Conditional Mutual Information and Quantum Markov Structure

Conditional mutual information answers an ordered question: after the mediator $B$ is retained, how much correlation between $A$ and $C$ remains? In finite dimensions, its exact vanishing is equivalent both to perfect recovery of $C$ from $B$ and to a precise direct-sum factorization of the state. A small positive value instead controls the accuracy of recovery; by itself it does not identify a nearby exact Markov state or a memoryless process in time.

Required background. Use Strong Subadditivity and Entropic Inequalities for the entropy inequality. Helpful background. Mutual Information and Regulator-Independent Correlations fixes the correlation language, while Markov generators and semigroups explains the different dynamical meaning of “Markov.”

The chapter’s task-comparison table separates this structural question from the extra metric, channel, locality, and energy data needed in a recovery problem.

Conditional correlation with an ordered mediator

For a density operator $\rho_{ABC}$ on a finite-dimensional tripartite Hilbert space, define

$$I(A{:}C\mid B)_\rho =S(AB)+S(BC)-S(B)-S(ABC).$$

Strong subadditivity gives $I(A{:}C\mid B)_\rho\geq0$. Two chain-rule forms explain the meaning of the quantity:

$$\begin{aligned} I(A{:}C\mid B)_\rho &=S(A\mid B)_\rho-S(A\mid BC)_\rho,\\ &=I(A{:}BC)_\rho-I(A{:}B)_\rho. \end{aligned}$$

Thus access to $C$ cannot reduce what an observer holding $B$ can learn about $A$. The quantity is symmetric under $A\leftrightarrow C$, but changing the conditioning system changes the question. In particular, $I(A{:}C\mid B)$ and $I(B{:}C\mid A)$ need not agree.

If $B$ is a classical register,

$$\rho_{ABC}=\sum_b p_b\,\rho^{(b)}_{AC}\otimes |b\rangle\!\langle b|_B,$$

then

$$I(A{:}C\mid B)_\rho =\sum_b p_b I(A{:}C)_{\rho^{(b)}}.$$

For a genuinely quantum $B$, there is no basis-independent ensemble of conditional states with this interpretation. The diagram below places the exact statement and the later recovery question on their respective branches.

$I(A{:}C\mid B)$ measures the correlation left between $A$ and $C$ when $B$ is retained. Exact zero has a structural theorem; a small positive value supports a separate quantitative recovery statement. Schematic.

Exact quantum Markov structure

For an arbitrary finite-dimensional state $\rho_{ABC}$, with no full-rank assumption, the following statements are equivalent:

1. $I(A{:}C\mid B)_\rho=0$.
2. There is a completely positive trace-preserving map $\mathcal R_{B\to BC}$ such that $$\rho_{ABC} =(\operatorname{id}_A\otimes\mathcal R_{B\to BC})(\rho_{AB}).$$
3. The Hilbert space of the mediator has a direct-sum tensor decomposition $$\mathcal H_B =\bigoplus_j \mathcal H_{b_j^L}\otimes\mathcal H_{b_j^R}$$ for which $$\rho_{ABC} =\bigoplus_j p_j\, \rho^{(j)}_{A b_j^L}\otimes \rho^{(j)}_{b_j^R C}.$$

This is the Hayden–Jozsa–Petz–Winter structure theorem Hayden et al. 2004, Eqs. (10)–(11) and Theorem 6. The label $j$ is classical information stored nondestructively in $B$. Once $j$ is fixed, $B$ splits into a left part that can correlate with $A$ and a right part that can correlate with $C$; there is no remaining correlation across $A b_j^L\mid b_j^R C$.

The block-entropy identity

$$S\!\left(\bigoplus_j p_j\sigma_j\right) =H(p)+\sum_j p_jS(\sigma_j)$$

makes sufficiency transparent: inserting the four reduced HJPW states into the definition of conditional mutual information cancels the four copies of $H(p)$ and every within-block entropy. The recovery implication is also short. Exact recovery and data processing imply

$$I(A{:}BC)_\rho \leq I(A{:}B)_{\rho_{AB}},$$

because $\mathcal R_{B\to BC}$ acts only on $B$. Since the difference of the two sides in the opposite order is $I(A{:}C\mid B)_\rho\geq0$, equality follows. The nontrivial direction from entropy equality to recovery is the equality case of relative-entropy monotonicity.

