Quantum Channel Capacities and Distinguishability Measures

A noisy quantum channel is a completely positive trace-preserving map $\mathcal{N}:\mathcal{B}(\mathcal{H}_A)\to\mathcal{B}(\mathcal{H}_B)$ — the open-system evolution that carries an input state through a physical medium, be it an optical fibre, a superconducting bus, or the idle time of a qubit in its own noisy environment. Quantum Shannon theory asks the operational question: given many independent uses of $\mathcal{N}$ , at what asymptotic rate can we push information through with vanishing error? The classical answer is one number, Shannon's capacity. The quantum answer is a family, because quantum mechanics traffics in several inequivalent currencies — classical bits, private bits, qubits, and shared entanglement — and a channel can be rich in one and bankrupt in another. Two threads run through this article. The first is capacity: the classical capacity $C$ (Holevo–Schumacher–Westmoreland) and the quantum capacity $Q$ (Lloyd–Shor–Devetak) are both regularized, multi-letter optimizations, because their entropic quantities are non-additive — entangling inputs across uses can strictly raise the rate, and two channels each with $Q=0$ can combine to give $Q>0$ , the purely quantum effect of superactivation. The entanglement-assisted capacity $C_E$ , by contrast, is single-letter and perfectly tame. The second thread is distinguishability: certifying that a device approximates an ideal gate needs a metric on channels, and the diamond norm is the operationally correct one because it accounts for entanglement with a protected reference. The average fidelity of benchmarking, the process fidelity of tomography, and the fault-tolerance thresholds of error correction are all quoted through these channel distances. This flagship synthesis takes the CPTP, Kraus, and specific-noise formalisms upstream and hands the results to error correction and benchmarking downstream.

Key takeaways

A channel has several inequivalent capacities — classical $C$ , private $P$ , quantum $Q$ , entanglement-assisted $C_E$ — each the optimal rate (bits or qubits per use) with vanishing error.
The HSW theorem gives $C=\lim_n \tfrac1n\chi(\mathcal{N}^{\otimes n})$ (regularized Holevo information); LSD gives $Q=\lim_n\tfrac1n\max I_c(\mathcal{N}^{\otimes n})$ (regularized coherent information).
Regularization is physics, not bookkeeping: coherent information is superadditive, single-letter formulas generally fail, and $Q$ can even be superactivated ( $Q_1=Q_2=0$ yet $Q_{1\otimes2}>0$ ).
Channel distance is the diamond norm $\|\mathcal{N}-\mathcal{M}\|_\diamond$ , stabilized by a reference, fixing the one-shot discrimination probability $p_{\mathrm{succ}}=\tfrac12+\tfrac14\|\mathcal{N}-\mathcal{M}\|_\diamond$ .
Quality is reported as average and process fidelity, tied by $\bar F=(d\,F_{\mathrm{pro}}+1)/(d+1)$ ; coherent errors make the diamond distance scale like $\sqrt{r}$ , so fidelity alone overstates margins.
Entanglement assistance collapses it to one formula, $C_E=\max_\rho I(A;B)$ — additive, never superactivated.
Key idea: Each resource defines its own capacity through an optimized entropic quantity — classical to Holevo information, quantum to coherent information, entanglement-assisted to quantum mutual information. Non-additivity forces regularization for $C$ and $Q$ but spares $C_E$ , while the diamond norm supplies the operational distance from a real channel to an ideal one.

The core insight

Each transmittable resource defines its own capacity through an optimized entropic quantity: classical → Holevo information, private/quantum → coherent information, entanglement-assisted → quantum mutual information. Non-additivity forces regularization for $C$ and $Q$ but spares $C_E$ , while the diamond norm supplies the operational distance that quantifies how close a real channel is to an ideal one.

