Convexity, the spectrum of free Laplacians and geometric consequences: a free KLS bound

The spectral projections of \Delta\otimes \mathrm{id} and \mathrm{id}\otimes\Delta are E(B)\otimes \mathrm{id} and \mathrm{id}\otimes E(C), and they therefore commute strongly. The joint spectral theorem gives the preceding identity together with the operator in (2.10). It remains first to verify the assertion concerning the core. For n\geq 0, put Q_n:=E([0,n])\otimes E([0,n]). For every x\in\mathrm{Dom}(\Delta^\otimes), the joint spectral calculus gives Q_nx\to x in the graph norm of \Delta^\otimes. On \mathrm{Ran}(Q_n), this graph norm is equivalent to the Hilbert-space norm, and

is dense in \mathrm{Ran}(Q_n) and contained in \mathrm{Dom}(\Delta)\otimes\mathrm{Dom}(\Delta). Hence, the algebraic tensor product is indeed a core. We now verify the projection estimate and set

Then

Since 0\leq \lambda+\mu\leq s implies 0\leq \lambda,\mu\leq s, we have the immediate inclusion T_s\subset[0,s]^2, and the monotonicity of projection-valued measures yields (2.10). The range of the projection in the right-hand side of (2.10) has dimension N_\Delta(s)^2, proving the desired rank estimate. Finally, since \lambda,\mu\geq0, the equality \lambda+\mu=0 is equivalent to \lambda=\mu=0, and (2.10) follows immediately from (2.9). This concludes the proof.

This was proved in [DiezObata, Lemma 2.19]. We nevertheless recall the short argument for the convenience of the reader. We denote by \dagger the isometric Hilbert–Schmidt involution on M\otimes M^{op}, given on simple tensors by (a\otimes b^{op})^\dagger=b^*\otimes(a^*)^{op}, and let \sigma(a\otimes b^{op})=b\otimes a^{op} denote the flip. The real property of the free difference quotients first gives (\bar\partial_j\xi_i)^\dagger=\bar\partial_j\xi_i, whereas the Schwarz symmetry of the conjugate system gives \bar\partial_j\xi_i=\sigma(\bar\partial_i\xi_j). Since \sigma is an involutive *–anti-automorphism, we obtain

This proves the self-adjointness of \mathscr J\xi.

For t\geq0, we denote by

the positive spectral subspace of \Delta below t. Fix f\in E_t, and denote by \nu_f its scalar spectral measure with respect to \Delta:

By definition, the measure \nu_f is supported in (0,t]. Thus, for every r>0, we get

The spectral characterization of the domains of the powers of a positive self-adjoint operator therefore shows that f\in\mathrm{Dom}(\Delta^r) for every r\geq0. In particular, f\in\mathrm{Dom}(\Delta^{3/2}) and \Delta f\in\mathrm{Dom}(\Delta^{1/2})=\mathrm{Dom}(\bar\partial). We may now apply Dabrowski’s almost-commutation formula from Lemma 2.15. It gives \bar\partial_i f\in\mathrm{Dom}(\Delta^\otimes) for every i=1,\ldots,n, together with the fundamental identity

Taking the scalar product of this identity with \bar\partial f, we obtain

But \Delta f\in\mathrm{Dom}(\bar\partial) and f\in\mathrm{Dom}(\Delta). Therefore, the duality for the closed Dirichlet form yields

We therefore have the identity

We now use in an essential way the fact that \nu_f is supported in (0,t]. Indeed,

On the other hand, the condition CD(c,\infty) combined with Proposition 2.12 gives

Combining all the preceding identities and estimates, we arrive at the following spectral-lowering estimate

We first consider the interval 0\leq t<c. The left-hand side of (Proof) is non-negative since \Delta^\otimes\geq0, and its right-hand side is strictly negative unless \bar\partial f=0. We thus necessarily have \bar\partial f=0, and (2.9) shows that f is scalar. But also

because E_t and \mathrm{Ker}(\Delta) are the ranges of the spectral projections corresponding to the disjoint Borel sets (0,t] and \{0\}. We obtain f=0, hence E_t=\{0\}, and consequently

This proves (3.1), and in particular already proves that the spectral gap is forced by the convexity criterion. We turn to the more interesting case t\geq c. Set s=t-c, and assume for the moment that

We consider the following natural lowering map:

The main point is that this map is injective. To see it, let f\in E_t and suppose that

For every i=1,\ldots,n, denote by

the scalar spectral measure of \bar\partial_i f with respect to \Delta^\otimes. Our assumption gives \nu_{i,f}([0,s])=0. Moreover, since \bar\partial_i f\in\mathrm{Dom}(\Delta^\otimes), the first moment of \nu_{i,f} is finite, and we have

Assume by contradiction that \bar\partial f\neq0. Then \bar\partial_i f\neq0 for at least one i, and consequently

Since (s,\infty)=\bigcup_{m\geq1}[s+m^{-1},\infty), we can find m\geq1 such that \nu_{i,f}([s+m^{-1},\infty))>0. It follows at once that

Thus, the right-hand side above is strictly positive, and we obtain

But s=t-c, and this is in direct contradiction with (Proof). Hence \bar\partial f=0. Using again (2.9), together with f\perp1, we get f=0. This proves that the map (Proof) is injective. The target space in (Proof) is the n–fold direct sum of the range of \mathds1_{[0,s]}(\Delta^\otimes). By Proposition 2.10 and our assumption N_\Delta(s)<\infty, this space is finite-dimensional. The injectivity proved above now gives

Moreover, the spectral calculus gives the orthogonal decomposition

and we already know that \dim\mathrm{Ker}(\Delta)=1. We have thus established

under the only condition that t\geq c and N_\Delta(t-c)<\infty. It only remains to prove that N_\Delta(t)<\infty for every finite t, and thus to remove this (apparently circular) finiteness assumption. We do this by a very simple induction. For k\geq0, consider the assertion