A useful corollary concerns pure states. If $\rho_{ABC}=|\psi\rangle\!\langle\psi|$ and $I(A{:}C\mid B)=0$, orthogonality of the direct-sum sectors permits only one nonzero $p_j$, and purity of the product block gives

$$|\psi\rangle_{ABC} =|\alpha\rangle_{A b^L}\otimes|\gamma\rangle_{b^R C}$$

on the occupied support of $B$.

Recovery on a rank-deficient support

Let $P_B$ project onto $\operatorname{supp}\rho_B$. One exact recovery witness is the Petz map

$$\mathcal R^{\rm P}_{B\to BC}(X) =\rho_{BC}^{1/2} \left[ \left(\rho_B^{-1/2}X\rho_B^{-1/2}\right)\otimes I_C \right] \rho_{BC}^{1/2},$$

where $\rho_B^{-1/2}$ is the Moore–Penrose inverse: it is the ordinary inverse on $P_B\mathcal H_B$ and zero on the kernel. The map is completely positive and trace preserving for inputs supported on $P_B$, because

$$\operatorname{Tr}\mathcal R^{\rm P}(P_BXP_B) =\operatorname{Tr}(P_BXP_B).$$

To obtain a channel on all of $\mathcal B(\mathcal H_B)$, choose any state $\tau_{BC}$ and set

$$\mathcal R_{\rm ext}(X) =\mathcal R^{\rm P}(P_BXP_B) +\operatorname{Tr}[(I_B-P_B)X]\,\tau_{BC}.$$

This extension is completely positive and trace preserving. Its arbitrary action outside the support cannot affect recovery of $\rho_{ABC}$, because $\rho_{AB}$ has no weight there. Equivalently, the HJPW decomposition can be restricted to the occupied support; unused orthogonal summands carry zero probability.

This support-qualified formula is the specialization of the equality condition for relative-entropy data processing to $\operatorname{Tr}_C$ Hayden et al. 2004, Theorem 3 and Eqs. (8), (10)–(11).

For a faithful state one may additionally write the operator equality

$$\log\rho_{ABC}+\log\rho_B =\log\rho_{AB}+\log\rho_{BC},$$

with identity operators on omitted factors understood. This logarithmic form should not be applied to a rank-deficient state without an explicit support restriction. For the operator-algebraic theorem and conditional-expectation formulation, see Sufficiency, Conditional Expectations, and Petz Recovery.

The word “Markov” here describes conditional independence in one state. It does not assert that the state arose from a time-local semigroup or that its dynamics has no memory.

What a small value does and does not imply

With natural logarithms and root fidelity $F(\rho,\sigma)=\lVert\sqrt\rho\sqrt\sigma\rVert_1$, a small conditional mutual information guarantees a channel satisfying

$$F\!\left( \rho_{ABC}, (\operatorname{id}_A\otimes\mathcal R_{B\to BC})(\rho_{AB}) \right) \geq e^{-I(A{:}C\mid B)_\rho/2}.$$

This is the Fawzi–Renner recovery theorem Fawzi and Renner 2015, Theorem 5.1. It is a statement about closeness to a recovered state, not necessarily to an exact Markov state. If

$$\sigma_{ABC} =(\operatorname{id}_A\otimes\mathcal R)(\rho_{AB}),$$

then generally

$$\sigma_{AB} =(\operatorname{id}_A\otimes\operatorname{Tr}_C\!\circ\mathcal R)(\rho_{AB}) \neq\rho_{AB}.$$

Therefore the same map has not reconstructed $\sigma_{ABC}$ from its own $AB$ marginal, and $I(A{:}C\mid B)_\sigma$ need not vanish. This distinction is genuinely quantum: if

$$\Delta_{\rm M}(\rho) =\inf_{\mu:\,I(A{:}C\mid B)_\mu=0}D(\rho\Vert\mu),$$

then $\Delta_{\rm M}(\rho)\geq I(A{:}C\mid B)_\rho$, but quantum examples make the ratio arbitrarily large and can require a logarithmic dimension factor Ibinson, Linden, and Winter 2008, Theorem 4 and Examples 6–7.

