Holevo Bound

Alice draws a classical symbol $X=x$ with probability $p_x$ , encodes it into a quantum state $\rho_x$ , and hands the system to Bob. Bob measures — a POVM $\{E_y\}$ — reads an outcome $Y=y$ , and tries to recover $x$ . How much of Alice's message can survive this quantum bottleneck? The figure of merit is the classical mutual information $I(X;Y)$ , and the best any measurement can do is the accessible information $I_{\mathrm{acc}}=\max_{\{E_y\}}I(X;Y)$ . The Holevo bound answers with a single, decisive inequality: whatever the measurement, $I(X;Y)\le\chi$ , where $\chi=S(\rho)-\sum_x p_x S(\rho_x)$ is a quantity built from the ensemble alone and blind to how Bob measures. The number $\chi$ — the Holevo information — is not an ad hoc bound but a mutual information in disguise. Package the ensemble as a classical–quantum state $\rho_{XQ}=\sum_x p_x\,|x\rangle\langle x|_X\otimes\rho_x$ ; then $\chi=I(X;Q)$ , the quantum mutual information between the label and the register. Bob's measurement is a quantum channel acting on $Q$ alone, and the data-processing inequality forbids a channel from raising mutual information. The Holevo bound is that one line: $I(X;Y)\le I(X;Q)=\chi$ . The consequence is stark. A $d$ -dimensional quantum system stores at most $\log_2 d$ classical bits, exactly as a classical $d$ -level system does, because $\chi\le S(\rho)\le\log d$ . Non-orthogonal alphabets, which seem to promise a denser packing of letters into Hilbert space, are punished twice: overlap lowers $\chi$ below $\log d$ , and measurement disturbance lowers $I_{\mathrm{acc}}$ below $\chi$ . Holevo's 1973 inequality is the impossibility half of quantum Shannon theory.

Key takeaways

For an ensemble $\{p_x,\rho_x\}$ the classical information any measurement extracts is capped by the Holevo quantity $\chi=S(\rho)-\sum_x p_x S(\rho_x)$ , with $\rho=\sum_x p_x\rho_x$ the average state.
$\chi$ is exactly the quantum mutual information $I(X;Q)$ of the classical–quantum encoding, so the bound $I(X;Y)\le\chi$ is a one-line corollary of the data-processing inequality.
Because $\chi\le S(\rho)\le\log d$ , no more than $\log_2 d$ classical bits can be read from one $d$ -dimensional system: a qubit carries at most one bit.
The bound is generally loose on a single copy — for non-commuting $\rho_x$ the accessible information is strictly below $\chi$ ; equality holds iff the states all commute.
$\chi$ is the converse half of the Holevo–Schumacher–Westmoreland theorem: maximized over inputs it equals the product-state classical capacity of a quantum channel, approached by collective measurements over many copies.
Key idea: A measurement is a CPTP map; data processing for quantum mutual information turns the bound $I(X;Y)\le I(X;Q)=\chi$ into an immediate corollary, and $\chi$ is the quantum mutual information of the classical–quantum encoding.

The core insight

A measurement is a CPTP map; data processing for quantum mutual information then turns the bound $I(X;Y)\le I(X;Q)=\chi$ into an immediate corollary. The Holevo $\chi$ is the quantum mutual information of the classical–quantum encoding.

Where this lives today

The Holevo bound is no longer a curiosity but a working tool, and its deepest modern question is whether $\chi$ is the whole story. The additivity conjecture long held that the single-letter Holevo capacity $C(\mathcal{N})$ in equation (7) equaled the true capacity — that entangling inputs across channel uses could not help. Shor showed in 2004 that additivity of Holevo capacity, of minimum output entropy, and of entanglement of formation were all equivalent; then in 2009 Hastings constructed a random channel violating additivity, proving that entangled inputs can raise the classical capacity and that the true capacity needs the regularized limit $\lim_n \tfrac1n C(\mathcal{N}^{\otimes n})$ . The single-letter $\chi$ remains the achievable product-state rate every experiment reports, but its ultimate optimality is now known to fail in general. On the applied side, $\chi$ is the security ledger of quantum key distribution: Eve's information about a sifted key is bounded by the Holevo quantity of her side-information ensemble, and the Devetak–Winter secret-key rate $r=I(X;B)-\chi(X;E)$ subtracts exactly that bound from Alice and Bob's mutual information, yielding provable rates for BB84 and continuous-variable protocols now running over fibre and satellite links. The bound also underwrites quantum random-access codes: because a qubit yields at most one accessible bit, an $m$ -into-one random-access code with perfect recovery is impossible for $m\ge2$ , and laboratory $2\to1$ and $3\to1$ codes report retrieval probabilities in precise agreement with the Holevo ceiling.