The case k=0 follows from N_\Delta(0)=\dim\mathrm{Ker}(\Delta)=1. Suppose that \mathcal P(k) holds for some k\geq0, and take 0\leq t\leq(k+1)c. If 0\leq t<c, then (3.1) gives N_\Delta(t)=1. If instead c\leq t\leq(k+1)c, then 0\leq t-c\leq kc, and the induction hypothesis gives N_\Delta(t-c)<\infty. We can therefore apply the conditional estimate (Proof), which yields

This proves \mathcal P(k+1). By induction, N_\Delta(t)<\infty for every finite t. In particular, the finiteness assumption in (Proof) is now automatically satisfied for every t\geq c, and (3.1) follows. We finally deduce compactness of the resolvent. For R>0, the spectral calculus gives

The first operator has finite rank, since N_\Delta(R)<\infty, while

We have proved that (1+\Delta)^{-1} is a norm limit of finite-rank operators. It is therefore compact, and the proof is complete.

Indeed, we have

Here both inequalities follow directly from the cumulative domination assumption, since s-u\in[0,s] for every u\in[0,s]. The middle identity follows from the commutativity of classical convolution and Tonelli’s theorem.

By Theorem 3.1, \Delta has compact resolvent. We can therefore consider its spectral counting measure (namely, the trace of its spectral measure), \mu_\Delta := \sum_{\lambda\in\sigma(\Delta)} \dim\mathrm{Ker}(\Delta-\lambda\:\mathrm{id})\,\delta_\lambda, so that N_\Delta(t)=\mu_\Delta([0,t]). The only change in the preceding proof occurs when estimating the dimension of the target of (Proof). Indeed, the joint spectral calculus recalled in Proposition 2.10 gives the precise identity

The injectivity of (Proof) therefore yields

The form of inequality (Proof) suggests looking for a discrete measure which satisfies the same relation with equality. Since each successive application of the descent procedure decreases the spectral level by c, it is natural to look for such a measure supported on the non-negative lattice c\mathbb Z_{\geq 0}=\{kc:k\geq0\}. Let us write \nu := \sum_{k=0}^{\infty}a_k\,\delta_{kc}, where a_k\geq0. We impose the fixed-point convolution equation \nu = \delta_0+n\,\delta_c*(\nu*\nu). Comparing first the masses at 0, and then at (k+1)c, we obtain

Since a locally finite measure supported on c\mathbb Z_{\geq0} is uniquely determined by the sequence of its point masses, this recursion determines \nu uniquely. The Catalan recursion therefore gives a_k=C_kn^k for every k\geq0. Consequently, the unique locally finite positive measure supported on c\mathbb Z_{\geq0} which satisfies the preceding fixed-point relation is \nu := \sum_{k=0}^{\infty}C_kn^k\,\delta_{kc}. It is locally finite since every bounded interval contains only finitely many points of c\mathbb Z_{\geq0}. The Catalan recursion C_{k+1}=\sum_{j=0}^{k}C_jC_{k-j} is equivalent to the convolution identity \nu = \delta_0+n\,\delta_c*(\nu*\nu). Thus, setting G(t):=\nu([0,t]), we have G(t)-1 = n(\nu*\nu)([0,t-c]), for t\geq c. We claim that N_\Delta(t)\leq G(t) for every t\geq0. By (3.1), this holds for 0\leq t<c, while (Proof) at t=c gives

Suppose that the claim holds on [0,kc] for some k\geq1, and take kc\leq t\leq(k+1)c. Set s=t-c, so that s\leq kc. For every 0\leq r\leq s, the induction hypothesis gives

Now using Lemma 3.2, applied with \alpha=\mu_\Delta and \beta=\nu, gives

Combining this with (Proof), we obtain

The assertion thus follows by induction. Using the elementary bound C_k\leq4^k, we conclude that

which proves the first estimate. It remains to establish the trace-class estimate of the heat semigroup. Fix t>0 (at this stage, all the quantities below initially are understood as elements of [0,\infty]). For every \lambda\geq0, one has

Since N_\Delta(r)\leq G(r) for every r\geq0, Tonelli’s theorem therefore gives

If C denotes the generating function of the Catalan numbers, then, for every 0<x\leq1/4, C(x):=\sum_{k=0}^{\infty}C_kx^k=\frac{1-\sqrt{1-4x}}{2x}. Taking x=ne^{-t c}, we finally obtain

when t\geq \frac{\log(4n)}{c}.

Let (S_t^{(i)})_{t\geq0}, 1\leq i\leq n, be free (standard) Brownian motions, free from M. Set

and \widetilde M=M*N, \widetilde M_t=M*N_t. The trace on \widetilde M is again denoted without confusion by \tau, and we have N\cong L(\mathbb F_\infty). Following [Dab14, Section 2.1], extend the polynomial free difference quotient by zero on the free Brownian variables:

where the free product is algebraic. This derivation is closable. Denote its joint closure by \bar\delta and put, for \alpha>0,

The subspace L^2(M) reduces \widetilde\Delta, whose restriction there is \Delta. We also write

Tensor operators act componentwise on the n-fold direct sum. As usual, we use the notation (a\otimes b^{op})\#dS_t^{(i)}=a\,dS_t^{(i)}b for stochastic integration. The filtration properties of these operators are those established in [Dab14, Section 2.1, Lemma 16].

