Tomita–Takesaki Modular Operators and Flow

Tomita–Takesaki theory says that a von Neumann algebra and a cyclic, separating vector determine a canonical one-parameter automorphism group. The construction requires neither a chosen Hamiltonian nor a trace. In QFT it is the rigorous origin of modular flow and its KMS property; a spacetime interpretation requires additional hypotheses.

Required background. The operator-algebra bridge supplies von Neumann algebras and commutants; modular Hamiltonian domains supplies the spectral logarithm and support conventions.

The chapter’s structure map shows what the standard pair determines intrinsically. The comparison table and validity guide separate those conclusions from wedge and conformal special cases.

From the Tomita map to polar decomposition

Let $\mathcal A\subset B(\mathcal H)$ and let $\Omega$ be cyclic and separating for $\mathcal A$. Cyclicity makes $\mathcal A\Omega$ dense. Separation makes the antilinear rule

$$S_0(A\Omega)=A^*\Omega, \qquad A\in\mathcal A,$$

well defined: $A\Omega=0$ implies $A=0$. The operator $S_0$ is closable; write $S$ for its closure. On its natural core, $S^2\subset\mathbf1$. Its polar decomposition is

$$S=J\Delta^{1/2},$$

where $J$ is antiunitary and $\Delta=S^*S$ is positive and self-adjoint. One also has

$$J^2=\mathbf1, \qquad J\Omega=\Omega, \qquad \Delta\Omega=\Omega.$$

The Tomita–Takesaki theorem gives

$$J\mathcal A J=\mathcal A', \qquad \Delta^{is}\mathcal A\Delta^{-is}=\mathcal A, \quad s\in\mathbb R.$$

Takesaki’s original treatment develops $\Delta$, the commutation theorem, and its automorphism group in Takesaki 1970, “The Modular Operator $\Delta$” through “The One-Parameter Automorphisms Defined by $\Delta$,” pp. 31–52. Thus

$$\sigma_s^\Omega(A)=\Delta^{is}A\Delta^{-is}$$

is a strongly continuous group of $*$-automorphisms of $\mathcal A$. With $K=-\log\Delta$,

$$\sigma_s(A)=e^{-isK}Ae^{isK}.$$

Both signs for modular time occur in the literature. A formula is unambiguous only when the definitions of $K$ and $\sigma_s$ are stated together.

A thermofield-double matrix benchmark

Represent $M_n(\mathbb C)$ on the Hilbert space of Hilbert–Schmidt matrices with

$$\langle X,Y\rangle_{\mathrm{HS}}=\operatorname{Tr}(X^\dagger Y).$$

Let the algebra act by left multiplication and choose $\Omega_\rho=\rho^{1/2}$ for a faithful density matrix $\rho=\sum_i p_i\lvert i\rangle\langle i\rvert$. Then

$$\Delta(X)=\rho X\rho^{-1}, \qquad J(X)=X^\dagger, \qquad S(X)=\rho^{-1/2}X^\dagger\rho^{1/2}.$$

On matrix units,

$$\Delta^{1/2}(E_{ij}) =\sqrt{\frac{p_i}{p_j}}\,E_{ij},$$

and applying $J$ gives $S(E_{ij})=\sqrt{p_i/p_j}\,E_{ji}$. This verifies $S=J\Delta^{1/2}$ without suppressing the right action.

Vectorize a matrix by

$$\lvert X\rangle\!\rangle =\sum_{ij}X_{ij}\lvert i\rangle_L\lvert j\rangle_R.$$

Left multiplication becomes $A\otimes\mathbf1$, right multiplication by $B$ becomes $\mathbf1\otimes B^T$, and

$$\lvert\Omega_\rho\rangle\!\rangle =\sum_i\sqrt{p_i}\,\lvert i\rangle_L\lvert i\rangle_R$$

is the canonical purification of the standard vector. When $\rho=Z^{-1}e^{-\beta H}$, it is the thermofield-double state; for a general faithful $\rho$, the same formula is best called the canonical purification. Witten 2018, § IV.A, pp. 17–20, eqs. (4.1)–(4.32) gives the finite construction in bipartite notation.