The physical picture

Picture Alice's label $X$ tethered to a quantum register $Q$ by a fixed amount of correlation, the quantum mutual information $I(X;Q)=\chi$ . Bob cannot touch $X$ ; he can only process $Q$ , and processing a system blurs correlations, never sharpens them. So whatever pointer $Y$ his apparatus produces, the correlation $I(X;Y)$ it retains cannot exceed the $\chi$ that was there to begin with. Holevo's inequality is conservation of information under processing, made precise. The first figure draws this cascade and stamps the relation onto it: correlation flows $X$ – $Q$ – $Y$ and can only shrink.

Figure

A measurement is a quantum channel — the classical label X is correlated with the quantum register Q with quantum mutual information I(X;Q) = χ, and Bob's POVM sends Q to a classical outcome Y. Data processing forbids the channel from raising the correlation, so I(X;Y) ≤ I(X;Q) = χ

The penalty for non-orthogonality is geometric. When the $\rho_x$ have orthogonal supports they are perfectly distinguishable, $\chi$ climbs to $H(X)$ , and a single projective measurement recovers everything. As the states crowd together their overlap makes them ambiguous to any apparatus: $\chi$ falls, and the accessible information falls further still, because measuring a non-commuting alphabet disturbs it. The second figure quantifies this for the cleanest case — two equiprobable pure qubits — showing the bound $\chi$ and the achievable $I_{\mathrm{acc}}$ pull apart as overlap grows.

Figure

The accessible-information gap for two equiprobable non-orthogonal pure qubits — the upper blue curve is the Holevo bound χ and the lower red curve is the accessible information I_acc, both versus the state overlap s. At overlap s = 0.707 the states carry χ = 0.601 bit yet any single measurement extracts at most I_acc = 0.399 bit, so the bound is loose by 0.202 bit

A one-line slogan carries the moral: $\log d$ dimensions, $\log d$ bits. A single qubit, however cleverly loaded, yields at most one classical bit per use. Genuine quantum advantage in communication comes from entanglement — superdense coding — or from coding over many channel uses, never from over-packing an unentangled single system.

The core equations

The Holevo information is the gap between the entropy of the average state and the average of the entropies. It measures how much the ensemble's mixing hides — how much order is lost when the labels are forgotten and only $\rho$ remains.

\chi(\{p_x,\rho_x\})=S\!\Big(\sum_x p_x\rho_x\Big)-\sum_x p_x\,S(\rho_x) \qquad (1)

where: $\chi$ — Holevo quantity (bits when $S$ uses $\log_2$ ); $p_x$ — prior of letter $x$ ; $\rho_x$ — density operator for $x$ ; $\rho=\sum_x p_x\rho_x$ — ensemble-average state; $S(\cdot)$ — von Neumann entropy $-\mathrm{Tr}(\rho\log\rho)$ . Reading: $\chi$ is large when the letters spread $\rho$ across many dimensions while each letter stays sharp. The bound itself says that this ensemble quantity caps every measurement outcome. The right-hand side has no reference to the POVM at all — it is fixed before Bob chooses his apparatus.

