Data-Processing Inequality

Every honest theory of information carries a version of the same warning: you cannot squeeze more knowledge out of data by mangling it. The data-processing inequality is the quantum statement of that warning made exact. Take two candidate states $\rho$ and $\sigma$ , run both through one identical physical process — a measurement, a noisy wire, a unitary that couples to an unwatched environment — and the two states can only become harder to tell apart, never easier. Written with the quantum relative entropy $D(\rho\|\sigma)=\mathrm{Tr}\,\rho(\log\rho-\log\sigma)$ , this near-tautology becomes one of the deepest theorems in the subject (Lindblad 1975; Uhlmann 1977): for every completely positive trace-preserving map $\mathcal{N}$ , $D(\mathcal{N}(\rho)\|\mathcal{N}(\sigma))\le D(\rho\|\sigma)$ . The reach of that single line is what makes it a headline rather than a footnote. Because the relative entropy is the parent of nearly every information measure — von Neumann entropy, mutual information, conditional entropy, the Holevo $\chi$ , fidelity, trace distance — each of their monotonicity laws is DPI with a channel or a reference state plugged in. Mutual information cannot grow under local processing; the Holevo quantity caps accessible information; trace distance contracts; fidelity, which measures closeness rather than distinguishability, can only rise. Strong subadditivity of the von Neumann entropy is not merely implied by DPI, it is logically equivalent to it. Two pictures organize the whole article. The Stinespring picture dilates any channel into an isometry — which preserves $D$ exactly — followed by discarding an environment, so all irreversibility lives in one act of forgetting. The recovery picture (Petz 1986; Fawzi–Renner 2015) turns the inequality into a quantitative statement: the gap in DPI lower-bounds how well a recovery channel can undo the processing, so equality means perfect reversibility. DPI is thus the second-law-like backbone of quantum Shannon theory.

Key takeaways

The data-processing inequality (DPI) says no physical processing can increase distinguishability: for any quantum channel $\mathcal{N}$ , $D(\mathcal{N}(\rho)\,\|\,\mathcal{N}(\sigma))\le D(\rho\,\|\,\sigma)$ .
Because almost every information measure is built from relative entropy, DPI is the master monotonicity theorem: mutual information, conditional entropy, Holevo $\chi$ , fidelity and trace distance all inherit their monotonicity from it.
DPI is mathematically equivalent to strong subadditivity of the von Neumann entropy — each is derivable from the other.
Fidelity runs the opposite way, $F(\mathcal{N}(\rho),\mathcal{N}(\sigma))\ge F(\rho,\sigma)$ , while trace distance contracts, $\tfrac12\|\mathcal{N}(\rho)-\mathcal{N}(\sigma)\|_1\le\tfrac12\|\rho-\sigma\|_1$ — the same physical fact read two ways.
The Stinespring dilation reduces the general theorem to monotonicity under a partial trace; the equality case (Petz recovery) says exactly when processing is reversible.
Key idea: distinguishability is monotone — any CPTP map $\mathcal{N}$ can only lower the quantum relative entropy, $D(\mathcal{N}(\rho)\|\mathcal{N}(\sigma))\le D(\rho\|\sigma)$ , and every other monotonicity statement is a corollary obtained by choosing the channel or the pair of states.

The core insight

Distinguishability is monotone: any physical process $\mathcal{N}$ (a CPTP map) cannot increase the quantum relative entropy, $D(\mathcal{N}(\rho)\|\mathcal{N}(\sigma))\le D(\rho\|\sigma)$ . Every other monotonicity statement — for mutual information, conditional entropy, fidelity and trace distance — is a corollary obtained by choosing the channel or the pair of states.

Where this lives today

DPI is the quiet workhorse behind almost every converse proof in quantum information, and its frontier keeps moving. In quantum Shannon theory it is the pivot of the "entropy method": bound the message-to-decoder correlation by the correlation at the channel output, then by the input, each step a data-processing inequality, and the achievable rate collapses onto the coherent or Holevo information — this is how the coding theorems are shown optimal. In quantum key distribution it guarantees that no classical post-processing Eve performs (error correction, privacy amplification) can raise her information above what the quantum signals already leaked, and the contractivity of trace distance under her operations underpins composable security. In quantum thermodynamics it is the single-shot second law: with $\sigma$ a Gibbs state, monotonicity of $D(\rho\|\sigma)$ under thermal operations says free energy can only fall, bounding extractable work and Landauer erasure. The equality case drives modern approximate quantum error correction — the Fawzi–Renner bound certifies a near-perfect Petz recovery whenever the DPI gap is small, and the same machinery structures quantum Markov chains and conditional independence. Most strikingly, in high-energy physics the positivity and monotonicity of relative entropy have become the sharp, ultraviolet-finite replacement for entanglement entropy: the Bousso–Casini–Fisher–Maldacena program encodes the quantum null energy condition and the Bekenstein bound as data-processing statements, carrying a 1975 theorem into arguments about black-hole interiors.