The Lipschitz conjugate variable hypothesis satisfies [Dab14, Assumption 1], as explained in the proof of [Dab14, Corollary 25] there. The results of [Dab14, Theorems 17 and 11(ii)], together with [Dab14, Propositions 23–24], therefore provide normal unital trace-preserving embeddings \Phi_t:M\to\widetilde M_t on this free product. For x\in M_0, the process Y_t=\Phi_t(x) is the mild solution

We use its continuous L^2 representative. In particular, Y\in\mathcal B_{2,\bar\delta}^{a}([0,T]) for every T<\infty, where the adapted graph spaces are those of [Dab14, Section 1.2] (the factors 1/2 in (?) come from Dabrowski’s normalization [Dab14], which uses the generator -\widetilde\Delta/2). For \epsilon\in(0,1), let Y^\epsilon be the regularized solution from [Dab14, Theorem 11(ii)]:

Since the two derivations in that theorem coincide here, Y^\epsilon is a strong solution and belongs to \mathcal B_{2,\widetilde\Delta}^{a}([0,T]) for every 0<T<\infty. The spaces \mathcal B_{2,\bar\delta}^{a}([0,T]) and \mathcal B_{2,\widetilde\Delta}^{a}([0,T]) are the completions of the corresponding simple adapted processes for the respective norms

and

These correspond to the choices \beta=0 and \beta=1, respectively, in the definitions of [Dab14, Section 1.2]. In particular, Y_s^\epsilon\in\mathrm{Dom}(\widetilde\Delta) for almost every s\in[0,T], and

Moreover, as \epsilon\downarrow0, Y^\epsilon converges to Y weakly in \mathcal B_{2,\bar\delta}^{a}([0,T]) and uniformly on [0,T] in L^2(\widetilde M). We first record the coercivity given by positive curvature, noting that the conjugate variables of the zero extension remain \xi_i. By using again Lemma 2.15, in its zero-extension form from [Dab14, Lemma 19(iii)],

including the corresponding domain assertion. The spatial inclusion M\bar\otimes M^{op}\subset\widetilde M\bar\otimes\widetilde M^{op} preserves positivity. By the representation properties recalled above in Proposition 2.12, we also have \mathcal R_{\mathscr J\xi}\geq c\,\mathrm{id} on the enlarged gradient space. The free Bochner formula (taking the inner product with \bar\delta u) therefore gives

For u\in\mathrm{Dom}(\widetilde\Delta), apply this inequality to \widetilde\rho_\alpha u\in\mathrm{Dom}(\widetilde\Delta^{3/2}). We have convergence to u in the graph norm of \widetilde\Delta, and the free difference quotients converge to \bar\delta u in L^2. Applying the same inequality to \widetilde\rho_\alpha u-\widetilde\rho_\beta u, we see that (\widetilde\Delta^\otimes\otimes I_n)^{1/2}\bar\delta\widetilde\rho_\alpha u is Cauchy in L^2. By closedness, we therefore obtain \bar\delta u\in\mathrm{Dom}((\widetilde\Delta^\otimes\otimes I_n)^{1/2}) and can pass to the limit in the inequality (see also [Dab14, proof of Theorem 17], explained in greater detail in [DiezObata, Lemma 3.6]). Thus, writing

we have

We fix \epsilon\in(0,1) and T>0. We apply [Dab14, Proposition 5, equation (6)] with B=\widetilde\Delta^{1/2} and \alpha'=\alpha, as in  [Dab14, Step 1 of the proof of Theorem 11 and Lemma 3(ii)–(iii)]. Together with the free Itô isometry of Biane–Speicher [BS], this gives the following identity (the real-part of the inner product in [Dab14, Proposition 5, equation (6)] can be omitted here, since the corresponding inner product equals a squared norm by spectral calculus):

By (?), the right-hand side is bounded by

We can let \alpha\to\infty by using monotone convergence. The (deterministic) semigroup satisfies

Moreover, \gamma(Y^\epsilon)_t\perp L^2(M), since its conditional expectation onto M is zero, while the deterministic term belongs to L^2(M). Since this subspace reduces \widetilde\Delta, these terms remain orthogonal after applying the spectral powers above. Adding the two energy identities and using (?), we obtain

Since 1-(1-\epsilon)^2\geq0, the coercivity estimate (?) yields

Omitting the first term and using weak lower semicontinuity in \mathcal B_{2,\bar\delta}^{a}([0,T]), we can reach the limit \epsilon\to0. Letting then T\to\infty, we obtain

We next identify the strong limit of \widetilde\varphi_r. Since \delta vanishes on the polynomials in the free Brownian variables, this is clear by density that L^2(N)\subset \mathrm{Ker}(\widetilde\Delta). We now prove the reverse inclusion using the word-space decomposition introduced before assumption \Gamma_{0u}(\omega) in [Dab14, p. 1409] and the tensor representation of the zero extension in [Dab14, Section 2.1.3]. We choose an orthonormal basis (e_j)_{j\geq1} of L^2(N)\ominus\mathbb C.1 obtained by applying Gram–Schmidt to a countable L^2-dense family of trace-zero non-commutative polynomials in the free Brownian variables at non-negative rational times. For k\geq1, set

with the middle factor omitted when k=1. On this space, \Delta^{\otimes(k+1)} denotes the positive self-adjoint closure of

on \mathrm{Dom}(\Delta)^{\otimes (k+1)}\cap \mathcal V_k (the algebraic tensor product). For each i=(i_1,\ldots,i_k)\in\mathbb N^k, write e_i=e_{i_1}\otimes\cdots\otimes e_{i_k}. We also use the multiplication map (a_0\otimes\cdots\otimes a_k)\#e_i =a_0e_{i_1}a_1\cdots e_{i_k}a_k, initially defined for a_0,\ldots,a_k\in M with \tau(a_j)=0 for 1\leq j\leq k-1. By freeness, this map extends to an isometry on \mathcal V_k, we also denote its (closed) range by setting \mathcal V_k\#e_i :=\{U\#e_i:U\in\mathcal V_k\}. By the standard Hilbert space decomposition of a tracial free product (see [AnanPopa, Section 5.3.2 and Remark 5.3.5])