I(X;Y)\ \le\ \chi(\{p_x,\rho_x\}) \qquad (2)

where: $I(X;Y)$ — classical mutual information between label $X$ and outcome $Y$ ; the measurement $\{E_y\}$ , obeying $E_y\ge0$ and $\sum_y E_y=\mathbb{1}$ , enters only through $p(y|x)=\mathrm{Tr}(E_y\rho_x)$ . Reading: no choice of $E_y$ beats $\chi$ — the ceiling is fixed by what was sent, not how it is read. The reason the bound is true, rather than merely plausible, is the identity that $\chi$ is a mutual information. Encode the ensemble as a classical–quantum state and its quantum mutual information equals $\chi$ exactly.

\chi=I(X;Q)_{\rho_{XQ}},\qquad \rho_{XQ}=\sum_x p_x\,|x\rangle\langle x|_X\otimes\rho_x \qquad (3)

where: $I(X;Q)=S(\rho_X)+S(\rho_Q)-S(\rho_{XQ})$ — quantum mutual information of the cq-state; $|x\rangle$ — orthonormal classical flags on register $X$ . Reading: $\chi$ inherits every property of a mutual information — non-negativity and monotonicity under channels — which makes the bound fall out for free. An equivalent form writes $\chi$ as the average quantum relative entropy of each letter from the mean, exposing it as an average distinguishability.

\chi=\sum_x p_x\,D\big(\rho_x\,\|\,\rho\big),\qquad D(\sigma\|\rho)=\mathrm{Tr}\,\sigma(\log\sigma-\log\rho) \qquad (4)

where: $D(\sigma\|\rho)\ge0$ — quantum relative entropy, zero iff $\sigma=\rho$ . Reading: $\chi$ is the average distance of the letters from the mean; non-negativity of $D$ gives $\chi\ge0$ , with equality iff every $\rho_x=\rho$ . Two chains of inequalities bracket what $\chi$ can be and pin where the bound is tight. The dimensional ceiling is the headline; the commutation condition is the fine print.

0\ \le\ I_{\mathrm{acc}}\ \le\ \chi\ \le\ S(\rho)\ \le\ \log d \qquad (5)

where: $I_{\mathrm{acc}}=\max_{\{E_y\}}I(X;Y)$ — best single-copy accessible information; $d$ — Hilbert-space dimension of the letter. Reading: the staircase caps classical storage at $\log_2 d$ bits per system, and each step can be strict; the gap $\chi-I_{\mathrm{acc}}$ is the price of non-orthogonality.

I_{\mathrm{acc}}=\chi\ \iff\ [\rho_x,\rho_{x'}]=0\ \ \forall\,x,x' \qquad (6)

where: $[\cdot,\cdot]$ — commutator; mutual commutation means the letters share one eigenbasis, i.e. a classical ensemble in disguise. Reading: the single-copy bound is achievable exactly when quantumness is absent; otherwise $\chi$ is only an asymptotic rate. That asymptotic meaning is the Holevo–Schumacher–Westmoreland capacity: maximize $\chi$ over input ensembles of channel outputs and you get the achievable classical rate.

C(\mathcal{N})=\max_{\{p_x,\rho_x\}}\chi\big(\{p_x,\mathcal{N}(\rho_x)\}\big) \qquad (7)

where: $\mathcal{N}$ — a quantum channel (CPTP map); $C(\mathcal{N})$ — its product-state (one-shot) classical capacity. Reading: inequality (2) is the converse and random coding with collective decoding is the achievability, making $C$ the quantum analogue of Shannon's capacity.

Derivation AChi equals the quantum mutual information of the classical quantum state

The engine of the entire subject is a single identity: the Holevo quantity of an ensemble is nothing other than the quantum mutual information of its classical–quantum encoding, $\chi=I(X;Q)$ . Establishing this converts an inequality about measurements into a statement about a fixed bipartite state, after which the data-processing inequality does all the remaining work. We compute the three entropies of $\rho_{XQ}=\sum_x p_x|x\rangle\langle x|_X\otimes\rho_x$ in turn and assemble the mutual information, watching the classical $H(X)$ terms cancel.

Step 1. Trace out $Q$ to get the marginal on the label register: $\rho_X=\mathrm{Tr}_Q\,\rho_{XQ}=\sum_x p_x|x\rangle\langle x|$ , a diagonal state whose von Neumann entropy is the Shannon entropy $S(\rho_X)=H(X)=-\sum_x p_x\log p_x$ . Why: the $|x\rangle$ are orthonormal, so the register is effectively classical and its quantum entropy collapses to the classical entropy of the priors.