The physical picture

Hand a referee two coins, one fair and one slightly biased, and let them examine each toss. Now interpose a noisy reporter who sometimes flips the reported result before passing it on. The referee's ability to decide which coin is which can only get worse — the noise blurs the two hypotheses together. DPI is the fully quantum, fully general version of that statement: passing both candidate states through the same channel never sharpens the distinction between them. The relative entropy is the exact currency of "how distinguishable," because by quantum Stein's lemma the optimal error in telling $\rho$ from $\sigma$ on $n$ copies falls like $2^{-nD(\rho\|\sigma)}$ ; a channel that lowers $D$ literally lowers that error exponent. The first figure makes the contraction concrete with real numbers.

Figure

A channel can only push two states together. A 50 percent depolarizing map sends ρ = diag(0.9, 0.1) and σ = I/2 toward the maximally mixed state, and the relative entropy falls from D = 0.531 bits to D = 0.119 bits, so D(N(ρ)‖N(σ)) = 0.119 ≤ 0.531 = D(ρ‖σ)

Where does the irreversibility actually come from? Every channel, no matter how complicated, can be built in two moves: first embed the input into a larger space with an isometry $V$ (a rotation into a system-plus-environment Hilbert space), then throw the environment away. The first move is reversible and preserves the relative entropy exactly; the second — forgetting $E$ — is the sole source of contraction. This Stinespring decomposition is the engine of the whole subject: it reduces the general DPI to the single statement that discarding a subsystem cannot increase $D$ . The second figure lays the mechanism out.

Figure

Where the irreversibility lives. Stinespring realizes any channel N as an isometry V into a larger space H_B ⊗ H_E followed by discarding the environment E. The isometry preserves relative entropy exactly, D(VρV†‖VσV†) = D(ρ‖σ), so all the contraction happens in the partial trace over E, giving D(ρ_BE‖σ_BE) ≥ D(N(ρ)‖N(σ))

Equality is the exceptional, reversible case. If processing loses no distinguishability between $\rho$ and $\sigma$ , then a single recovery channel — the Petz transpose map — exists that undoes $\mathcal{N}$ on both states at once. The amount by which DPI is strict measures how badly the process scrambles the pair beyond repair, and the modern Fawzi–Renner bound turns that intuition into a hard inequality: a small gap in DPI guarantees a good recovery.

The core equations

The master statement is one line, and its content is a direction: the divergence on the output side is never larger than on the input side. Everything else in this article is a specialization or a refinement of it.

D(\mathcal{N}(\rho)\,\|\,\mathcal{N}(\sigma)) \;\le\; D(\rho\,\|\,\sigma) \qquad (1)

where: $\rho,\sigma$ — density operators on the input space, $\rho,\sigma\ge 0$ , $\mathrm{Tr}\,\rho=\mathrm{Tr}\,\sigma=1$ , with $\mathrm{supp}\,\rho\subseteq\mathrm{supp}\,\sigma$ ; $\mathcal{N}$ — any CPTP (completely positive trace-preserving) map, the most general physical channel; $D(\rho\|\sigma)=\mathrm{Tr}[\rho(\log\rho-\log\sigma)]$ — the quantum relative entropy. Read it as "processing cannot increase distinguishability." The object being made monotone is the relative entropy itself, whose non-negativity (Klein's inequality) is the $\mathcal{N}=\mathrm{id}$ special case and the base camp for everything else.