For a reproducible qubit benchmark, take

$$\rho=\operatorname{diag}(3/4,1/4).$$

Then

$$\operatorname{spec}(\Delta)=\{1,3,1/3\}, \qquad \operatorname{spec}(-\log\Delta)=\{0,-\log3,\log3\},$$

with multiplicities inherited from the four matrix units. All domains are the full four-dimensional Hilbert–Schmidt space, and

$$\sigma_s(E_{01})=3^{is}E_{01}.$$

If each $p_i$ is known within $\eta<\delta$, where $p_i\geq\delta$, the uncertainty in a modular eigenvalue obeys

$$\left\lvert\delta\log\frac{p_i}{p_j}\right\rvert \leq\frac{2\eta}{\delta-\eta}.$$

Here $\delta=1/4$. The finite benchmark is therefore controlled by the probability tolerance $\eta$ and the distance $\delta$ from loss of faithfulness.

KMS characterization and its analytic domain

Let $A$ and $B$ first be entire analytic elements for $\sigma_s$. Then

$$F_{A,B}(z)=\langle\Omega,A\Delta^{iz}B\Omega\rangle$$

is analytic for $-1<\operatorname{Im}z<0$, continuous on the boundary, and satisfies

$$\langle\Omega,A\,\sigma_{s-i}(B)\Omega\rangle =\langle\Omega,\sigma_s(B)A\Omega\rangle.$$

For arbitrary bounded $A,B$, the KMS statement is the existence of the corresponding bounded strip-analytic continuation with these boundary values. It is not a claim that the literal vector $\Delta^{iz}B\Omega$ belongs to every unbounded-operator domain. Takesaki proves the modular KMS boundary condition in Takesaki 1970, “The Modular Automorphism Group and the Kubo–Martin–Schwinger Boundary Condition,” pp. 63–72.

Reversing the sign of modular time moves the strip to $0<\operatorname{Im}z<1$. The invariant statement is a unit-width KMS condition; modular correlators develops its ordering conventions and analytic continuations.

The KMS property also fixes the parametrization of the modular group once the inverse-temperature convention is chosen. A proposed geometric flow must therefore reproduce the $2\pi$ normalization, not merely preserve the region.

Thermofield double versus a vacuum wedge

For the right Rindler wedge and the convention $\sigma_s=\operatorname{Ad}\Delta^{is}$, the Bisognano–Wichmann theorem gives

$$\Delta_W^{is}=U(\Lambda_W(-2\pi s)).$$

The original general-field theorem and its wedge-duality consequences are stated in Bisognano and Wichmann 1976, Theorem 1 and § V, pp. 303–321. The algebraic theorem alone does not imply that $\sigma_s$ moves spacetime points.

Structure	Qubit thermofield double	Vacuum wedge
Algebra–state pair	Left $M_2(\mathbb C)$ and $\rho^{1/2}$	$\mathcal A(W)$ and the vacuum
Tomita domain	Entire Hilbert–Schmidt space	Closure of the Tomita core $\mathcal A(W)\Omega$
Modular spectrum	$1,3,1/3$	Spectrum of the boost representation; generally unbounded at generator level
Flow	$\rho^{is}A\rho^{-is}$	Boost rapidity $-2\pi s$
Normalization	Fixed by the eigenvalue ratios of normalized $\rho$	Fixed by the theorem’s $2\pi$ factor; no local trace
Control	Exact matrix identities plus $(\eta,\delta)$	Vacuum, covariance, locality, spectrum condition, and field-generation hypotheses

The wedge row carries no numerical error bar: it is exact when its hypotheses hold and unlicensed when they do not. Comparing a regulator with it requires a stated cutoff sequence and convergence test; the qubit model by itself is not such an extrapolation.

