Existence

A non-zero trace φ exists on a two-sided ideal J of operators on a separable Hilbert space if the co-dimension of its commutator subspace is not zero. There are ideals that admit infinitely many linearly independent non-zero singular traces. For example, the commutator subspace of the ideal of weak trace-class operators contains the ideal of trace class operators and every positive operator in the commutator subspace of the weak trace class is trace class.^[12] Consequently, every trace on the weak trace class ideal is singular and the co-dimension of the weak trace class ideal commutator subspace is infinite.^[6]: 191 Not all of the singular traces on the weak trace class ideal are Dixmier traces.^[6]: 316

Lidskii formulation

The trace of a square matrix is the sum of its eigenvalues. Lidskii's formula extends this result to functional analysis and states that the trace of a trace class operator A is given by the sum of its eigenvalues,^[13]

${\displaystyle {\rm {Tr}}(A)=\sum _{n=0}^{\infty }\lambda (n,A)=\sum (\lambda (A)).}$

The characterisation (1) of a trace φ on positive operators of a two-ideal J as a symmetric functional applied to singular values can be improved to the statement that the trace φ on any operator in J is given by the same symmetric functional applied to eigenvalue sequences, provided that the eigenvalues of all operators in J belong to the Calkin sequence space j.^[14] In particular, if a bounded operator A belongs to J whenever there is a bounded operator B in J such that

${\displaystyle \prod _{k=0}^{n}\mu (k,A)\leq \prod _{k=0}^{n}\mu (k,B)}$

(2)

for every natural number n, then for each trace φ on J there is a unique symmetric functional f on the Calkin space j with

${\displaystyle \varphi (A)={\rm {f}}(\lambda (A))}$

(3)

where λ(A) is the sequence of eigenvalues of an operator A in J rearranged so that the absolute value of the eigenvalues is decreasing. If A is quasi-nilpotent then λ(A) is the zero sequence. Most two-sided ideals satisfy the property (2), including all Banach ideals and quasi-Banach ideals.

Equation (3) is the precise statement that singular traces measure asymptotic spectral behaviour of operators.

Fredholm formulation

The trace of a square matrix is the sum of its diagonal elements. In functional analysis the corresponding formula for trace class operators is

${\displaystyle {\rm {Tr}}(A)=\sum _{n=0}^{\infty }\langle Ae_{n},e_{n}\rangle =\sum (\{\langle Ae_{n},e_{n}\rangle \}_{n=0}^{\infty })}$

where { e_n }_n=0^∞ is an arbitrary orthonormal basis of the separable Hilbert space H. Singular traces do not have an equivalent formulation for arbitrary bases. Only when φ(A)=0 will an operator A generally satisfy

${\displaystyle \varphi (A)={\rm {f}}(\{\langle Ae_{n},e_{n}\rangle \}_{n=0}^{\infty })}$

for a singular trace φ and an arbitrary orthonormal basis { e_n }_n=0^∞ .^[6]: 242

The diagonal formulation is often used instead of the Lidskii formulation to calculate the trace of products, since eigenvalues of products are hard to determine. For example, in quantum statistical mechanics the expectation of an observable S is calculated against a fixed trace-class energy density operator T by the formula

${\displaystyle \langle S\rangle ={\rm {Tr}}(ST)=\sum _{n=0}^{\infty }\langle Se_{n},e_{n}\rangle \lambda (n,T)=v_{T}(\{\langle Se_{n},e_{n}\rangle \}_{n=0}^{\infty })}$

where v_T belongs to (l_∞)_* ≅ l₁. The expectation is calculated from the expectation values ⟨Se_n, e_n⟩ and the probability ⟨P_n⟩ = λ(n,T) of the system being in the bound quantum state e_n. Here P_n is the projection operator onto the one-dimensional subspace spanned by the energy eigenstate e_n. The eigenvalues of the product, λ(n,ST), have no equivalent interpretation.

There are results for singular traces of products.^[15] For a product ST where S is bounded and T is selfadjoint and belongs to a two-sided ideal J then

${\displaystyle \varphi (ST)={\rm {f}}(\{\langle Se_{n},e_{n}\rangle \lambda (n,T)\}_{n=0}^{\infty })=v_{\varphi ,T}(\{\langle Se_{n},e_{n}\rangle \}_{n=0}^{\infty })}$

for any trace φ on J. The orthonormal basis { e_n }_n=0^∞ must be ordered so that Te_n = μ(n,T)e_n, n=0,1,2... . When φ is singular and φ(T)=1 then v_φ,T is a linear functional on l_∞ that extends the limit at infinity on the convergent sequences c. The expectation ⟨S⟩ = φ(ST) in this case has the property that ⟨P_n⟩= 0 for each n, or that there is no probability of being in a bound quantum state. That

${\displaystyle \langle S\rangle ={\text{``limit at infinity''}}\langle Se_{n},e_{n}\rangle }$

has led to a link between singular traces, the correspondence principle, and classical limits,.^{[6]: ch 12}

Ideals without traces

• Bounded operators. Paul Halmos showed in 1954 that every bounded operator on a separable infinite-dimensional Hilbert space is the sum of two commutators.^[17] That is, Com(B(H)) = B(H) and the co-dimension of the commutator subspace of B(H) is zero. The bounded linear operators admit no everywhere defined traces. The qualification is relevant; as a von Neumann algebra B(H) admits semifinite (strong-densely defined) traces.