D(\rho\,\|\,\sigma)=\mathrm{Tr}\big[\rho\,(\log\rho-\log\sigma)\big]\;\ge\;0,\qquad D=0\iff\rho=\sigma \qquad (2)

where: $\log$ — operator logarithm via spectral calculus, base 2 for bits or $e$ for nats. The ordering ( $\log\rho$ first) is exactly what makes $D\ge 0$ ; it diverges to $+\infty$ when $\rho$ has support where $\sigma$ vanishes. Isometric invariance is the second irreducible fact: reversible embeddings change nothing. Together with (2) it is all the analytic input DPI needs.

D(V\rho V^\dagger\,\|\,V\sigma V^\dagger)=D(\rho\,\|\,\sigma),\qquad V^\dagger V=\mathbb{I} \qquad (3)

where: $V$ — an isometry (a unitary is the square case $VV^\dagger=\mathbb{I}$ too); embedding faithfully preserves all eigenvalue and overlap data entering $D$ , so no distinguishability is created or destroyed by a rotation into a larger space. Applied to a bipartite state with a local channel on one side, (1) becomes the monotonicity of mutual information — local processing cannot manufacture correlations.

I(A;B)_\rho \;\ge\; I(A;B')_{(\mathrm{id}_A\otimes\mathcal{N}_{B\to B'})\rho},\qquad I(A;B)_\rho=D(\rho_{AB}\|\rho_A\otimes\rho_B) \qquad (4)

where: $I(A;B)_\rho=S(\rho_A)+S(\rho_B)-S(\rho_{AB})$ — quantum mutual information, written as a divergence from the product of marginals; $\mathcal{N}_{B\to B'}$ — a channel acting only on $B$ . Since the product reference maps to $\rho_A\otimes\mathcal{N}(\rho_B)$ , DPI of the divergence is exactly this inequality. The same monotonicity, read through the two most-used state distances, gives a contracting trace distance and — running the opposite way — a non-decreasing fidelity.

\tfrac12\,\|\mathcal{N}(\rho)-\mathcal{N}(\sigma)\|_1\;\le\;\tfrac12\,\|\rho-\sigma\|_1,\qquad F\big(\mathcal{N}(\rho),\mathcal{N}(\sigma)\big)\;\ge\;F(\rho,\sigma) \qquad (5)

where: $\|X\|_1=\mathrm{Tr}\sqrt{X^\dagger X}$ — trace norm, with $\tfrac12\|\rho-\sigma\|_1$ the trace distance in $[0,1]$ ; $F(\rho,\sigma)=\|\sqrt\rho\sqrt\sigma\|_1$ — Uhlmann fidelity, equal to $1$ iff $\rho=\sigma$ . Distinguishability measures shrink; closeness measures grow. Finally the recovery refinement makes DPI quantitative: the gap is bounded below by the fidelity with which the Petz map undoes the channel, so the inequality is tight exactly when recovery is perfect.

D(\rho\|\sigma)-D(\mathcal{N}(\rho)\|\mathcal{N}(\sigma))\;\ge\;-2\log F\big(\rho,\ \mathcal{R}_{\sigma,\mathcal{N}}(\mathcal{N}(\rho))\big)\;\ge\;0 \qquad (6)

where: $\mathcal{R}_{\sigma,\mathcal{N}}$ — the Petz recovery channel for reference $\sigma$ and channel $\mathcal{N}$ ; the right side vanishes iff $\mathcal{R}\circ\mathcal{N}$ restores $\rho$ perfectly, upgrading DPI from an inequality into an approximate-reversibility statement (Fawzi–Renner 2015).

Derivation ADPI from Stinespring dilation and monotonicity under partial trace

The cleanest proof splits the theorem into a reversible part and a single irreversible part. Realize the channel as an isometry into an enlarged space, which preserves the relative entropy exactly, followed by tracing out the environment, which can only decrease it. This reduces the general statement about arbitrary CPTP maps to the one special case of monotonicity under a partial trace — the irreducible analytic input, itself a consequence of the joint convexity of relative entropy (Lindblad 1975). Let $\mathcal{N}:\mathcal{B}(\mathcal{H}_A)\to\mathcal{B}(\mathcal{H}_B)$ be CPTP; by Stinespring there is an isometry $V:\mathcal{H}_A\to\mathcal{H}_B\otimes\mathcal{H}_E$ with $\mathcal{N}(\cdot)=\mathrm{Tr}_E[V(\cdot)V^\dagger]$ .