Where this lives today

Channel capacities and the diamond norm are now the working vocabulary of quantum-device engineering. In fault-tolerant architectures the depolarizing hashing bound is the yardstick for code performance — the surface code's circuit threshold near $p\sim1\%$ against the capacity that noise leaves on the table — and certification increasingly targets $\|\mathcal{E}-\mathrm{id}\|_\diamond$ rather than fidelity, since the diamond distance is what enters threshold theorems. Benchmarking extracts $\bar F$ and converts it through $\bar F=(dF_{\mathrm{pro}}+1)/(d+1)$ ; because coherent errors inflate the diamond distance relative to $r=1-\bar F$ , randomized compiling tailors coherent noise into stochastic Pauli noise. In communication, $C_E$ sets ultimate rates over lossy links — the pure-loss channel of transmissivity $\eta$ has secret-key capacity $-\log_2(1-\eta)$ (the PLOB bound), whose linear scaling at high loss is what repeaters are built to beat. And a live open problem anchors it all: $Q$ has no known single-letter formula and is not known to be computable in general.

The physical picture

Every capacity here is a contest between two receivers. Purify the channel: by Stinespring's theorem any $\mathcal{N}$ is a unitary $U$ on the input and a fresh environment, after which one output is discarded. The intended receiver keeps $B$ ; the environment steals $E$ ; and whatever leaks into $E$ is, by no-cloning, unavailable at $B$ . The quantity that scores this tug-of-war is the coherent information $I_c=S(B)-S(E)$ , the receiver's entropy advantage over the environment, and it can be negative. When the environment wins for every input — the channel is antidegradable — no quantum information survives and $Q=0$ . The first figure is this picture.

Figure

The dilation picture: a channel is a unitary U on system plus environment, then one output is discarded. The receiver keeps B, the environment steals E, and the coherent information I_c = S(B) − S(E) scores who wins — for the identity channel S(E) = 0, so I_c = S(B) = 1 qubit per use

Turning $I_c$ into a capacity exposes the strangeness of the quantum theory. Classically, capacity is single-letter — one entropic quantity for one use suffices, because Shannon's mutual information is additive. Quantumly, coherent information is superadditive: correlating inputs across uses, or pairing two channels, can unlock rate no single use reveals. So $Q$ is a regularized limit over ever-longer blocks, generally uncomputable, and in the extreme it superactivates. The concrete anchor is the depolarizing channel — the second figure is its exactly-computable coherent information, the number to remember here.

Figure

The verified punchline: single-letter coherent information of the qubit depolarizing channel, I_c(p) = 1 − H(1−3p/4, p/4, p/4, p/4). It begins at 1 qubit per use for the noiseless channel, falls to 0.497 at p = 0.10, and crosses zero at p* ≈ 0.2524 — beyond which the hashing bound is zero and the channel is antidegradable

Distinguishability tells the same story from the metric side. Telling two channels apart is harder than telling two states apart, because you also choose how to probe — and the best probe is often one half of an entangled pair whose partner you keep protected. The diamond norm bakes this in, maximizing the output trace distance over all inputs including entangled ones; that stabilization is why entangled probes can strictly help and why the naive trace norm of outputs is the wrong measure. The same reference-system trick that makes coherent information the right capacity quantity makes the diamond norm the right distance.

The core equations

The classical currency comes first. The Holevo quantity measures how much classical information an optimal input ensemble pushes through in one use — the output entropy of the average state minus the average output entropy. It is the achievable single-letter rate of the HSW theorem.

\chi(\mathcal{N})=\max_{\{p_x,\rho_x\}}\Big[S\Big(\mathcal{N}\big(\textstyle\sum_x p_x\rho_x\big)\Big)-\sum_x p_x\,S\big(\mathcal{N}(\rho_x)\big)\Big] \qquad (1)

where: $\chi(\mathcal{N})$ — Holevo quantity, maximized over ensembles; $\{p_x,\rho_x\}$ — signal states $\rho_x$ with probabilities $p_x$ ; $S(\rho)=-\mathrm{tr}(\rho\log_2\rho)$ — von Neumann entropy in bits. Reading it: the spread of the average output minus the noise added to each signal is the classical information the receiver can decode. The classical capacity is the regularized limit of $\chi$ : achievability per use is HSW, but the converse plus superadditivity force the block limit.