Step 2. Trace out $X$ to get the marginal on the quantum register: $\rho_Q=\mathrm{Tr}_X\,\rho_{XQ}=\sum_x p_x\rho_x=\rho$ , so $S(\rho_Q)=S(\rho)$ . Why: forgetting the label mixes the letters into their weighted average, which is exactly the ensemble-average state $\rho$ that appears in the definition of $\chi$ .

Step 3. Compute the joint entropy. Because $\rho_{XQ}$ is block-diagonal across the orthogonal flags $|x\rangle$ , its spectrum is the union of the spectra of the blocks $p_x\rho_x$ , and the entropy splits as $S(\rho_{XQ})=H(X)+\sum_x p_x S(\rho_x)$ . Why: a block-diagonal state with known block weights has entropy equal to the classical entropy of the weights plus the weighted entropies within each block — the quantum version of the grouping axiom for entropy.

Step 4. Assemble the mutual information: $I(X;Q)=S(\rho_X)+S(\rho_Q)-S(\rho_{XQ})=H(X)+S(\rho)-\big[H(X)+\sum_x p_x S(\rho_x)\big]$ . The two $H(X)$ terms cancel, leaving $I(X;Q)=S(\rho)-\sum_x p_x S(\rho_x)=\chi$ . Why: the label entropy enters both the marginal and the joint term identically, so it drops out, and what remains is precisely the Holevo quantity.

Step 5. Read off the payoff: $\chi\ge0$ automatically, since a quantum mutual information is non-negative (subadditivity $S(\rho_{XQ})\le S(\rho_X)+S(\rho_Q)$ ), reproducing the concavity bound $S(\rho)\ge\sum_x p_x S(\rho_x)$ . Why: the identity does not merely rename $\chi$ — it hands us non-negativity and, next, monotonicity for free, which are the two facts the bound needs.

Go deeper · Derivation B · 1 minThe Holevo bound as data processing

Derivation BThe Holevo bound as data processing

With the identity $\chi=I(X;Q)$ in hand, the bound $I(X;Y)\le\chi$ is a one-step consequence of the data-processing inequality — the monotonicity of quantum mutual information under a local channel, itself equivalent to strong subadditivity of the von Neumann entropy. The goal here is to make that step explicit: model Bob's measurement as a CPTP map on $Q$ , apply it to the classical–quantum state, and show the surviving correlation is $I(X;Y)$ , which cannot exceed $\chi$ .

Step 1. Represent the measurement as a channel. A POVM $\{E_y\}$ defines the CPTP map $\mathcal{M}(\sigma)=\sum_y \mathrm{Tr}(E_y\sigma)\,|y\rangle\langle y|_Y$ , sending any input state to a classical pointer with distribution $p(y|x)=\mathrm{Tr}(E_y\rho_x)$ when fed $\rho_x$ . Why: measurement outcomes are the paradigm of classical data extracted from a quantum system, and Kraus operators $\sqrt{E_y}$ show this map is completely positive and trace-preserving — a legitimate channel.

Step 2. Apply the channel to the $Q$ half of $\rho_{XQ}$ , leaving $X$ untouched: $\rho_{XY}=(\mathrm{id}_X\otimes\mathcal{M})\,\rho_{XQ}=\sum_{x,y}p_x\,p(y|x)\,|x\rangle\langle x|\otimes|y\rangle\langle y|$ . Both registers are now classical, so the quantum mutual information of $\rho_{XY}$ equals the Shannon mutual information $I(X;Y)$ of the joint distribution $p_x\,p(y|x)$ . Why: a classical–classical state's quantum mutual information reduces to the ordinary Shannon quantity, so the abstract inequality below becomes a statement about the measurement's information yield.