Step 1. Use isometric invariance, $D(V\rho V^\dagger\,\|\,V\sigma V^\dagger)=D(\rho\,\|\,\sigma)$ . Why: an isometry embeds the input support faithfully and preserves every eigenvalue and overlap entering $D$ , so the reversible half of the channel moves no distinguishability at all — it is a bookkeeping rotation.

Step 2. Name the joint states $\rho_{BE}=V\rho V^\dagger$ and $\sigma_{BE}=V\sigma V^\dagger$ , so that $\mathcal{N}(\rho)=\mathrm{Tr}_E\,\rho_{BE}$ and $\mathcal{N}(\sigma)=\mathrm{Tr}_E\,\sigma_{BE}$ . Why: this isolates the only irreversible operation in the whole construction — the partial trace over the environment — and packages the channel's action as "discard $E$ ."

Step 3. Invoke monotonicity under partial trace, $D(\rho_{BE}\,\|\,\sigma_{BE})\ge D(\mathrm{Tr}_E\rho_{BE}\,\|\,\mathrm{Tr}_E\sigma_{BE})$ . Why: this is the Lindblad–Uhlmann inequality, provable from the joint convexity of $(\rho,\sigma)\mapsto D(\rho\|\sigma)$ or from operator concavity; it is the one deep fact the argument cannot avoid.

Step 4. Chain the pieces: $D(\rho\|\sigma)=D(\rho_{BE}\|\sigma_{BE})\ge D(\mathcal{N}(\rho)\|\mathcal{N}(\sigma))$ . Why: combining the exact equality of Step 1 with the one-sided drop of Step 3 gives the master DPI, and shows all irreversibility is concentrated in forgetting the environment.

Go deeper · Derivation B · 1 minStrong subadditivity is equivalent to the data-processing inequality

Derivation BStrong subadditivity is equivalent to the data-processing inequality

Strong subadditivity (SSA) of the von Neumann entropy, $S(\rho_{ABC})+S(\rho_B)\le S(\rho_{AB})+S(\rho_{BC})$ , is not a separate theorem from DPI but the same theorem in different clothing. The goal here is to obtain SSA in a single line from the monotonicity of mutual information under a partial trace, and then to note that the implication runs both ways, so that the deepest entropy inequality in the subject and the master monotonicity law are logically interchangeable. Write mutual information across the cut $A:BC$ as a relative entropy and discard $C$ .

Step 1. Express the correlation across the larger cut as a divergence, $I(A;BC)=D(\rho_{ABC}\|\rho_A\otimes\rho_{BC})$ . Why: casting a correlation measure as a divergence from the uncorrelated product is what lets the general monotonicity theorem act on it.

Step 2. Choose the channel $\mathcal{N}=\mathrm{Tr}_C$ , a bona fide CPTP map, and apply DPI. Tracing out $C$ sends $\rho_{ABC}\mapsto\rho_{AB}$ and $\rho_A\otimes\rho_{BC}\mapsto\rho_A\otimes\rho_B$ , so $I(A;B)\le I(A;BC)$ . Why: discarding a subsystem is the most physical channel there is, and it is exactly the operation that shrinks the $A:BC$ cut to the $A:B$ cut, so "throwing away $C$ cannot increase what $A$ knows" is literal.

Step 3. Expand both mutual informations, $S(\rho_A)+S(\rho_B)-S(\rho_{AB})\le S(\rho_A)+S(\rho_{BC})-S(\rho_{ABC})$ , cancel $S(\rho_A)$ , and rearrange to $S(\rho_{ABC})+S(\rho_B)\le S(\rho_{AB})+S(\rho_{BC})$ . Why: the $S(\rho_A)$ terms are spectators; once they drop out, monotonicity of mutual information is verbatim SSA.

Step 4. Run the implication backward. Given any $\mathcal{N}$ , Stinespring-dilate it on a purifying register and apply SSA to the resulting tripartite state; it reproduces monotonicity under $\mathcal{N}$ . Why: this closes the loop and establishes the two-way equivalence — SSA and DPI carry exactly the same content. *Note: the conditional mutual information identity $I(A;C|B)=S(AB)+S(BC)-S(ABC)-S(B)\ge0$ underlies the bookkeeping in Step 3; equality holds if and only if $\rho_{ABC}$ is a quantum Markov chain $A-B-C$ (Hayden–Jozsa–Petz–Winter).*

Go deeper · Derivation C · 1 minThe Holevo bound as a corollary of data processing