Recovery handoff. Recovery and Approximate Markovianity develops the error bounds, metrics, and physical qualifications.

Map handoff. Petz, Rotated, and Universal Recovery Maps compares explicit recovery channels and their support and modular properties.

Neither a small dimensionless number nor high fidelity alone makes the channel local, energy limited, state independent, or experimentally implementable.

Contiguous equal-time intervals in a 1+1-dimensional CFT

Consider a unitary nonchiral CFT on the infinite line, use natural logarithms and units with propagation speed one, and denote the central charge by $c_{\rm CFT}$. At equal time, let

$$A=[x_0,x_1],\qquad B=[x_1,x_2],\qquad C=[x_2,x_3]$$

have lengths $a$, $b$, and $d$, respectively. The letter $d$ avoids confusing the length of region $C$ with the central charge. Each region in the entropy combination—$AB$, $BC$, $B$, and $ABC$—is a single interval, so no disjoint-interval twist-field correlator is needed.

Vacuum

For one interval of length $\ell$,

$$S_0(\ell) =\frac{c_{\rm CFT}}{3}\ln\frac{\ell}{\epsilon}+s_1.$$

Using the same cutoff $\epsilon$ and endpoint prescription in all four terms gives

$$I_0(A{:}C\mid B) =\frac{c_{\rm CFT}}{3} \ln\frac{(a+b)(b+d)}{b(a+b+d)}.$$

Both the cutoff and the nonuniversal constant $s_1$ cancel. The identity

$$(a+b)(b+d)-b(a+b+d)=ad$$

shows that the logarithm is positive whenever $a,b,d>0$. Equivalently, with

$$\eta=\frac{ad}{(a+b)(b+d)},$$

one has

$$I_0(A{:}C\mid B) =-\frac{c_{\rm CFT}}{3}\ln(1-\eta).$$

The endpoint and thick-mediator limits are

$$I_0\longrightarrow0 \quad\text{as}\quad a\to0\ \text{or}\ d\to0,$$ $$I_0 =\frac{c_{\rm CFT}}{3}\frac{ad}{b^2} +O(b^{-3}) \quad (b\to\infty),$$

and

$$I_0 =\frac{c_{\rm CFT}}{3} \ln\frac{ad}{b(a+d)}+o(1) \quad (b\to0^+).$$

For $c_{\rm CFT}>0$, ordinary nondegenerate equal-time intervals therefore do not saturate the common-regulator entropy form of strong subadditivity; this is the regulated sense in which they are non-Markov here. The divergence as $b\to0^+$ is the singular limit in which the separated outer intervals become adjacent Calabrese and Cardy 2004, Eq. (16) and its $n\to1$ derivative.

Thermal state

At inverse temperature $\beta$,

$$S_\beta(\ell) =\frac{c_{\rm CFT}}{3} \ln\left[ \frac{\beta}{\pi\epsilon} \sinh\left(\frac{\pi\ell}{\beta}\right) \right]+s_1.$$

Set $x=\pi a/\beta$, $y=\pi b/\beta$, and $z=\pi d/\beta$. Direct substitution gives

$$I_\beta(A{:}C\mid B) =\frac{c_{\rm CFT}}{3} \ln\frac{\sinh(x+y)\sinh(y+z)} {\sinh y\,\sinh(x+y+z)}.$$

The hyperbolic identity

$$\sinh(x+y)\sinh(y+z) -\sinh y\,\sinh(x+y+z) =\sinh x\,\sinh z$$

proves positivity and gives the numerically stable form

$$I_\beta =\frac{c_{\rm CFT}}{3} \ln\left[ 1+\frac{\sinh x\,\sinh z} {\sinh y\,\sinh(x+y+z)} \right].$$