Failure injection: remove separatingness

Let $\mathcal A=M_2(\mathbb C)$ act on $\mathbb C^2$ and choose $\Omega=\lvert0\rangle$. This vector is cyclic but not separating. For

$$B=\lvert0\rangle\langle1\rvert,$$

one has $B\Omega=0$ but $B^*\Omega=\lvert1\rangle$. The formal rule would demand

$$S_0(0)=\lvert1\rangle,$$

so $S_0$ is not well defined. Cyclicity alone preserves the dense set $\mathcal A\Omega$; without separatingness there is no Tomita operator from this rule and hence no $J$ or $\Delta$ obtained from its polar decomposition.

Even when $\Omega$ is separating, the closure $S$ need not act on every vector. A limit $A_n\Omega\to\xi$ belongs to $\mathcal D(S)$ only if $A_n^*\Omega$ also converges. Imaginary modular time similarly begins on analytic elements, whereas real modular time is implemented by bounded unitaries on all of $\mathcal H$.

Common pitfalls

Calling modular flow the system’s time evolution. Modular time is intrinsic to $(\mathcal A,\Omega)$. Agreement with physical or geometric time is an extra KMS or covariance statement.

Replacing closures by matrix intuition. The matrix benchmark has trivial domains. The continuum theorem concerns closed unbounded operators and cannot inherit that simplification.

Using imaginary modular time on arbitrary operators. Real-time unitaries exist globally, but $\sigma_{s-i}(A)$ requires analytic continuation.

Exercises

1. Verify $S=J\Delta^{1/2}$ for the qubit benchmark on all four matrix units, and identify the commutant action after vectorization.

Solution

For $p_0=3/4$ and $p_1=1/4$,

$$\Delta^{1/2}(E_{ij})=\sqrt{p_i/p_j}\,E_{ij}, \qquad J(E_{ij})=E_{ji}.$$

Therefore $J\Delta^{1/2}(E_{ij})=\sqrt{p_i/p_j}\,E_{ji}$, which agrees with $S(E_{ij})$. Under vectorization, $L_A$ acts as $A\otimes\mathbf1$ and $R_B$ as $\mathbf1\otimes B^T$; every right action commutes with every left action, so the right matrix algebra is the commutant.

2. In the finite standard representation, prove the $s=0$ KMS boundary identity for arbitrary matrices $A,B$ directly from cyclicity of the trace.

Solution

Since $\sigma_{-i}(B)=\rho B\rho^{-1}$,

$$\operatorname{Tr}\!\left(\rho A\sigma_{-i}(B)\right) =\operatorname{Tr}(\rho A\rho B\rho^{-1}).$$

Cyclically moving the final $\rho^{-1}$ to the front gives

$$\operatorname{Tr}(A\rho B) =\operatorname{Tr}(\rho BA),$$

which is the required boundary value $\omega(BA)$. Faithfulness is used when writing $\rho^{-1}$; the support-reduced version is required otherwise.

3. For the nonseparating example, explain why no redefinition of $S_0(0)$ repairs the problem while retaining $S_0(A\Omega)=A^*\Omega$. State the strongest surviving conclusion.

Solution

Both the zero operator and $B=\lvert0\rangle\langle1\rvert$ produce the same input vector, $0\Omega=B\Omega=0$. The rule assigns outputs $0$ and $B^*\Omega=\lvert1\rangle$, respectively. An operator must assign a unique output to each input, so no choice of $S_0(0)$ satisfies both assignments. What survives is only that $\Omega$ is cyclic and $\mathcal A\Omega$ is dense. Separation, closability of the Tomita rule, its polar decomposition, and modular flow do not follow.

References

• Bisognano, Joseph J., and Eyvind H. Wichmann. “On the Duality Condition for Quantum Fields.” Journal of Mathematical Physics 17 (1976): 303–321. DOI.
• Takesaki, Masamichi. Tomita’s Theory of Modular Hilbert Algebras and Its Applications. Lecture Notes in Mathematics 128. Berlin: Springer, 1970. DOI.
• Witten, Edward. “Notes on Some Entanglement Properties of Quantum Field Theory.” Reviews of Modern Physics 90 (2018): 045003. DOI; Accepted manuscript PDF; arXiv.