The last assertion of [Dab14, Corollary 15] verifies his condition \Gamma_{0u}(\omega)\:e.2) for these tensor extensions of \Delta: for U\in\mathrm{Dom}(\Delta^{\otimes(k+1)}), one has U\#e_i\in\mathrm{Dom}(\widetilde\Delta) and

Applying this to U=\bigl(\alpha+\Delta^{\otimes(k+1)}\bigr)^{-1}(W) with \alpha>0 gives

Each closed word space is therefore invariant under the self-adjoint resolvent, and hence reduces \widetilde\Delta, whose restriction there is unitarily equivalent to \Delta^{\otimes(k+1)}. Since \ker\Delta=\mathbb C.1, the kernel of \Delta^{\otimes 2} on \mathcal V_1 is \mathbb C.(1\otimes1), whose inserted copies are the e_j’s. For k\geq2, a centered internal factor makes the kernel on \mathcal V_k zero. Consequently, we deduce that \ker\widetilde\Delta=L^2(N) and denoting by E_N the trace-preserving conditional expectation onto N, also viewed as the orthogonal projection onto L^2(N), the spectral theorem gives

We now use (?) to pass to the limit in the mild equation. Fix T\geq0. We split the stochastic integral in (?) at T and, using the semigroup identity of [Dab14, Lemma 3(ii)], we obtain, for t\geq T,

The free Itô isometry and L^2-contractivity of the semigroup associated to \widetilde\Delta^{\otimes} give

For fixed T, the deterministic term converges to E_NY_T by (?). Hence

The triangle inequality gives \underset{t,t'\to\infty}{\limsup}\|Y_t-Y_{t'}\|_2 \leq2\left(\int_T^\infty\|\bar\delta Y_s\|_2^2\,ds\right)^{1/2}. The right-hand side tends to zero as T\to\infty, so Y_t is Cauchy in L^2(\widetilde M). Its limit then belongs to L^2(N), since (?) approximates it by E_NY_T\in L^2(N) with an error tending to zero. We thus have proved that

The L^2 isometry of the maps \Phi_t and density of M_0 extend this convergence to every x\in M. Moreover, we see that E_N\Phi_t(x)\in N converges to \Phi_\infty(x) in L^2 and has operator norm at most \|x\|. Since the operator-norm balls in N are closed in the L^2 topology, we get \Phi_\infty(x)\in N and \|\Phi_\infty(x)\|\leq\|x\|. For x,y\in M,

Since \Phi_t(xy)=\Phi_t(x)\Phi_t(y) and \Phi_t(xy)\to\Phi_\infty(xy) in L^2, uniqueness of the limit gives

Therefore, we see that multiplicativity is preserved at the limit. The same is true for linearity, the involution, the unit and the trace. Injectivity is ensured by preservation of the faithful trace. The map is also normal: indeed, if 0\leq x_\alpha\uparrow x, then the positive element \Phi_\infty(x)-\underset{\alpha}{\sup}\:\Phi_\infty(x_\alpha) has trace zero by trace preservation and normality of the traces, hence vanishes by faithfulness. Finally, the subgroup generated freely by \{b^kab^{-k}:k\in\mathbb Z\}\subset\mathbb F_2=\langle a,b\rangle gives a trace-preserving inclusion L(\mathbb F_\infty)\subset L(\mathbb F_2). Composing the two embeddings gives the conclusion.

The tuple X has Lipschitz conjugate variables, and its von Neumann algebra M=W^*(X_1,\ldots,X_n) is now known to be R^\omega–embeddable, so Dabrowski [Dab14, Corollary 25] gives \delta_0(X)=n.

For m\geq1, set \rho_m=m(m+\Delta)^{-1}. By Proposition 2.7, each map \rho_m:M\to M is normal, u.c.p. and \tau-preserving, and \left\lVert \rho_m(x)-x\right\rVert_2\to0 for every x\in M. Since \Delta has compact resolvent, the L^2-implementation of \rho_m is compact. Thus (\rho_m)_{m\geq1} satisfies the conditions of the preceding definition. This argument uses only compactness of the resolvent and the completely Dirichlet form, not the embedding.

The algebraic free difference quotient \partial:M_0\to(L^2(M)\bar\otimes L^2(M^{op}))^{\oplus n} is a real closable derivation, and its closure is \bar\partial. Indeed, we have (\partial_i x)^\dagger=\partial_i(x^*) for every x\in M_0, and by Remark 2.2 this identity passes to the closure. The target is a finite direct sum of coarse M–M correspondences (hence it is weakly contained in the coarse correspondence). Moreover, the positive operator associated with our derivation is exactly \Delta=\partial^*\bar\partial. The first assertion follows directly from Ozawa–Popa [OP2, Remark 5.4], since \Delta has compact resolvent. We now recall how solidity is obtained. Take a diffuse von Neumann subalgebra Q\subset M, choose a diffuse abelian von Neumann subalgebra A\subset Q, and set \mathcal G:=\mathcal U(Q'\cap M). Every element of \mathcal G commutes with A, and thus the conjugation action \mathcal G\curvearrowright A is trivial. The algebra A is amenable, while the trivial action is compact and hence weakly compact by Ozawa–Popa [OP1, Proposition 3.2]. Applying the first part of the theorem with P=A shows that \mathcal G'' is amenable. But

Thus Q'\cap M is amenable, which is exactly solidity of M.

The first part is a well-known fact. For the second part, L(\mathbb F_2) is weakly exact, and weak exactness passes to the range of a normal conditional expectation by Ozawa [OzawaWeak2, Corollary 3 and the subsequent discussion]. Hence Q is weakly exact.

Take P\subset M diffuse and amenable. By Theorem 4.13, the conjugation action \mathcal N_M(P)\curvearrowright P is weakly compact. We can therefore apply Theorem 4.10 with \mathcal G=\mathcal N_M(P), and we get that \mathcal N_M(P)'' is amenable.