It passes three useful checks:

$$\lim_{\beta\to\infty}I_\beta=I_0,$$ $$I_\beta =I_0-\frac{c_{\rm CFT}\pi^2ad}{9\beta^2} +O(\beta^{-4}) \quad (\beta\to\infty),$$

and, for $b\gg\beta$ with $a,d$ fixed,

$$I_\beta \sim\frac{c_{\rm CFT}}{3} (1-e^{-2\pi a/\beta}) (1-e^{-2\pi d/\beta}) e^{-2\pi b/\beta}.$$

At fixed positive $a,b,d$, the high-temperature limit is the special case

$$I_\beta \sim\frac{c_{\rm CFT}}{3}e^{-2\pi b/\beta} \quad (\beta\to0).$$

Shrinking either outer interval sends $I_\beta$ to zero. Shrinking the mediator instead produces the adjacent-interval divergence

$$I_\beta =\frac{c_{\rm CFT}}{3} \ln\left[ \frac{\sinh x\,\sinh z} {y\,\sinh(x+z)} \right]+o(1) \quad (b\to0^+).$$

The vacuum mediator suppresses the conditional correlation algebraically, as $b^{-2}$; a thermal correlation length changes this to exponential suppression. The reproducible benchmark record evaluates both formulas with stable logarithms and checks positivity and the large-$\beta$ limit. For example, with $c_{\rm CFT}=1$, $a=d=1$, and $b=2$,

$$I_0=\frac13\ln\frac98\approx0.039261,$$

whereas at $\beta=4$ the exact thermal expression gives

$$I_{\beta=4}\approx0.0093323.$$

These single-interval vacuum and thermal formulas follow from Calabrese and Cardy 2004, Eqs. (16) and (18).

Ordering and regulator stress tests

Two simple adversarial tests prevent false Markov conclusions.

Reorder the mediator. Let $A$ and $B$ be independent fair classical bits and let $C$ be a copy of $B$:

$$p(a,b,c)=\frac14\,\delta_{c,b}.$$

Conditioned on $B$, the variables $A$ and $C$ are independent, so

$$I(A{:}C\mid B)=0.$$

Conditioned on $A$, however, $B$ and $C$ remain identical fair bits, so

$$I(B{:}C\mid A)=1\ \text{bit}.$$

For the spatial CFT arrangement, the same illicit reorder replaces the four single-interval entropies by

$$I(B{:}C\mid A) =S(AB)+S(AC)-S(A)-S(ABC),$$

where $AC$ is disconnected. The preceding single-interval formula therefore cannot be reused.

Mismatch the regulators. If the four CFT entropies use cutoffs $\epsilon_{AB}$, $\epsilon_{BC}$, $\epsilon_B$, and $\epsilon_{ABC}$, the reported result is shifted by

$$\delta I =\frac{c_{\rm CFT}}{3} \ln\frac{\epsilon_B\epsilon_{ABC}} {\epsilon_{AB}\epsilon_{BC}}.$$

For the numerical vacuum example above, taking

$$\epsilon_{AB}=\epsilon_{BC}=\epsilon_{ABC}=\epsilon, \qquad \epsilon_B=\frac89\epsilon$$

adds $-(1/3)\ln(9/8)$ and falsely reports $I=0$. Taking $\epsilon_B<8\epsilon/9$ even fabricates a negative value. A common cutoff and endpoint prescription are therefore part of the statement, not optional numerical details.

The failure map summarizes what must remain fixed when the exact theorem or an approximate recovery bound is applied.

Exact and approximate Markov statements concern one ordered tripartite state. Reordering the mediator, changing support, mixing regulator prescriptions, or postselecting a different state changes the conditional information and its associated recovery problem. Schematic.