C(\mathcal{N})=\lim_{n\to\infty}\frac{1}{n}\,\chi\!\left(\mathcal{N}^{\otimes n}\right) \qquad (2)

where: $C(\mathcal{N})$ — classical capacity in bits per use; $\mathcal{N}^{\otimes n}$ — $n$ parallel uses. Reading it: if $\chi$ were additive this would collapse to $\chi(\mathcal{N})$ , but Hastings' 2009 counterexample shows $\chi(\mathcal{N}_1\otimes\mathcal{N}_2)>\chi(\mathcal{N}_1)+\chi(\mathcal{N}_2)$ can occur, so the limit is genuine. The quantum currency is scored by coherent information — the environment-subtracted output entropy.

I_c(\rho,\mathcal{N})=S\big(\mathcal{N}(\rho)\big)-S\big(\mathcal{N}^c(\rho)\big)=S(B)-S(E) \qquad (3)

where: $I_c$ — coherent information of $\rho$ through $\mathcal{N}$ ; $\mathcal{N}^c$ — the complementary channel to the environment; $S(B),S(E)$ — output and environment entropies. Reading it: $I_c>0$ means the receiver holds more of the input's quantum correlations than the environment; $I_c<0$ means the environment has won. The quantum capacity is the regularized, optimized coherent information (LSD).

Q(\mathcal{N})=\lim_{n\to\infty}\frac{1}{n}\max_{\rho^{(n)}} I_c\!\left(\rho^{(n)},\mathcal{N}^{\otimes n}\right) \qquad (4)

where: $Q(\mathcal{N})$ — quantum capacity in qubits per use; $\rho^{(n)}$ — an input over $n$ uses, possibly entangled across them. Reading it: the single-letter $\max_\rho I_c(\rho,\mathcal{N})$ is only the hashing bound $Q\ge I_c$ ; degenerate codes beat it, so the block limit is essential. Give the sender and receiver free pre-shared entanglement and the theory collapses to one line — the quantum mutual information of the output.

C_E(\mathcal{N})=\max_{\rho}\,I(A;B)_\omega,\qquad \omega_{AB}=(\mathrm{id}\otimes\mathcal{N})(|\phi_\rho\rangle\langle\phi_\rho|) \qquad (5)

where: $C_E$ — entanglement-assisted classical capacity; $I(A;B)=S(A)+S(B)-S(AB)$ — quantum mutual information; $|\phi_\rho\rangle$ — a purification of $\rho$ . Reading it: since $I(A;B)$ is concave in $\rho$ and additive, this is a clean convex problem with a single-letter answer. Now the distinguishability side. The diamond norm is the completely bounded trace norm — the largest output trace distance an entangled probe can achieve on a reference of dimension $\ge\dim A$ .

\|\mathcal{N}-\mathcal{M}\|_\diamond=\max_{\rho_{RA}}\big\|(\mathrm{id}_R\otimes\mathcal{N})(\rho_{RA})-(\mathrm{id}_R\otimes\mathcal{M})(\rho_{RA})\big\|_1 \qquad (6)

where: $\|\cdot\|_\diamond$ — diamond-norm distance; $\mathrm{id}_R$ — identity on the reference $R$ ; $\|X\|_1=\mathrm{tr}\sqrt{X^\dagger X}$ — trace norm. Reading it: the reference $R$ is the whole point — dropping it can strictly underestimate the distance. The diamond norm earns its name operationally: it sets the best one-shot probability of telling which channel acted.

p_{\mathrm{succ}}=\frac{1}{2}+\frac{1}{4}\,\|\mathcal{N}-\mathcal{M}\|_\diamond \qquad (7)

where: $p_{\mathrm{succ}}$ — optimal probability of naming which channel acted, one use, equal prior. Reading it: identical channels give distance $0$ and a coin-flip $\tfrac12$ ; maximally different channels give distance $2$ and perfect discrimination. In practice hardware reports fidelities, not diamond distances, and the average gate fidelity and the process (entanglement) fidelity are rigidly linked by a dimension factor.