Step 3. Invoke data processing. Acting with a channel on one subsystem cannot increase the quantum mutual information: $I(X;Y)_{\rho_{XY}}\le I(X;Q)_{\rho_{XQ}}$ . Why: this monotonicity is the operational content of strong subadditivity (Lieb–Ruskai); processing $Q$ can destroy correlation with $X$ but never manufacture it, because a channel is a partial trace over an isometric dilation.

Step 4. Substitute the identity from the previous derivation, $I(X;Q)_{\rho_{XQ}}=\chi$ , to conclude $I(X;Y)\le\chi$ for every POVM. Maximizing the left side over measurements gives $I_{\mathrm{acc}}\le\chi$ . Why: the inequality holds pointwise in the measurement, so it survives the maximum that defines accessible information — the bound is universal over apparatus.

Go deeper · Derivation C · 1 minWhen the bound is tight

Derivation CWhen the bound is tight

The single-copy bound is not always achievable, and characterizing when it is recovers exactly the classical world inside the quantum one. The claim: equality $I_{\mathrm{acc}}=\chi$ holds if and only if the states $\rho_x$ all mutually commute — share a common eigenbasis. The forward direction is a construction — a projective measurement in the shared eigenbasis makes the ensemble a classical channel attaining $\chi$ ; the converse traces the strict inequality to the impossibility of simultaneously diagonalizing non-commuting operators.

Step 1. Suppose all $\rho_x$ commute, so they share an eigenbasis $\{|e_k\rangle\}$ with $\rho_x=\sum_k \lambda^{(x)}_k|e_k\rangle\langle e_k|$ . Measure in that basis, $E_k=|e_k\rangle\langle e_k|$ , giving outcome statistics $p(k|x)=\lambda^{(x)}_k$ . Why: when a common eigenbasis exists, one measurement reads every letter without disturbance, so no quantum ambiguity remains.

Step 2. The resulting classical channel has mutual information equal to $\chi$ . Here $S(\rho_x)=H(\lambda^{(x)})$ and $S(\rho)=H(\sum_x p_x\lambda^{(x)})$ , so $\chi=S(\rho)-\sum_x p_x S(\rho_x)$ is precisely the classical mutual information $I(X;K)$ of the joint distribution $p_x\lambda^{(x)}_k$ . Why: for a commuting ensemble the Holevo quantity is a classical mutual information in disguise, and the eigenbasis measurement realizes it exactly, so the bound saturates.

Step 3. Conversely, if some pair fails to commute, $[\rho_x,\rho_{x'}]\ne0$ , no single basis diagonalizes all letters, and equality in the data-processing step fails: the measurement is not a sufficient statistic for $X$ . Hence $I_{\mathrm{acc}}<\chi$ strictly on one copy. Why: saturation of data processing demands that the channel be reversible on the relevant states (a recoverability condition), which non-commuting letters cannot satisfy — measuring one basis blurs the others.

Step 4. The gap is genuine and closes only asymptotically. For four BB84-style qubit states $\chi$ can reach one bit while $I_{\mathrm{acc}}$ stays well below it; only collective measurements over many copies drive the per-copy accessible information up to $\chi$ , which is the achievability half of HSW. Why: $\chi$ is an asymptotic rate, not a single-shot yield — the strict single-copy gap is the whole reason quantum Shannon theory needs block coding.

Limiting cases

Case	Result	Meaning
Orthogonal pure states	$S(\rho_x)=0,\ S(\rho)=H(X)$ so $\chi=H(X)$	Perfectly distinguishable letters; the bound saturates and $I_{\mathrm{acc}}=\chi=H(X)$ via one projective measurement — fully classical.
Identical states	$\rho_x=\sigma\ \forall x$ so $\chi=0$	The average equals every letter, $S(\rho)=\sum p_x S(\rho_x)$ ; nothing about $X$ can be learned, and $I_{\mathrm{acc}}=0$ .
Two non-orthogonal pure qubits	$\chi=H_2\!\big(\tfrac{1+s}{2}\big)<1$ , $s=\vert\langle\psi_0\vert\psi_1\rangle\vert$	Overlap $s>0$ pushes $\chi$ below one bit, and $I_{\mathrm{acc}}=1-H_2\!\big(\tfrac{1+\sqrt{1-s^2}}{2}\big)$ is strictly smaller still.
Maximal ensemble	$\chi\le S(\rho)\le\log d$	For any $d$ -dimensional letter $\chi$ never exceeds $\log_2 d$ bits — the hard ceiling on classical storage in a quantum system.