Continuum scope and null Markov geometry

In a continuum QFT, the four individual entropies may be divergent, and local algebras are generally type III rather than matrix tensor factors. One must not interpret a formal $\infty+\infty-\infty-\infty$ as conditional mutual information. A common regulator can define the combination in a controlled limit if convergence is proved; split constructions can supply type-I intermediates but still require a separate finiteness argument. Alternatively, one can work directly with relative entropy and modular inclusions.

For separable type-I tensor-product systems, conditional mutual information has a lower-semicontinuous extension based on finite-rank truncations, and the recovery theorem survives even when marginal entropies are infinite Shirokov 2016, Theorem 2 and Proposition 13. This does not turn the finite-dimensional HJPW direct sum into a classification of type-III local algebras.

There is nevertheless a special continuum vacuum Markov property. For causal regions $R_1$ and $R_2$ whose future horizons are cuts of the same null plane—and, in a CFT, for their conformal images on a common null cone—the modular Hamiltonians obey an inclusion–exclusion identity. With one compatible null-surface regulator, the corresponding entropy combination saturates strong subadditivity:

$$S(R_1)+S(R_2) -S(R_1\cap R_2)-S(R_1\cup R_2)=0.$$

This is the null-surface result of Casini, Testé, and Torroba 2017, Eqs. (1.9)–(1.10) and §7. It concerns a special family of null cuts. It does not contradict the strictly positive equal-time interval formulas above, and it should not be translated into an HJPW factorization without a type-I regulator.

Common pitfalls

Adding a faithfulness hypothesis to the exact theorem. Exact finite-dimensional recovery and the HJPW structure hold for rank-deficient states. Faithfulness is needed only for inverse or logarithmic formulas written without support qualifiers.

Calling the recovered approximation an exact Markov state. A channel may recover a state close to $\rho_{ABC}$ while changing its own $AB$ marginal. Recoverability and distance to the HJPW set are different optimization problems.

Treating a spatial statement as dynamics. Conditional independence of ordered subalgebras says nothing by itself about a Lindblad generator, temporal divisibility, or memory kernels.

Subtracting continuum entropies term by term. The finite answer depends on a common limiting prescription. Region order, endpoints, state, and regulator must be held fixed throughout the combination.

Exercises

1. A classical sector stored in the mediator

Let

$$\rho_{ABC} =\sum_jp_j\, \rho_A^{(j)}\otimes|j\rangle\!\langle j|_B \otimes\rho_C^{(j)}.$$

Show that $I(A{:}C\mid B)=0$ and construct an exact recovery channel.

Solution

For each reduced state, use

$$S\!\left(\sum_jp_j|j\rangle\!\langle j|\otimes\sigma_j\right) =H(p)+\sum_jp_jS(\sigma_j).$$

The four entropies are

$$\begin{aligned} S(ABC)&=H(p)+\sum_jp_j[S(\rho_A^{(j)})+S(\rho_C^{(j)})],\\ S(AB)&=H(p)+\sum_jp_jS(\rho_A^{(j)}),\\ S(BC)&=H(p)+\sum_jp_jS(\rho_C^{(j)}),\\ S(B)&=H(p). \end{aligned}$$

They cancel in $I(A{:}C\mid B)$. A channel can first dephase in the $|j\rangle$ basis and then append $\rho_C^{(j)}$:

$$\mathcal R(X) =\sum_j\langle j|X|j\rangle\, |j\rangle\!\langle j|_B\otimes\rho_C^{(j)}.$$

This is completely positive and trace preserving and recovers the stated $\rho_{ABC}$.

2. Why the support extension is trace preserving

Let $P_B=\operatorname{supp}\rho_B$. Verify that $\mathcal R^{\rm P}$ preserves trace on $P_B\mathcal H_B$, and then verify that $\mathcal R_{\rm ext}$ preserves trace on all inputs.