\bar F(\mathcal{N})=\frac{d\,F_{\mathrm{pro}}(\mathcal{N})+1}{d+1},\qquad F_{\mathrm{pro}}(\mathcal{N})=\langle\Phi|(\mathrm{id}\otimes\mathcal{N})(|\Phi\rangle\langle\Phi|)|\Phi\rangle \qquad (8)

where: $\bar F$ — fidelity to the identity averaged over pure inputs; $F_{\mathrm{pro}}$ — process fidelity, the Choi-state overlap with $|\Phi\rangle$ ; $d$ — dimension. Reading it: benchmarking measures $\bar F$ , tomography measures $F_{\mathrm{pro}}$ , and this exact relation converts between them. Finally, the bridge to fault tolerance: fidelity bounds the diamond distance, but loosely, and the looseness is where coherent errors hide.

1-F_{\mathrm{pro}}(\mathcal{N})\le \tfrac{1}{2}\|\mathcal{N}-\mathrm{id}\|_\diamond\le \sqrt{1-F_{\mathrm{pro}}(\mathcal{N})^2} \qquad (9)

where: $\mathrm{id}$ — the ideal target channel. Reading it: the lower bound is linear in the infidelity, the upper bound scales like its square root, so a gate excellent by fidelity can still be diamond-far from ideal.

Derivation ACoherent information and the quantum capacity of the depolarizing channel

The goal is to compute from scratch the single-letter coherent information of the qubit depolarizing channel and read off the hashing lower bound on its quantum capacity, grounding the second figure in an explicit calculation rather than a quoted result. We exploit the channel's Pauli symmetry to fix the optimal input, evaluate $S(B)$ and $S(E)$ separately, and locate the threshold where the bound vanishes.

Step 1. Write the channel in Pauli form. Using $\tfrac{I}{2}=\tfrac14(\rho+X\rho X+Y\rho Y+Z\rho Z)$ , the channel $\mathcal{D}_p(\rho)=(1-p)\rho+p\,\tfrac{I}{2}$ becomes $\mathcal{D}_p(\rho)=(1-\tfrac{3p}{4})\rho+\tfrac{p}{4}(X\rho X+Y\rho Y+Z\rho Z)$ — a Pauli channel with probabilities $(1-\tfrac{3p}{4},\tfrac{p}{4},\tfrac{p}{4},\tfrac{p}{4})$ . Why: the Pauli form makes the environment diagonal, so $S(E)$ is a Shannon entropy of these four numbers.

Step 2. Fix the optimal input by symmetry. The channel is covariant under every single-qubit unitary, so the optimal $\rho$ must be invariant under that group, forcing the maximally mixed $\rho=I/2$ fed as one half of a maximally entangled pair. Why: a group-symmetric optimization is solved at a fixed point of the group.

Step 3. Evaluate the receiver's entropy. For the maximally mixed input $\mathcal{D}_p(I/2)=I/2$ , so $S(B)=\log_2 2=1$ . Why: the receiver sees a full qubit of entropy regardless of $p$ ; all the $p$ -dependence lives in $S(E)$ .

Step 4. Environment entropy and combine. The complementary output has eigenvalues equal to the Pauli probabilities, so $S(E)=H\big(1-\tfrac{3p}{4},\tfrac{p}{4},\tfrac{p}{4},\tfrac{p}{4}\big)$ and therefore

I_c(\mathcal{D}_p)=S(B)-S(E)=1+\big(1-\tfrac{3p}{4}\big)\log_2\big(1-\tfrac{3p}{4}\big)+\tfrac{3p}{4}\log_2\tfrac{p}{4}

the hashing rate $Q\ge I_c$ . Why: the environment's Shannon entropy subtracted from the receiver's one bit is exactly $S(B)-S(E)$ .