Modern examination of the commutator subspace involves checking its spectral characterisation. The following ideals have no traces since the Cesàro means of positive sequences from the Calkin corresponding sequence space belong back in the sequence space, indicating that the ideal and its commutator subspace are equal.

• Compact operators. The commutator subspace Com(K(H)) = K(H) where K(H) denotes the compact linear operators. The ideal of compact operators admits no traces.
• Schatten p-ideals. The commutator subspace Com(L_p) = L_p, p > 1, where L_p denotes the Schatten p-ideal,

${\displaystyle L_{p}=\{A\in K(H):\left(\sum _{n=0}^{\infty }\mu (n,A)^{p}\right)^{\frac {1}{p}}<\infty \},}$

and μ(A) denotes the sequence of singular values of a compact operator A. The Schatten ideals for p > 1 admit no traces.

• Lorentz p-ideals or weak-L_p ideals. The commutator subspace Com(L_p,∞) = L_p,∞, p > 1, where

${\displaystyle L_{p,\infty }=\{A\in K(H):\mu (n,A)=O(n^{-{\frac {1}{p}}})\}}$

is the weak-L_p ideal. The weak-L_p ideals, p > 1, admit no traces. The weak-L_p ideals are equal to the Lorentz ideals (below) with concave function ψ(n)=n^1−1/p.

Ideals with traces

• Finite rank operators. It is checked from the spectral condition that the kernel of the operator trace Tr and the commutator subspace of the finite rank operators are equal, ker Tr = Com(F(H)). It follows that the commutator subspace Com(F(H)) has co-dimension 1 in F(H). Up to scaling Tr is the unique trace on F(H).
• Trace class operators. The trace class operators L₁ have Com(L₁) strictly contained in ker Tr. The co-dimension of the commutator subspace is therefore greater than one, and is shown to be infinite.^[18] Whilst Tr is, up to scaling, the unique continuous trace on L₁ for the norm ||A||₁ = Tr(|A|), the ideal of trace class operators admits infinitely many linearly independent and non-trivial singular traces.

• Weak trace class operators. Since Com(L_1,∞)₊ = (L₁)₊ the co-dimension of the commutator subspace of the weak-L₁ ideal is infinite. Every trace on weak trace class operators vanishes on trace class operators, and hence is singular. The weak trace class operators form the smallest ideal where every trace on the ideal must be singular.^[18] Dixmier traces provide an explicit construction of traces on the weak trace class operators.

${\displaystyle {\rm {Tr}}_{\omega }(A)=\omega \left(\left\{{\frac {1}{\log(1+n)}}\sum _{k=0}^{n}\lambda (k,A)\right\}_{n=0}^{\infty }\right),\quad A\in L_{1,\infty }.}$

This formula is valid for every weak trace class operator A and involves the eigenvalues ordered in decreasing absolute value. Also ω can be any extension to l_∞ of the ordinary limit, it does not need to be dilation invariant as in Dixmier's original formulation. Not all of the singular traces on the weak trace class ideal are Dixmier traces.^[6]: 316

• k-tensor weak trace class ideals. The weak-L_p ideals, p > 1, admit no traces as explained above. They are not the right setting for higher order factorisations of the traces on the weak trace class ideal L_1,∞. For a natural number k ≥ 1 the ideals

${\displaystyle E_{\otimes k}=\{A\in K(H):\mu (n,A)=O(\log ^{k-1}(n)/n)\}}$

form the appropriate setting. They have commutator subspaces of infinite co-dimension that form a chain such that E_⊗k-1 ⊂ Com(E_⊗k) (with the convention that E₀ = L₁). Dixmier traces on E_⊗k have the form

${\displaystyle {\rm {Tr}}_{\omega }^{k}(A)=\omega \left(\left\{{\frac {1}{\log ^{k}(1+n)}}\sum _{j=0}^{n}\lambda (j,A)\right\}_{n=0}^{\infty }\right),\quad A\in E_{\otimes k}.}$

• Lorentz ψ-ideals. The natural setting for Dixmier traces is on a Lorentz ψ-ideal for a concave increasing function ψ : [0,∞) → [0,∞),

${\displaystyle L_{\psi }=\{A\in K(H):{\frac {1}{\psi (1+n)}}\sum _{j=0}^{n}\mu (n,A)<\infty \}.}$

There are some ω that extend the ordinary limit to l_∞ such that

${\displaystyle {\rm {Tr}}_{\omega }^{\psi }(A)=\omega \left(\left\{{\frac {1}{\psi (1+n)}}\sum _{j=0}^{n}\lambda (j,A)\right\}_{n=0}^{\infty }\right),\quad A\in L_{\psi }}$

is a singular trace if and only if^[6]: 225

${\displaystyle \liminf _{n\to \infty }{\frac {\psi (2n)}{\psi (n)}}=1.}$

The principal ideal generated by any compact operator A with μ(A)=ψ' is called the 'small ideal' inside L_ψ. The k-tensor weak trace class ideal is the small ideal inside the Lorentz ideal with ψ=log^k.

• Fully symmetric ideals generalise Lorentz ideals. Dixmier traces form all the fully symmetric traces on a Lorentz ideal up to scaling, and form a weak* dense subset of the fully symmetric traces on a general fully symmetric ideal. It is known the fully symmetric traces are a strict subset of the positive traces on a fully symmetric ideal.^[6]: 109 Therefore, Dixmier traces are not the full set of positive traces on Lorentz ideals.