The free difference quotient is a closable real derivation into a finite direct sum of coarse correspondences, and its associated positive generator is \Delta=\partial^*\bar\partial. By the compact-resolvent assumption,

Non-amenability of M excludes the central state in Ding–Kunnawalkam Elayavalli–Peterson [DKEP, Proposition 9.4 and Lemma 9.5] and therefore gives proper proximality.

For every \alpha>0, compactness of the resolvent gives \rho_\alpha := \alpha (\alpha+\Delta)^{-1} \in\mathbb K(L^2(M)). The first assertion therefore follows directly from Ding–Peterson [DP, Theorem 7.11], applied to the free difference quotient case. If C^*(X_1,\ldots,X_n) is exact, then its ultraweak closure M is weakly exact [OzawaWeak2]. Since every von Neumann subalgebra has a trace-preserving conditional expectation, its bi-exactness follows from [DP, Proposition 6.10]. The proper-proximality alternative is [DP, Corollary 6.11].

By Corollary 4.12, M is weakly exact. Theorem 3.1 gives compact resolvent, so Corollary 4.17 applies. When n\geq2, M is a non-amenable factor by Corollary 4.15; it has no amenable direct summand. The proper-proximality assertions follow from [DP, Corollary 6.11].

Take f\in\mathrm{Ker}(\Delta-c\:\mathrm{id}). By spectral calculus, f\in\mathrm{Dom}(\Delta^r) for every r\geq0, and Lemma 2.15 gives

Taking the scalar product with \bar\partial f and subtracting c\left\lVert \bar\partial f\right\rVert_2^2, we get

Both terms on the right-hand side are non-negative, and their sum vanishes. Hence \bar\partial f\in\mathrm{Ker}(\Delta^\otimes\otimes I_n). By (2.10), there exist a_1,\ldots,a_n\in\mathbb C such that \bar\partial_jf=a_j.(1\otimes 1) for every j=1,\ldots,n. We therefore get f-\sum_ja_jX_j\in\mathrm{Ker}(\bar\partial)=\mathbb C\,1. Since c>0, the identities \Delta f=cf and \Delta1=0 imply that f\perp1. Centering this affine expression gives (5.1). If f=f^*, all the coefficients a_j are real (see [DiezObata, Lemma 3.14] for the complete details).

Since \Delta has compact resolvent and \mathrm{Ker}(\Delta)=\mathbb C.1, the bottom \lambda_1 of its positive spectrum is an eigenvalue. Moreover, C_P(X)=\lambda_1^{-1}, and hence C_P(X)=1 gives \lambda_1=1. We can therefore apply the multidimensional free Obata theorem of [DiezObata, Theorem 3.15], which gives the desired conclusion. The second isomorphism comes again from Open AI’s result [openai]. Finally, tracial free products are associative and the uniqueness characterization of tracial free products in [AnanPopa, Proposition 5.3.3], interchanging the order of the two components gives a canonical normal trace-preserving isomorphism. Using Open AI’s isomorphism result [openai] then gives the free absorption by L(\mathbb F_2).

By Theorem 3.1, the operator \Delta has compact resolvent and \ker(\Delta)=\mathbb C.1. Consequently, its minimal non-zero spectral value is an eigenvalue. We denote it by \lambda := \min(\sigma(\Delta)\setminus\{0\}). The free Poincaré constant is therefore given by C_P(X)=\lambda^{-1}. We choose a normalized centered eigenfunction f\in L^2(M):

Note that \tau(f)=0 is immediate as f belongs to the orthogonal complement of \mathrm{Ker}(\Delta)=\mathbb C.1. Moreover, since f belongs to an eigenspace of \Delta, it belongs to \operatorname{Dom}(\Delta^r) for every r>0. In particular, f\in\operatorname{Dom}(\Delta^{3/2}). Dabrowski’s almost-commutation formula 2.15 therefore applies to f, and \bar\partial_i f\in\operatorname{Dom}(\Delta^\otimes), for all i=1,\ldots,n. Let P_0 denote the orthogonal projection of L^2(M) onto \mathbb C.1, and use the same notation for the corresponding projection of L^2(M^{op}) onto \mathbb C.1. By a slight abuse of notation, the identity operators on L^2(M), on L^2(M^{op}), and on L^2(M)\bar\otimes L^2(M^{op}) are all denoted by the same symbol \mathrm{id}. By the definition of \lambda and the spectral theorem, we have

in the sense of quadratic forms. Recalling that

it follows that

Indeed, one can check on the core \mathrm{Dom}(\Delta)\otimes \mathrm{Dom}(\Delta) (here again \otimes denotes the algebraic tensor product) that

Since P_0\otimes P_0, initially defined on simple tensors, extends to the orthogonal projection onto \mathbb C.(1\otimes1), we conclude that, for every \eta\in\operatorname{Dom}(\Delta^\otimes),

The same conclusion can also be seen directly by diagonalizing \Delta. Indeed, if 0=\lambda_0<\lambda_1=\lambda\leq\lambda_2\leq\cdots are its eigenvalues, counted with multiplicity, then the eigenvalues of the tensor extension \Delta^\otimes (which thus has discrete spectrum) are the sums \lambda_i+\lambda_j. The only zero sum is for \lambda_0+\lambda_0, and every other sum is at least \lambda. For i=1,\ldots,n, set m_i := \langle \bar\partial_i f,1\otimes1 \rangle_2 =: (\tau\otimes\tau)(\bar\partial_i f). Applying the preceding estimate to each \bar\partial_i f and summing over i, we obtain

Since \lVert\bar\partial f\rVert_2^2 = \langle\Delta f,f\rangle_2 = \lambda, this becomes