Derivation CThe Holevo bound as a corollary of data processing

The Holevo bound — that the classical information any measurement can extract about which member of an ensemble was prepared cannot exceed the Holevo $\chi$ quantity — is DPI applied to the measurement channel. The goal is to show a "physical" communication theorem descending from a single monotonicity in four steps: package the ensemble as a classical–quantum state whose mutual information is $\chi$ , recognize the receiver's measurement as a CPTP map, and let DPI cap the accessible information by $\chi$ . This is why $\chi$ sets the capacity of a classical-quantum channel. Take an ensemble $\{p_x,\rho_x\}$ with average $\bar\rho=\sum_x p_x\rho_x$ .

Step 1. Encode the ensemble in a classical–quantum state $\rho_{XB}=\sum_x p_x\,|x\rangle\langle x|_X\otimes(\rho_x)_B$ , with an orthonormal register $X$ holding the label and $B$ the quantum system. Why: promoting the classical index to a quantum register lets one mutual information hold both "which state was sent" and the states themselves.

Step 2. Compute $I(X;B)_\rho=S(\bar\rho)-\sum_x p_x S(\rho_x)=\chi$ . Why: orthogonality of the $|x\rangle$ splits the joint entropy cleanly, and what survives is exactly the Holevo quantity — the pre-measurement correlation.

Step 3. Model the receiver as a measurement channel $\mathcal{M}:(\rho_x)_B\mapsto\sum_y\mathrm{Tr}[E_y\rho_x]\,|y\rangle\langle y|_Y$ built from a POVM $\{E_y\}$ . This map is CPTP. Why: every physically realizable readout is some POVM, so bounding all POVMs at once is what makes the Holevo bound universal.

Step 4. Apply DPI for mutual information under $\mathrm{id}_X\otimes\mathcal{M}$ : $I(X;Y)\le I(X;B)=\chi$ . Why: measuring is processing, and processing cannot create correlation, so the classical information read out is capped by the quantum $\chi$ you started with. *Note: $I(X;B)=\chi$ for the classical–quantum state constructed in Steps 1–2, and $\mathcal{M}$ is indeed CPTP; the bound is saturable only when the signal states $\rho_x$ commute.*

Limiting cases

DPI specializes to a whole family of monotone quantities depending on the choice of channel, states, or divergence.

Case	Result	Meaning
Partial trace	$D(\rho_{AB}\,\\|\,\sigma_{AB})\ge D(\rho_A\,\\|\,\sigma_A)$	Choosing $\mathcal{N}=\mathrm{Tr}_B$ recovers monotonicity under discarding a subsystem — the building block from which the general case follows by Stinespring.
Mutual information	$I(A;B)_\rho\ge I(A;B')_{(\mathrm{id}\otimes\mathcal{N})\rho}$	Local processing of $B$ alone cannot increase correlations with $A$ ; write $I=D(\rho_{AB}\\|\rho_A\otimes\rho_B)$ and apply DPI to $\mathrm{id}\otimes\mathcal{N}$ .
Fidelity (reversed)	$F(\mathcal{N}(\rho),\mathcal{N}(\sigma))\ge F(\rho,\sigma)$	Fidelity measures closeness, so it can only rise: channels bring states together. Here $F(\rho,\sigma)=\\|\sqrt\rho\sqrt\sigma\\|_1$ .
Trace distance	$\tfrac12\\|\mathcal{N}(\rho)-\mathcal{N}(\sigma)\\|_1\le\tfrac12\\|\rho-\sigma\\|_1$	Single-shot distinguishability (Helstrom) cannot improve; this is the contractivity of the trace norm under CPTP maps.
Rényi range	Petz $\alpha\in[0,2]$ ; sandwiched $\alpha\ge\tfrac12$	DPI for the Petz–Rényi divergence holds only for $\alpha\in[0,2]$ , and for the sandwiched version only for $\alpha\ge 1/2$ ; outside these ranges monotonicity fails.

Common trapAssuming DPI holds for any map, in any direction, for any divergence

DPI applies to genuine CPTP maps acting identically on both arguments — not to arbitrary linear maps, not to post-selection, and not to maps that differ between $\rho$ and $\sigma$ . Post-selection renormalizes after discarding events, which can increase distinguishability (you keep only the favorable cases) and openly violates the inequality; trace preservation is the physical constraint that forbids gaining information for free. Mind the direction, too: relative entropy and trace distance DECREASE under channels, while fidelity INCREASES — the same physical fact read as distinguishability versus closeness. And do not assume monotonicity for every Rényi parameter: the Petz $D_\alpha$ is monotone only for $\alpha\in[0,2]$ , the sandwiched $\widetilde D_\alpha$ only for $\alpha\ge 1/2$ .

