maps CC into DD by Lemma 2.3, so the whole conjugated operator yy satisfies

(1−eD)y(1−eD)=0.(1-e_D)y(1-e_D)=0.

Apply Lemma 1.1 to yy and uj=λrju_j=\lambda_{r^j}. Its norm is ∥x∥\|x\|, and this gives (16). □\square

The estimate is for the whole polynomial, rather than a separate estimate summed over its coefficients. One value of mm works for its entire finite support.

Corollary 3.2. For every x∈Ax\in A, τ(x)1\tau(x)1 is in the operator-norm closure of the convex hull of {λgxλg∗:g∈Γ}\{\lambda_gx\lambda_g^*:g\in\Gamma\}.

Proof. Given η>0\eta>0, choose a polynomial zz with ∥z−x∥<η\|z-x\|<\eta, and put z0=z−τ(z)1z_0=z-\tau(z)1. Then

∥z0−(x−τ(x)1)∥<2η.\|z_0-(x-\tau(x)1)\|<2\eta.

The average TNT_N from Theorem 3.1 is a unital contraction and fixes scalars. Consequently

∥TN(x)−τ(x)1∥<2η+2∥z0∥N.(17)\|T_N(x)-\tau(x)1\| <2\eta+\frac{2\|z_0\|}{\sqrt N}. \tag{17}

First choose η\eta small, then NN large. Each TN(x)T_N(x) is an actual finite convex combination of the required conjugates. □\square

4. Simplicity and the only trace

For the arbitrary generator set in this lesson, here is the full trace check on AA. On group polynomials x=∑gxgλgx=\sum_g x_g\lambda_g, y=∑gygλgy=\sum_g y_g\lambda_g,

τ(xy)=∑gxgyg−1=∑gygxg−1=τ(yx).\tau(xy)=\sum_g x_g y_{g^{-1}}=\sum_g y_g x_{g^{-1}}=\tau(yx).

Both sums are finite. Norm approximation by polynomials and continuity of the vector state extend this identity to every x,y∈Ax,y\in A. For faithfulness, the right translations Rhδg=δgh−1R_h\delta_g=\delta_{gh^{-1}} commute with all of AA. If z∈A+z\in A_+ and τ(z)=0\tau(z)=0, then z1/2δe=0z^{1/2}\delta_e=0. Commutation gives z1/2Rhδe=0z^{1/2}R_h\delta_e=0 for every hh. These vectors include every basis vector and have dense linear span in ℓ2(Γ)\ell^2(\Gamma), so z1/2=0z^{1/2}=0 and z=0z=0. The argument requires no countability of the basis.

Theorem 4.1. The reduced C∗C^*-algebra of a free group with at least two generators is simple, and its canonical trace is its unique tracial state.

Proof. Every tracial state ρ\rho is invariant under conjugation by the group unitaries. It has the same value on xx and every convex average in Corollary 3.2, so continuity gives

ρ(x)=ρ(τ(x)1)=τ(x).\rho(x)=\rho(\tau(x)1)=\tau(x).

Let JJ be a nonzero closed two-sided ideal. Choose 0≠z∈J+0\ne z\in J_+; for example take z=y∗yz=y^*y for nonzero y∈Jy\in J. Faithfulness gives τ(z)>0\tau(z)>0. After rescaling, assume τ(z)=1\tau(z)=1. Corollary 3.2 provides an element v∈Jv\in J with ∥v−1∥<1\|v-1\|<1, because every conjugate of zz remains in JJ. The Neumann series makes vv invertible in AA, hence 1=v−1v∈J1=v^{-1}v\in J. Therefore J=AJ=A. □\square

The distinction between the two norms matters. A trace 22-norm approximation to 11 need not produce an invertible element; the argument above uses operator norm.

5. Exercises with complete solutions

Exercise 1. Give a norm-equality counterexample with orthogonal ranges satisfying (2).

Solution. Take the two projections onto the coordinate axes of C2\mathbb C^2. Their ranges are orthogonal; their sum is the identity. Its squared norm is 11, while the sum of their squared norms is 22. They are the two conjugates of v=ev=e under 11 and the coordinate flip, and eu1∗u2e=0eu_1^*u_2e=0. This verifies (7) under the full hypotheses.

Exercise 2. Prove the coefficient N−1/2N^{-1/2} in the estimate for v=evv=ev cannot be replaced by N−1N^{-1}.

Solution. On a Hilbert space with orthonormal e0,e1,…,eNe_0,e_1,\ldots,e_N, let ee project onto e1e_1 and vξ=e1⟨e0,ξ⟩v\xi=e_1\langle e_0,\xi\rangle. Choose uju_j fixing e0e_0 and carrying e1e_1 to eje_j. Their conjugates of vv send e0e_0 to the orthogonal unit vectors eje_j. Thus the sum has norm N\sqrt N, and its average has norm N−1/2N^{-1/2}. The projections ujeuj∗u_jeu_j^* are orthogonal, so (2) holds.

Exercise 3. Which part of the proof of (9) requires exclusion of an initial bb-block of exponent 11?