We now apply Dabrowski’s almost-commutation formula 2.15:

Taking the inner product with \bar\partial f, and using \Delta f=\lambda f, we get

Combining the last two estimates gives

Equivalently,

This is the free analogue of the “preferred-direction” estimate in Klartag’s terminology (see the discussion after [Klartaglog, Proposition 2.3]). For i=1,\ldots,n, we denote X_i^\circ:=X_i-\tau(X_i).1. Since the free difference quotients are taken with respect to the variables X_1,\ldots,X_n, we have

Integration by parts therefore gives

Hence

Consider the operator

We can check for every a,b\in\mathbb C^n, using the self-adjointness of the X_i^\circ, that we have

Hence, the Gram matrix of T is \operatorname{Cov}(X). Therefore, we have T^*T=\operatorname{Cov}(X), so that \lVert T\rVert^2 = \lVert\operatorname{Cov}(X)\rVert_{\mathrm{op}}. Since \lVert f\rVert_2=1, it follows that

We then get \sum_{i=1}^n\lvert m_i\rvert^2 \leq \lambda^2 \lVert\operatorname{Cov}(X)\rVert_{\mathrm{op}}. Combining this with the preferred-direction estimate gives

Finally, the CD(c,\infty) condition implies \mathcal R_{\mathscr J\xi} \geq c\,\mathrm{id}. Consequently,

Since \lambda>0, this yields

and therefore we deduce

It remains to establish the lower bound. It is easy to see that testing the free Poincaré inequality on linear functions T(a):=a_1X_1^{\circ}+\ldots+a_nX_n^{\circ} and taking the supremum over all unit vectors a\in\mathbb C^n gives

Finally, the free Lichnerowicz inequality under CD(c,\infty) gives

Together with \lVert\operatorname{Cov}(X)\rVert_{\mathrm{op}}\leq C_P(X), this implies

and completes the proof.

By the spectral theorem, since \ker(\Delta)=\mathbb C.1, the free Poincaré inequality is equivalently written for f\in\operatorname{Dom}(\Delta) as \lVert\bar\partial f\rVert_2^2 \leq C_P(X)\lVert\Delta f\rVert_2^2. Using again the free Bochner formula, for every f\in\operatorname{Dom}(\Delta^{3/2}),

The last inequality is obtained by applying the spectral-gap estimate for \Delta^\otimes to each component \bar\partial_i f, that is, by applying (Proof). Since \lVert\operatorname{Cov}(X)\rVert_{\mathrm{op}}>0, we may center the tuple and apply a dilation to normalize its maximal variance to one. Thus, without loss of generality, we assume that \tau(X_i)=0, for all i=1,\ldots,n, and \lVert\operatorname{Cov}(X)\rVert_{\mathrm{op}}=1. In the sequel, we keep the same notation for the tuple and all the associated objects. Under this dilation, the free Laplacian and the curvature constant are multiplied by the original value of \lVert\operatorname{Cov}(X)\rVert_{\mathrm{op}}, while the Poincaré constant is divided by the same quantity. We continue to denote the (new) curvature constant by c. Thus the normalized tuple satisfies CD(c,\infty), and the equality assumption becomes

Testing the free Poincaré inequality on a linear combination of the variables in a direction of maximal variance gives 1\leq C_P(X)=c^{-1/2}, and hence c\leq1. The goal is to prove that c=1. This will show that equality in the improved free Lichnerowicz inequality implies equality in the usual free Lichnerowicz inequality, and the conclusion will then follow from the free Obata rigidity theorem [DiezObata, Theorem 3.15]. By Theorem 3.1, the spectral gap is attained.

Since \Delta commutes with the Tomita conjugation (see the proof of [DiezObata, Corollary 3.4]), we can choose a self-adjoint eigenfunction f\in L^2(M) satisfying \Delta f=\sqrt c\,f, \tau(f)=0, and \lVert f\rVert_2=1. As f\in\ker(\Delta-\sqrt c\,\mathrm{id}), it belongs to \operatorname{Dom}(\Delta^r) for every r>0. We also have \lVert\bar\partial f\rVert_2^2 = \langle\Delta f,f\rangle_2 = \sqrt c, and \lVert\Delta f\rVert_2^2=c. For every i=1,\ldots,n, integration by parts gives, using the self-adjointness of \Delta f and X_i,

If \left( \langle \bar\partial_i f,1\otimes1 \rangle_2 \right)_{i=1}^n =0, then applying (Proof) to f would yield c\geq c+c^{3/2}, which is impossible. We may therefore define \theta := \frac{ ( \langle \bar\partial_i f,1\otimes1 \rangle_2 )_{i=1}^n }{ \left( \sum_{i=1}^n \lvert \langle \bar\partial_i f,1\otimes1 \rangle_2 \rvert^2 \right)^{1/2} }. Then \theta\in\mathbb R^n and \lvert\theta\rvert=1. Moreover,

Combining this estimate with (Proof), applied to f, gives

Hence, equality holds at every step, and in particular in (?) and (?). Equality in (?) means that \theta is a direction of unit variance, and hence a direction of maximal variance under the normalization under consideration. Equivalently,

Equality in (?) implies that \Delta f = \lambda \sum_{i=1}^n\theta_iX_i for some \lambda\in\mathbb R. Taking the L^2-norms of both sides then gives \lvert\lambda\rvert=\sqrt c. Since \Delta f=\sqrt c\,f, we deduce that f = \pm \sum_{i=1}^n\theta_iX_i. Therefore, \bar\partial_jf = \pm\theta_j(1\otimes1), for all j=1,\ldots,n, and hence

Thus c=1, and consequently C_P(X)=1 in the normalized setting. The normalized tuple therefore saturates the usual free Lichnerowicz inequality. The conclusion follows from its equality case, that is, the free Obata rigidity theorem [DiezObata, Theorem 3.15]. Replacing \theta by -\theta, if necessary, and removing the normalization, we obtain

where c again denotes the curvature constant of the original tuple, and \theta is a direction of maximal variance. After an orthogonal change of variables whose first direction is \theta, the first centered variable is semicircular with variance 1/c, it is free from the remaining variables, and, when n\geq2, the remaining tuple satisfies CD(c,\infty) by the conditional-expectation argument in the proof of Corollary 5.4.