Worked examples

1 · Distinguishability, trace distance and fidelity under a bit-flip (fully worked)

A bit-flip channel $\mathcal{N}(\cdot)=(1-q)(\cdot)+q\,X(\cdot)X$ with $q=0.25$ acts on the pure state $\rho=|0\rangle\langle 0|$ and the reference $\sigma=I/2$ . Track the relative entropy, the trace distance and the fidelity through the channel and check that all three move in the DPI directions.

Step 1. The bit flip is unital, so it fixes the maximally mixed state, $\mathcal{N}(I/2)=I/2$ , and it sends $\rho=\mathrm{diag}(1,0)$ to $\mathcal{N}(\rho)=(1-q)\,\mathrm{diag}(1,0)+q\,\mathrm{diag}(0,1)=\mathrm{diag}(0.75,0.25)$ . Both input and output pairs are diagonal, so every divergence reduces to a classical sum.

Step 2. Relative entropy. Input: $D(\mathrm{diag}(1,0)\,\|\,I/2)=1\cdot\log_2\tfrac{1}{0.5}=\mathbf{1\ \text{bit}}$ (a pure state is maximally distinguishable from the uniform reference). Output: $D(\mathrm{diag}(0.75,0.25)\,\|\,I/2)=\log_2 2-S(\mathrm{diag}(0.75,0.25))=1-0.8113=\mathbf{0.1887\ \text{bits}}$ . So $D$ drops from $1$ to $0.189$ bits — DPI holds, and $0.1887$ is the very number computed for this state in the relative-entropy article.

Step 3. Trace distance. Input: $\tfrac12\|\mathrm{diag}(0.5,-0.5)\|_1=\mathbf{0.5}$ . Output: $\tfrac12\|\mathrm{diag}(0.25,-0.25)\|_1=\mathbf{0.25}$ . It contracts by exactly $1-2q=0.5$ , the known bit-flip factor — again a decrease, consistent with (5).

Step 4. Fidelity, which must go the other way. Input: $F(|0\rangle\langle 0|,I/2)=\sqrt{\langle 0|(I/2)|0\rangle}=\sqrt{0.5}=\mathbf{0.707}$ . Output: $F(\mathrm{diag}(0.75,0.25),I/2)=\tfrac{1}{\sqrt2}(\sqrt{0.75}+\sqrt{0.25})=\mathbf{0.966}$ . Fidelity rises as distinguishability falls. All three agree: $D:1\to0.189$ , trace distance $0.5\to0.25$ , fidelity $0.707\to0.966$ — one physical fact, three consistent monotones.

2 · The Holevo bound in numbers: accessible information is capped by χ

Alice sends one of two equiprobable non-orthogonal states, $|0\rangle$ or $|+\rangle$ . Compute the Holevo $\chi$ and compare it to the classical mutual information Bob obtains from a computational-basis measurement, verifying $I(X;Y)\le\chi$ .

Step 1. The ensemble average is $\bar\rho=\tfrac12|0\rangle\langle 0|+\tfrac12|+\rangle\langle +|=\left(\begin{smallmatrix}0.75&0.25\\0.25&0.25\end{smallmatrix}\right)$ , with eigenvalues $\lambda_\pm=\tfrac12(1\pm\tfrac{1}{\sqrt2})=0.8536,\ 0.1464$ .

Step 2. Since the signal states are pure, $S(\rho_x)=0$ , so $\chi=S(\bar\rho)=-0.8536\log_2 0.8536-0.1464\log_2 0.1464=\mathbf{0.601\ \text{bits}}$ . No measurement whatsoever can extract more than $0.601$ bits about which state was sent — that is the Holevo bound of Derivation 3.

Step 3. Bob measures in the $Z$ basis. If $|0\rangle$ was sent he always reads $y=0$ ; if $|+\rangle$ was sent he reads $y=0$ or $y=1$ with equal probability. The joint distribution is $P(0,0)=0.5,\ P(1,0)=0.25,\ P(1,1)=0.25$ , with marginals $P(y{=}0)=0.75,\ P(y{=}1)=0.25$ .