Common trapConfusing the Holevo bound with the accessible information

The Holevo $\chi$ is an upper bound on the accessible information, not the accessible information itself. For non-commuting $\rho_x$ the strict inequality $I_{\mathrm{acc}}<\chi$ always holds on a single copy — no measurement realizes $\chi$ letter by letter, and $\chi$ is only approached asymptotically with collective measurements (this is the HSW achievability, not the Holevo bound). Two further traps: $\chi$ depends on the ensemble decomposition $\{p_x,\rho_x\}$ , not on the average state $\rho$ alone; and $\chi\le\log_2 d$ does not say a qubit "contains" only one bit — it says only one classical bit can be reliably read out, even though the state has continuously many parameters you cannot retrieve as classical data.

Worked examples

1 · Two non-orthogonal pure states — the bound and the gap (fully worked)

Alice sends $|\psi_0\rangle=|0\rangle$ or $|\psi_1\rangle=|+\rangle=\tfrac{1}{\sqrt2}(|0\rangle+|1\rangle)$ , each with probability $\tfrac12$ . Find $\chi$ , and compare it with the accessible information $I_{\mathrm{acc}}$ .

Step 1. The overlap is $s=|\langle\psi_0|\psi_1\rangle|=|\langle 0|+\rangle|=\tfrac{1}{\sqrt2}=0.7071$ . The average state is $\rho=\tfrac12|0\rangle\langle0|+\tfrac12|+\rangle\langle+\rangle=\begin{pmatrix}0.75 & 0.25\\ 0.25 & 0.25\end{pmatrix}$ , whose eigenvalues are $\tfrac{1\pm s}{2}=0.8536$ and $0.1464$ .

Step 2. Each letter is pure, so $S(\rho_x)=0$ and $\chi=S(\rho)=H_2(0.8536)$ . Evaluating the binary entropy, $\chi=-0.8536\log_2 0.8536-0.1464\log_2 0.1464=\mathbf{0.601\ bit}$ . This is the ceiling on any measurement's yield.

Step 3. The optimal single-copy measurement (a projective measurement in the basis symmetric to the two states) gives $I_{\mathrm{acc}}=1-H_2\!\big(\tfrac{1+\sqrt{1-s^2}}{2}\big)=1-H_2(0.8536)=\mathbf{0.399\ bit}$ . The bound is loose by $\chi-I_{\mathrm{acc}}=\mathbf{0.202\ bit}$ : overlapping states advertise $0.601$ bit of distinguishability, but a single measurement can cash in only $0.399$ bit.

Step 4. Sanity checks at the extremes. Orthogonal ( $s=0$ ): $\chi=1$ bit and $I_{\mathrm{acc}}=1$ bit — everything readable. Identical ( $s=1$ ): both vanish. The bound $\chi=0.601$ bit is real information, but only $I_{\mathrm{acc}}=0.399$ bit is single-shot accessible.

2 · A mixed-state ensemble — computing χ with an impure letter

Alice sends the pure state $|0\rangle$ with probability $\tfrac12$ and the maximally mixed state $\tfrac{\mathbb{1}}{2}$ with probability $\tfrac12$ . Find $\chi$ .

Step 1. The average state is $\rho=\tfrac12|0\rangle\langle0|+\tfrac12\cdot\tfrac{\mathbb{1}}{2}=\tfrac34|0\rangle\langle0|+\tfrac14|1\rangle\langle1|$ , diagonal with eigenvalues $0.75$ and $0.25$ , so $S(\rho)=H_2(0.75)=\mathbf{0.8113\ bit}$ .