Voiculescu’s free Poincaré inequality 2.8 gives a positive spectral gap on L^2_0(M):=L^2(M)\ominus\mathbb C.1. Writing \lambda:=\inf\sigma(\Delta|_{L^2_0(M)}), the spectral theorem gives

Here and below, \Delta^{-1} acts on L^2_0(M). We do not assume that \lambda is an eigenvalue. Choose a centered self-adjoint f\in\mathrm{Dom}(\Delta^{3/2}). Recall that the free Poincaré inequality is equivalently written as \|\bar\partial f\|_2^2\leq C_P(X)\|\Delta f\|_2^2. Our aim is to prove the following estimate:

Since C_P(X)=1/\lambda, the coefficient 7/8<1 will allow us to absorb the term involving C_P(X) and obtain the upper bound in the theorem. We divide the proof into five steps.

Step 1: Let P_0 be the orthogonal projection of L^2(M) onto \mathbb C.1. The spectral theorem gives \Delta\geq\lambda(\mathrm{id}-P_0), so that

as quadratic forms. If \eta\in\mathrm{Dom}(\Delta^\otimes), the joint spectral calculus gives \eta\in\mathrm{Dom}(\Delta\otimes\mathrm{id}). Hence, using the tensor Dirichlet form identity (see, for instance, [DiezObata, Proposition 3.5]),

For a simple tensor, (P_0\otimes\mathrm{id})(x\otimes y) =\tau(x).1\otimes y =1\otimes(\tau\otimes\mathrm{id})(x\otimes y). Both sides extend continuously to the Hilbert tensor product. Orthogonality therefore gives

Together with (?), this proves

Dabrowski’s almost commutation formula in Lemma 2.15 gives, for u\in\mathrm{Dom}(\Delta^{3/2}),

Define the tensor Dirichlet energy by

Taking the inner product of (?) with \bar\partial u, and using CD(0,\infty), yields

Co-associativity also holds on this domain:

Indeed, approximate u by polynomials in the graph norm of \bar\partial and test the polynomial identity against V\in M_0^{\otimes3}. Voiculescu’s formula for the adjoint [Voic, Proposition 4.1] and density of M_0^{\otimes3} then give (?). In particular, the two tensor energies are equal:

Step 2: Reduction to the centered partial traces of the free gradients For the function f fixed above, set

Thus \tau(a_i)=0. Moreover, a_i\in\mathrm{Dom}(\Delta). This follows from Dabrowski’s Lemma 2.15(ii). One can also see it directly: \mathrm{id}\otimes P_0 commutes with the spectral projections of \Delta^\otimes and its range is L^2(M)\bar\otimes\mathbb C.1. The restriction of \Delta^\otimes to this range is \Delta\otimes\mathrm{id}. Applying this observation to \bar\partial_i f\in\mathrm{Dom}(\Delta^\otimes) gives the conclusion. The real property of free difference quotients gives \bar\partial_i(f^*)=(\bar\partial_i f)^\dagger. Since f=f^*, we have m_i\in\mathbb R. Using Mai’s result [Mai, Lemma 4.6] again, for any u\in L^2(M)\bar\otimes L^2(M^{op}) one has (\mathrm{id}\otimes \tau)(u)^\ast = (\tau\otimes \mathrm{id})(u^\dagger), so that

In particular, by orthogonality, since \tau(a_i)=0, we have a_i^*\perp1 in L^2(M). Pythagoras’ theorem and traciality therefore give

Apply (?) to \bar\partial_i f and sum over i=1,\ldots,n. By (?),

Writing X_i^\circ:=X_i-\tau(X_i).1, integration by parts gives m_i=\langle\Delta f,X_i^\circ\rangle_2. Since \operatorname{Cov}(X) is the Gram matrix of the X_i^\circ, Cauchy–Schwarz yields

It remains to estimate A. Applying Poincaré directly to each a_i would give no improvement in (?). We instead use the inverse Laplacian.

Step 3: The Poisson equation and exterior Hessian traces. To express \|a_i\|_2^2 by integration by parts, we solve the Poisson equation. For each i=1,\ldots,n, let

Integration by parts and (Proof) give

Each term \langle a_i,\Delta^{-1}a_i\rangle_2 is precisely the squared H^{-1} norm of a_i. Thus (?) controls the energy of the Poisson solutions, while \Delta h_i=a_i gives \|a_i\|_2^2=\langle\bar\partial a_i,\bar\partial h_i\rangle_2. This is the identity we will use to pair the Hessian of f with the gradients of the h_i. Since a_i\in\mathrm{Dom}(\Delta), in particular a_i\in\mathrm{Dom}(\Delta^{1/2}). On L^2_0(M), spectral calculus gives \Delta^{3/2}\Delta^{-1}a_i=\Delta^{1/2}a_i, so h_i\in\mathrm{Dom}(\Delta^{3/2}). The free Bochner formula and co-associativity therefore apply to each h_i. Define the Hessian

By Dabrowski [Dab14, Lemma 18], the “sliced” derivatives, i.e. composing the conditional expectation \mathrm{id}\otimes\tau with the free difference quotients: (\mathrm{id}\otimes\tau)\circ \partial_j, extend to bounded operators on L^2(M). The same holds for (\tau\otimes\mathrm{id})\circ\partial_i by applying the involution. As in [Dab14, Lemma 19(ii)], co-associativity on polynomials, approximation in the graph norm of \bar\partial, and closedness of \bar\partial therefore give