Solution

For $X=P_BXP_B$, cyclicity of the trace and $\operatorname{Tr}_C\rho_{BC}=\rho_B$ give

$$\begin{aligned} \operatorname{Tr}\mathcal R^{\rm P}(X) &=\operatorname{Tr}\left[ \rho_B^{-1/2}X\rho_B^{-1/2}\rho_B \right]\\ &=\operatorname{Tr}(XP_B) =\operatorname{Tr}X. \end{aligned}$$

For arbitrary $X$,

$$\begin{aligned} \operatorname{Tr}\mathcal R_{\rm ext}(X) &=\operatorname{Tr}(P_BXP_B) +\operatorname{Tr}[(I_B-P_B)X]\\ &=\operatorname{Tr}X. \end{aligned}$$

Both summands in the extension are completely positive maps, so the extension is CPTP.

3. Vacuum positivity and the large-separation law

Starting from the vacuum interval entropy, derive $I_0(A{:}C\mid B)$ and show that it is positive. Then obtain the first nonzero term for $b\gg a,d$.

Solution

Substitution cancels $s_1$ and $\epsilon$ and yields

$$I_0 =\frac{c_{\rm CFT}}{3} \ln\frac{(a+b)(b+d)}{b(a+b+d)}.$$

The ratio equals

$$1+\frac{ad}{b(a+b+d)}>1,$$

so $I_0>0$. At large $b$,

$$\frac{ad}{b(a+b+d)} =\frac{ad}{b^2}+O(b^{-3}).$$

Using $\ln(1+u)=u+O(u^2)$ gives

$$I_0 =\frac{c_{\rm CFT}}{3}\frac{ad}{b^2} +O(b^{-3}).$$

4. Thermal screening

Use the thermal interval formula to prove $I_\beta>0$, recover the vacuum as $\beta\to\infty$, and find the leading behavior for $b\gg\beta$.

Solution

With $x=\pi a/\beta$, $y=\pi b/\beta$, and $z=\pi d/\beta$, the hyperbolic identity on this page rewrites the entropy combination as

$$I_\beta =\frac{c_{\rm CFT}}{3} \ln\left[ 1+\frac{\sinh x\sinh z} {\sinh y\sinh(x+y+z)} \right]>0.$$

When $\beta\to\infty$, all three arguments are small and $\sinh u\sim u$, which recovers the vacuum length ratio. For $y\gg1$,

$$\frac{\sinh x\sinh z} {\sinh y\sinh(x+y+z)} \sim (1-e^{-2x})(1-e^{-2z})e^{-2y}.$$

This ratio is small, so expanding the logarithm gives

$$I_\beta \sim\frac{c_{\rm CFT}}{3} (1-e^{-2\pi a/\beta}) (1-e^{-2\pi d/\beta}) e^{-2\pi b/\beta}.$$

References

• Calabrese, Pasquale, and John Cardy. “Entanglement Entropy and Quantum Field Theory.” Journal of Statistical Mechanics: Theory and Experiment 2004, no. 6 (2004): P06002. DOI. Open preprint.
• Casini, Horacio, Eduardo Testé, and Gonzalo Torroba. “Modular Hamiltonians on the Null Plane and the Markov Property of the Vacuum State.” Journal of Physics A: Mathematical and Theoretical 50, no. 36 (2017): 364001. DOI. Open preprint.
• Fawzi, Omar, and Renato Renner. “Quantum Conditional Mutual Information and Approximate Markov Chains.” Communications in Mathematical Physics 340, no. 2 (2015): 575–611. DOI. Open preprint.
• Hayden, Patrick, Richard Jozsa, Dénes Petz, and Andreas Winter. “Structure of States Which Satisfy Strong Subadditivity of Quantum Entropy with Equality.” Communications in Mathematical Physics 246, no. 2 (2004): 359–374. DOI. Open preprint.
• Ibinson, Ben, Noah Linden, and Andreas Winter. “Robustness of Quantum Markov Chains.” Communications in Mathematical Physics 277, no. 2 (2008): 289–304. DOI. Open preprint.
• Shirokov, M. E. “Measures of Quantum Correlations in Infinite-Dimensional Systems.” Sbornik: Mathematics 207, no. 5 (2016): 724–768. DOI. Open preprint.