Solution. For negative kk, the terminal b−1b^{-1} of rkr^k can cancel against an initial positive bb-block. Exponent 11 would remove that block completely and could expose an aa-letter which cancels further into the prefix. Excluding exponent 11 leaves a nonzero bb-block, preserving the initial a−1a^{-1}. For example r−1(ba)=e∉Cr^{-1}(ba)=e\notin C; the input baba is in CC, not in DD.

Exercise 4. Apply the stripping construction (11) to w=(ab)−2c(ab)w=(ab)^{-2}c(ab), where cc is a third generator.

Solution. Here p=2,q=1,v=cp=2,q=1,v=c, and ℓ(w)=7\ell(w)=7. For m>7m>7, the conjugate is (ab)m−2c(ab)−(m−1)(ab)^{m-2}c(ab)^{-(m-1)}. No join cancels, so it starts with aa and ends with a−1a^{-1}, as required.

Exercise 5. Explain why q0q_0 contains b2b^2 rather than bb, and test the weaker printed condition with d=bcd=bc.

Solution. The initial b2b^2 puts the result outside CC; the final b−2b^{-2} leaves one b−1b^{-1} after multiplying by an input whose initial block is b1b^1. Replacing dd by bcbc satisfies d≠e,bd\ne e,b but begins with exponent 11 of bb. With w=a,m=2w=a,m=2, the conjugated word and its product with a∈Ca\in C retain that initial block, giving (15) in CC. Thus the weaker condition fails in rank three.

Exercise 6. For x=λa+2λb−λabx=\lambda_a+2\lambda_b-\lambda_{ab}, explain how a single averaging family controls the whole polynomial.

Solution. The identity coefficient is zero, and the largest support length is 22. Choose m=3m=3, q0=b2ab2q_0=b^2ab^2, and the conjugators rjq0t3r^jq_0t^3. Every transformed supporting word sends CC into DD, so the compression of the whole transformed polynomial to ℓ2(C)\ell^2(C) is zero. Formula (16) gives 2∥x∥/N2\|x\|/\sqrt N, without replacing ∥x∥\|x\| by a sum of coefficient magnitudes.

Exercise 7. Does one averaging family work for several polynomials at once?

Solution. Yes. Choose mm exceeding the lengths in the union of their finite supports after removing identity coefficients. The sets C,DC,D, the word q0q_0 and the powers rjr^j then work for every polynomial. The bound for the ii-th polynomial is 2∥xi−τ(xi)1∥/N2\|x_i-\tau(x_i)1\|/\sqrt N. Approximating a finite set of arbitrary elements by polynomials and applying (17) gives simultaneous norm approximation as well.

Exercise 8. Why does the simplicity proof need a positive element and a faithful trace?

Solution. A nonzero element can have trace zero, so rescaling it to trace 11 may be impossible. If y≠0y\ne0 lies in the ideal, y∗yy^*y is a nonzero positive element in it. Faithfulness then ensures τ(y∗y)>0\tau(y^*y)>0. Rescaling and norm averaging produce an invertible ideal element and force the ideal to contain 11.

Exercise 9. Give a finite-dimensional example showing that closeness to 11 in normalized trace 22-norm does not imply invertibility.

Solution. In MdM_d, let pp project onto the first coordinate and set x=1−px=1-p. It is singular, but ∥x−1∥2=∥p∥2=d−1/2\|x-1\|_2=\|p\|_2=d^{-1/2}, tending to zero as dd increases. Its operator-norm distance from 11 is 11, so it never satisfies the Neumann criterion.

Exercise 10. For the free group on one generator, show that the conclusion of Theorem 4.1 fails.

Solution. Let S=λ1S=\lambda_1 on ℓ2(Z)\ell^2(\mathbb Z). The algebra generated by SS is commutative. For ζ∈{1,−1}\zeta\in\{1,-1\}, the unit vectors (2N+1)−1/2∑k=−NNζ−kδk(2N+1)^{-1/2}\sum_{k=-N}^N\zeta^{-k}\delta_k satisfy ∥SξN−ζξN∥→0\|S\xi_N-\zeta\xi_N\|\to0. Their vector states on Laurent polynomials tend to p(S)↦p(ζ)p(S)\mapsto p(\zeta), giving bounded characters of the norm closure. The character at 11 has a proper kernel containing the nonzero element S−1S-1; the two characters are distinct tracial states. Commutativity also makes every conjugate of xx equal to xx, preventing scalar averaging for a nonscalar element.

Reading and attribution

Pierre de la Harpe, On simplicity of reduced C*-algebras of groups, arXiv:math/0509450v1, Section 3, Definition 9 and full Theorem 14 proof, pp.7 and 10–11, supplies the Powers partition and orthogonal-range averaging method. The exact 2/N2/\sqrt N estimate and invertible-ideal step were actually read. Appendix IX, p.19, discusses scalar convex averaging by reference; the complete scalar averaging and unique-trace arguments are proved here.

Sections 1–4 give the complete finite-word construction, correct range inequality, fixed b2ab2b^2ab^2, simultaneous finite-support estimate and ideal proof. The local two-range decomposition works for non-self-adjoint inputs, and invertibility uses the Neumann series in 1−v1-v. The argument applies to every free generator set of size at least two, with no countability hypothesis or reliance on a boundary-action theorem. Exact transitive Hilbert-space, group trace and C*-algebra foundations remain pending; no source expression was imported.