Step 2. The average of the letter entropies is $\sum_x p_x S(\rho_x)=\tfrac12\,S(|0\rangle\langle0|)+\tfrac12\,S(\tfrac{\mathbb{1}}{2})=\tfrac12\cdot 0+\tfrac12\cdot 1=\mathbf{0.5\ bit}$ , since a pure state has zero entropy and the maximally mixed qubit has entropy $\log_2 2=1$ .

Step 3. Therefore $\chi=S(\rho)-\sum_x p_x S(\rho_x)=0.8113-0.5=\mathbf{0.311\ bit}$ . The second letter's own entropy is subtracted off — mixedness in the letters costs Holevo information, which is why noisy channels shrink capacity. $\chi=0.311$ bit, well under the $\log_2 2=1$ bit ceiling.

Physics of the portrait

The Holevo bound is where classical information theory and quantum mechanics were first forced to reconcile: Shannon's bits meet Hilbert space, and the meeting has a hard limit.

Origins · 1 minAlexander Holevo and how this was discovered

Alexander Holevo

b. 1943 · quantum information theory and the capacity of quantum channels

Alexander Semenovich Holevo, born in 1943, is the Russian mathematician who founded the quantitative theory of information transmission over quantum channels. In 1973, at the Steklov Mathematical Institute in Moscow, he proved the inequality now bearing his name — establishing, against a then-common intuition that non-orthogonal states might pack unlimited data into a single system, that the accessible information is capped by $\chi$ . The result was decades ahead of the field it would anchor: only in the 1990s, when quantum information became a discipline, did its full weight land. Holevo went on to contribute the converse and, in 1998, part of the achievability of the Holevo–Schumacher–Westmoreland theorem, identifying maximized $\chi$ as the product-state classical capacity of a quantum channel — the exact analogue of Shannon's 1948 noisy-channel coding theorem. His monograph Probabilistic and Statistical Aspects of Quantum Theory built the rigorous foundations of quantum measurement and channel theory. The bound is his signature: a clean entropic inequality that turned "how much classical information fits in a quantum state" from intuition into a theorem.

History

1948

— Shannon founds information theory with the noisy-channel coding theorem

1973

— Holevo proves the bound χ ≥ I(X;Y) for quantum ensembles

1973

— Lieb and Ruskai prove strong subadditivity of von Neumann entropy

1996

— Hausladen, Jozsa, Schumacher, Westmoreland and Wootters achieve χ for pure-state ensembles

1998

— Schumacher, Westmoreland and Holevo establish χ as the product-state classical channel capacity

Going further

Texts. Nielsen and Chuang, Quantum Computation and Quantum Information, section 12.1.1 (Holevo bound) · Holevo, "Bounds for the quantity of information transmitted by a quantum communication channel," Problemy Peredachi Informatsii 9, 3 (1973) · Wilde, Quantum Information Theory, 2nd ed., chapters 11–13 · Watrous, The Theory of Quantum Information, chapter 5 · Preskill, Lecture Notes for Physics 219, chapter 10 · Schumacher and Westmoreland, Phys. Rev. A 56, 131 (1997); Holevo, IEEE Trans. Inf. Theory 44, 269 (1998); Hastings, Nature Physics 5, 255 (2009). The key entropic identities to keep in mind: the block-diagonal split $S(\rho_{XQ})=H(X)+\sum p_x S(\rho_x)$ yields $I(X;Q)=\chi$ ; the relative-entropy form gives $\chi=\sum p_x D(\rho_x\|\rho)\ge0$ ; and the staircase $0\le I_{\mathrm{acc}}\le\chi\le S(\rho)\le\log d$ holds in full generality. The two-pure-state anchor values at $s=1/\sqrt2$ are $\chi=H_2(\tfrac{1+s}{2})=0.601$ bit and $I_{\mathrm{acc}}=1-H_2(\tfrac{1+\sqrt{1-s^2}}{2})=0.399$ bit; the mixed-ensemble value is $\chi=H_2(0.75)-0.5=0.311$ bit. Note: the correct single-copy Holevo quantity for two pure states is $\chi=H_2(\tfrac{1+s}{2})$ in the overlap $s$ ; the form with $\sin\theta$ instead of $\cos\theta$ gives the accessible information, not $\chi$ .