Step 4. The accessible information is $I(X;Y)=\sum_{x,y}P(x,y)\log_2\tfrac{P(x,y)}{P(x)P(y)}=0.2075-0.1462+0.25=\mathbf{0.311\ \text{bits}}$ . Since $0.311\le 0.601$ , DPI is satisfied: measuring is a channel, and it cannot lift the classical mutual information above the quantum $\chi$ . The gap between $0.311$ and $0.601$ is the information the $Z$ measurement leaves on the table — a cleverer POVM narrows it but can never cross $\chi$ .

Physics of the portrait

The data-processing inequality is where a statistician's intuition became an operator-algebra theorem: the classical version was folklore, but the quantum monotonicity of relative entropy waited for someone fluent in unbounded operators to prove it.

Origins · 1 minGöran Lindblad and how this was discovered

Göran Lindblad

b. 1940 · Swedish theoretical physicist · open quantum systems and quantum information

Göran Lindblad, working at the Royal Institute of Technology in Stockholm, is best known for the 1976 Lindblad master equation that governs Markovian open-system dynamics — but a year earlier, in 1975, he proved the result on which this whole article turns: that the quantum relative entropy is monotone under every completely positive trace-preserving map. His starting point was the 1973 breakthrough of Elliott Lieb and Mary Beth Ruskai, who had established strong subadditivity and the joint convexity of relative entropy; Lindblad recognized that joint convexity, combined with the invariance of $D$ under the reversible parts of a channel, forced the full data-processing inequality for finite-dimensional systems. Armin Uhlmann extended the monotonicity to the general von Neumann algebra setting in 1977, so the theorem is often called Lindblad–Uhlmann monotonicity. What was proved as a structural fact about noncommutative entropy turned out to be the load-bearing beam of quantum Shannon theory: converse bounds on capacity, the Holevo bound, the second law of quantum thermodynamics and the theory of approximate error correction all rest on it. Decades later Dénes Petz characterized exactly when the inequality is tight — via the recovery channel that now bears his name — and Omar Fawzi and Renato Renner turned that equality condition into a robust, quantitative bound. The arc is characteristic of the field: a lemma about operator algebras becomes, in hindsight, the principle itself.

History

1951

— Kullback and Leibler define the classical relative entropy, the KL divergence

1973

— Lieb and Ruskai prove strong subadditivity and joint convexity of relative entropy

1975

— Lindblad proves monotonicity of relative entropy under CPTP maps — the data-processing inequality

1986

— Petz characterizes the equality case and the recovery channel

2015

— Fawzi and Renner lower-bound the data-processing gap by recovery fidelity

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §11.3–11.4 & §12.1 · Wilde, Quantum Information Theory, 2nd ed., Ch. 11–13 · Watrous, The Theory of Quantum Information, Ch. 5 · Preskill, Lecture Notes for Physics 219, Ch. 10 · Lindblad, Comm. Math. Phys. 40, 147 (1975) · Uhlmann, Comm. Math. Phys. 54, 21 (1977) · Petz, Rev. Math. Phys. 15, 79 (2003) · Fawzi & Renner, Comm. Math. Phys. 340, 575 (2015). The SSA equivalence rests on the identity $I(A;C|B)=S(AB)+S(BC)-S(ABC)-S(B)\ge0$ . Key numerical benchmarks: bit-flip $q=0.25$ on $|0\rangle\langle0|$ vs $I/2$ gives $D:1\to0.1887$ bits, trace distance $0.5\to0.25$ , fidelity $0.707\to0.966$ ; the Holevo ensemble $\{|0\rangle,|+\rangle\}$ equiprobable gives $\chi=0.601$ bits with $Z$ -measurement $I(X;Y)=0.311$ bits. Rényi monotonicity holds for the Petz $D_\alpha$ with $\alpha\in[0,2]$ and for the sandwiched $\widetilde D_\alpha$ with $\alpha\ge1/2$ (Müller-Lennert et al. 2013; Wilde–Winter–Yang 2014).

Prerequisites. Quantum Relative Entropy · Quantum Channels and CPTP Maps · Von Neumann Entropy · Strong Subadditivity of Quantum Entropy. Leads to. Holevo Bound · Complementary Channels and Degradability · Stinespring Dilation Theorem · Schumacher Quantum Data Compression.