Here the “bracketed” compositions of conditional expectations with derivatives denote their bounded extensions on L^2(M). For the last equalities, the corresponding identities hold on tensors of non-commutative polynomials, which form a core for the closed operators \overline{\mathrm{id}\otimes\partial_j} and \overline{\partial_i\otimes\mathrm{id}} in (?). They therefore extend by approximation of \bar\partial_i f and \bar\partial_j f in the respective graph norms, using boundedness of these extensions and contractivity of the L^2-extensions of the conditional expectations \mathrm{id}\otimes\mathrm{id}\otimes\tau and \tau\otimes\mathrm{id}\otimes\mathrm{id}. Since constants have zero derivative, the definition of a_j and (?) yield

The adjoints of these conditional expectation maps are respectively given by x\mapsto x\otimes1 and x\mapsto1\otimes x. Thus, by integration by parts, we see immediately that

Since \Delta commutes with the Tomita involution, we have \Delta(h_i^*)=a_i^*. Consequently, using traciality,

Adding the two representations gives

Each outer trace of the Hessian produces one copy of A. Co-associativity allows us to combine them in the same scalar product. Since \sum_{i,j=1}^n\|H_{ij}\|_2^2=\mathcal E_2(f)/2 by (?), Cauchy–Schwarz yields

Step 4: Estimating the mixed term. A direct calculation on simple tensors, extended by linearity and continuity, gives the following identity (recall that our inner products are linear in the first variable): \langle x\otimes1,1\otimes y\rangle_2 =\left\langle(\tau\otimes\mathrm{id})x, (\mathrm{id}\otimes\tau)y\right\rangle_2.

Set s_{ij}:=(\tau\otimes\mathrm{id})(\bar\partial_i h_j). Using Mai’s result [Mai, Lemma 4.6] once more gives (\mathrm{id}\otimes\tau)\bar\partial_j(h_i^*)=s_{ji}^*. Hence, Cauchy–Schwarz implies

Expanding the squared norms, and using \|\bar\partial(h_j^*)\|_2=\|\bar\partial h_j\|_2 (since ^\dagger is an isometric involution), we obtain

The first sum is bounded by A/\lambda by (?).

For the second sum, the domains in (?) and the trace identity (?), applied to h_j, give s_{ij}\in\mathrm{Dom}(\bar\partial), with

The partial trace is contractive. Thus, using (?), (?), and \Delta h_j=a_j, we obtain

The h_j need not be self-adjoint: the Bochner estimate and (?) hold for arbitrary complex u. The means of the s_{ij} are also controlled. Integration by parts gives

As in (?), this yields

We now apply the free Poincaré inequality to the centered family s_{ij}-\tau(s_{ij})\,1. Using (?) and (?) gives

Combining this with (?) and (?) yields

Substituting into (?), we obtain

Dividing by 4A when A>0 gives the following bound, which also holds trivially when A=0:

Step 5: Self-improvement of the Poincaré constant. Combining (?), (?), and (?) gives

By (?), we get

which is (Proof).

We have proved this for centered self-adjoint f\in\mathrm{Dom}(\Delta^{3/2}), and both sides are unchanged by adding a scalar. Since \Delta commutes with the Tomita involution f^*:=Jf (see [DiezObata, Corollary 3.4]), its spectral calculus does as well. For f=\Re f+i\Im f, with \Re f=(f+f^*)/2 and \Im f=(f-f^*)/(2i), the real property of the free difference quotient, stability under adjunction 2.2 and the reality of the corresponding inner products give \|\bar\partial f\|_2^2=\|\bar\partial\Re f\|_2^2 +\|\bar\partial\Im f\|_2^2, and the same decomposition holds for \|\Delta f\|_2^2. Thus (Proof) holds for every f\in\mathrm{Dom}(\Delta^{3/2}). For f\in\mathrm{Dom}(\Delta), consider the spectral truncations

Then f_N\in\mathrm{Dom}(\Delta^r) for every r>0, and f_N\to f, \Delta f_N\to\Delta f in L^2(M). Moreover,

Thus (Proof) extends to all f\in\mathrm{Dom}(\Delta). Now take g\in L^2_0(M) and apply this estimate to f=\Delta^{-1}g. Integration by parts gives

Taking the supremum over unit vectors g and using the positivity of \Delta^{-1} yields

Rearranging gives \frac1{8\lambda}\leq\frac54\|\operatorname{Cov}(X)\|_{\mathrm{op}}, and hence

which concludes the proof.

Since X is centered and isotropic, \tau(\mathbb{X})=n and

The free Poincaré inequality and the definition of C give

Moreover, functional calculus yields

Taking the trace proves the second assertion and concludes the proof.

If X has law \nu_u, then its conjugate variable is \xi=u'(X) and

and on the diagonal \partial u'(x,x)=u''(x). Thus X satisfies CD(c,\infty), and Theorem 3.1 gives compact resolvent. Moreover, the free Caffarelli contraction theorem [Di4, Theorem 7] gives \operatorname{Lip}(T)\leq c^{-1/2}. The map f\mapsto f\circ T is an isometry from L^2(\nu_u) into L^2(\eta), and the chain rule for the free difference quotient yields, first for a polynomial f, the following comparison of the Dirichlet forms \mathcal E_\eta and \mathcal E_{\nu_u}:

Standard approximation for f and closedness then extend this inequality to the full form domain. The min–max principle therefore yields the following comparison of eigenvalues: \lambda_k(\Delta_{\nu_u})\geq c\lambda_k(\Delta_\eta)=ck for every k\geq0. The reader can consult similar details in E. Milman [MilmanSpectral, Section 3] and [Di4] for precise explicit computations of the free Laplacian in this one-dimensional case.