Prerequisites. Von Neumann Entropy · Quantum Joint, Conditional Entropy and Mutual Information · Data-Processing Inequality · Shannon Entropy and Classical Information Measures. Leads to. Holevo–Schumacher–Westmoreland Theorem · Quantum Channels and Completely Positive Trace-Preserving Maps · Schumacher Quantum Data Compression · Strong Subadditivity of Quantum Entropy.

Ask the tutor: Want to see why data processing is equivalent to strong subadditivity, or work through the accessible information of a specific ensemble by hand?

Is the Holevo bound the same as the channel capacity?

Not quite — it is one half of it. The Holevo bound (2) is the converse: no measurement extracts more than $\chi$ from a single copy. The Holevo–Schumacher–Westmoreland capacity (7) adds the achievability: over many channel uses, random coding with a collective decoding measurement attains a rate approaching $\chi$ maximized over inputs. The bound tells you the ceiling; the capacity theorem tells you the ceiling is reachable asymptotically. On a single copy the bound is usually loose, which is exactly why the capacity needs a many-copy, block-coding argument.

Why is χ not simply the accessible information?

Because measuring a non-commuting alphabet disturbs it, and no single apparatus is simultaneously optimal for letters that do not share an eigenbasis. $\chi$ counts the correlation in the classical–quantum state, $I(X;Q)$ ; the accessible information counts what a real measurement retrieves, $\max I(X;Y)$ . Data processing forces $I_{\mathrm{acc}}\le\chi$ , but the inequality is strict whenever the $\rho_x$ fail to commute — as the worked example shows, $\chi=0.601$ bit while $I_{\mathrm{acc}}=0.399$ bit. Only over many copies, with collective measurements, does the per-copy accessible information climb to $\chi$ .

How can χ depend on the ensemble if the average state ρ is fixed?

This is a genuine subtlety. The same density matrix $\rho$ can be written as $\sum_x p_x\rho_x$ in infinitely many ways, and different decompositions give different $\chi$ . The maximally mixed qubit $\tfrac{\mathbb{1}}{2}$ is an equal mix of $\{|0\rangle,|1\rangle\}$ with $\chi=1$ bit, or of infinitely many overlapping states with smaller $\chi$ . The Holevo quantity is a property of the preparation — the labelled ensemble — not of the state alone, which is why $\chi=\sum_x p_x D(\rho_x\|\rho)$ references both the letters and their average.

Does the bound mean a qubit "holds" only one bit of information?

Only one classical bit can be reliably read out, which differs from what the qubit contains. A qubit's state is a point on the Bloch sphere — continuous parameters, so in a loose sense infinite information. But you cannot retrieve them as classical data: any single measurement collapses the state and returns at most $\log_2 2=1$ bit by the Holevo bound. The extra structure powers quantum algorithms and entanglement-based protocols, but as a channel for classical letters a qubit is exactly as capacious as a classical bit.

What role does strong subadditivity play here?

It is the deep engine underneath. The Holevo bound rests on the data-processing inequality — that a local channel cannot increase quantum mutual information — and that monotonicity is mathematically equivalent to strong subadditivity (SSA) of the von Neumann entropy, $S(ABC)+S(B)\le S(AB)+S(BC)$ , proved by Lieb and Ruskai in 1973. So every use of the Holevo bound is, at bottom, an application of SSA — and any proof of SSA immediately re-proves Holevo.

How does the Holevo bound secure quantum key distribution?

In QKD, an eavesdropper Eve holds a quantum system correlated with the raw key. Her total knowledge is capped by the Holevo quantity $\chi(X;E)$ of her side-information ensemble — she can never learn more, even with a quantum memory and delayed collective attacks. The Devetak–Winter theorem gives a provable secret-key rate $r=I(X;B)-\chi(X;E)$ : Alice and Bob keep the mutual information they share, minus the Holevo bound on what Eve could know. Because $\chi(X;E)$ is estimated from the observed error rate, the bound converts measured noise directly into a guaranteed number of secure bits — the heart of QKD security proofs.