Ask the tutor: Ask me to derive strong subadditivity from data processing in one line, or to check the DPI directions for any channel and pair of qubit states you name.

Is the data-processing inequality really just "you can't gain information by processing"?

That is the honest one-line summary, but the precise content is sharper. DPI does not say your processed data are less useful for every purpose — it says the two hypotheses $\rho$ and $\sigma$ become no easier to tell apart, measured by relative entropy. You might process raw data into a form that is far more convenient to work with; what you cannot do is make the underlying states more distinguishable than they were at the input. The inequality is about the best-case statistical separation of the pair, not about the readability of any particular representation. This is why it powers converse (impossibility) proofs: it caps what any downstream decoder could possibly achieve.

Why does fidelity increase under a channel while relative entropy decreases — isn't that a contradiction?

It is the same statement read from opposite ends. Relative entropy and trace distance are measures of distinguishability: they are large when the states are easy to tell apart and shrink as a channel blurs them together. Fidelity is a measure of closeness: it equals $1$ when the states coincide and grows as they are pushed together. So a channel that lowers $D$ and lowers the trace distance must, by the very same physical action, raise the fidelity. Worked example 1 shows all three in one calculation: $D$ and trace distance fall while fidelity rises. There is no contradiction — only two vocabularies for "the states got closer."

How can strong subadditivity, a statement purely about entropies, be equivalent to DPI, a statement about channels?

Because a partial trace is a channel, and mutual information is a relative entropy. Write $I(A;BC)=D(\rho_{ABC}\|\rho_A\otimes\rho_{BC})$ ; apply DPI with the channel "discard $C$ "; the result is $I(A;B)\le I(A;BC)$ , which expands directly into SSA once the shared $S(\rho_A)$ terms cancel. Conversely, dilating an arbitrary channel with a purifying register and applying SSA reproduces monotonicity under that channel. The two are genuinely interchangeable — historically SSA (Lieb–Ruskai 1973) came first and Lindblad used its companion, joint convexity, to derive general monotonicity two years later.

When does DPI hold with equality, and why does anyone care?

Equality $D(\mathcal{N}(\rho)\|\mathcal{N}(\sigma))=D(\rho\|\sigma)$ holds if and only if a single recovery channel — the Petz map $\mathcal{R}_{\sigma,\mathcal{N}}$ — perfectly undoes $\mathcal{N}$ on both states simultaneously: $\mathcal{R}\circ\mathcal{N}(\rho)=\rho$ and $\mathcal{R}\circ\mathcal{N}(\sigma)=\sigma$ . Physically, no distinguishability was lost, so the processing was reversible on this pair and the states form a sufficient statistic for the channel. People care because approximate reversibility is the whole game in error correction: the Fawzi–Renner bound (6) says a small DPI gap guarantees a near-perfect recovery, so you can certify that lost quantum information is almost recoverable just by checking how little the relative entropy dropped.

Why can't I apply DPI to a post-selected or measurement-conditioned map?

Because such a map is not trace-preserving. Post-selection keeps only the runs with a favorable outcome and renormalizes, which is exactly conditioning on a measurement result — and that can increase distinguishability, since you are deliberately discarding the ambiguous cases. Trace preservation is the mathematical encoding of "no information for free": it forbids the selective throwing-away of events that would let you sharpen the two hypotheses. If you want to describe a measurement honestly, include all outcomes as a genuine CPTP channel into a classical register (as in the Holevo derivation); then DPI applies and correctly caps the accessible information.

Does DPI hold for the Rényi and sandwiched relative entropies too?

Only within specific parameter ranges, so this is a common trap. The Petz–Rényi relative entropy $D_\alpha$ is monotone under CPTP maps for $\alpha\in[0,2]$ , while the sandwiched (minimal) version $\widetilde D_\alpha$ is monotone for $\alpha\ge 1/2$ ; both reduce to the Umegaki relative entropy of this article at $\alpha=1$ , which is the case Stein's lemma singles out. Outside those windows monotonicity genuinely fails, so you cannot assume a data-processing inequality for an arbitrary divergence just because it looks entropy-like. The sandwiched family governs strong-converse exponents and the Petz family governs error tradeoffs, which is why refined, finite-block-length analyses need the ranges kept straight.