Step 5. Locate the threshold and check limits. At $p=0$ , $I_c=1$ (noiseless qubit); solving $I_c=0$ gives $p^*\approx 0.2524$ , beyond which $\mathcal{D}_p$ is antidegradable and $Q=0$ ; as $p\to 1$ , $I_c\to -1$ . Why: the endpoints reproduce the noiseless and maximally-noisy channels, and the crossing is the antidegradability boundary in the figure. *These values check out directly from $I_c=1-H(1-3p/4,p/4,p/4,p/4)$ : $I_c(0)=1$ , $I_c(0.10)=0.497$ , and the zero-crossing is at $p^*=0.2524$ . Near threshold the true $Q$ exceeds this hashing bound because degenerate codes can exploit the channel's structure (DiVincenzo–Shor–Smolin).*

Go deeper · Derivation B · 1 minWhy entanglement assistance gives a single-letter capacity

Derivation BWhy entanglement assistance gives a single-letter capacity

The goal is to show why pre-shared entanglement turns the messy regularized capacity problem into the cleanest formula in quantum Shannon theory — the quantum mutual information — needing no block limit. We identify the three entropies of the mutual information on a purified input, invoke the entanglement-assisted coding theorem, use additivity to kill the regularization, and read off the ordering it imposes.

Step 1. Expand the mutual information on a purification. Send half of $|\phi_\rho\rangle_{AA'}$ through the channel; then $S(A)=S(\rho)$ , $S(B)=S(\mathcal{N}(\rho))$ , and since $ABE$ is pure, $S(AB)=S(E)=S(\mathcal{N}^c(\rho))$ , giving $I(A;B)=S(\rho)+S(\mathcal{N}(\rho))-S(\mathcal{N}^c(\rho))$ . Why: purity trades the joint entropy $S(AB)$ for the environment entropy $S(E)$ .

Step 2. Attach the operational meaning. The Bennett–Shor–Smolin–Thapliyal theorem shows that with unlimited shared entanglement the sender attains exactly $I(A;B)$ bits per use. Why: this makes $I(A;B)$ a capacity, not just an entropic expression.

Step 3. Remove the regularization. Quantum mutual information is additive, $I(\mathcal{N}_1\otimes\mathcal{N}_2)=I(\mathcal{N}_1)+I(\mathcal{N}_2)$ , and $I(A;B)$ is concave in $\rho$ , so product inputs are optimal and $C_E=\max_\rho I(A;B)$ with no limit. Why: the sub- and super-additive terms cancel exactly — the reason $C_E$ is single-letter where $C$ and $Q$ are not.

Step 4. Read off the ordering. Since assistance can only help, $Q(\mathcal{N})\le\tfrac12 C_E(\mathcal{N})$ and $C(\mathcal{N})\le C_E(\mathcal{N})$ , and $C_E$ is never superactivated. Why: this makes $C_E$ the unique well-behaved capacity and a computable ceiling on the others. *For the noiseless qubit channel $S(\mathcal{N}^c)=0$ , so $\max_\rho I(A;B)=2\log_2 2=2$ bits — the superdense-coding value $C_E(\mathrm{id}_2)=2$ .*

Go deeper · Derivation C · 1 minDiamond distance versus average gate fidelity

Derivation CDiamond distance versus average gate fidelity

The goal is to connect the fidelity a benchmarking experiment reports to the worst-case diamond distance a fault-tolerance threshold requires, exposing the gap as the fingerprint of coherent (unitary) errors. We invert the average-to-process fidelity relation, bound the diamond distance from both sides, and quantify how badly fidelity can flatter a gate whose errors are coherent rather than stochastic.

Step 1. Define the error rate. With $\mathcal{E}=\mathcal{N}\circ\mathrm{id}^{-1}$ the error channel, invert equation (8): $r\equiv 1-\bar F=\tfrac{d+1}{d}\,(1-F_{\mathrm{pro}})$ , linear in the process infidelity. Why: benchmarking delivers $r$ , and everything downstream must be phrased in it.

Step 2. Bound it both ways. Fidelity controls the diamond distance only within a band,

\frac{d-1}{d}\,r \;\le\; \tfrac12\|\mathcal{E}-\mathrm{id}\|_\diamond \;\le\; \sqrt{d(d+1)\,r}

the lower bound linear in $r$ , the upper bound scaling like $\sqrt{r}$ . Why: a single number $r$ cannot pin the diamond distance — the same fidelity spans a range of worst-case behaviours.

Step 3. Attribute the gap to error type. Stochastic Pauli noise saturates the lower bound ( $\sim r$ ), while purely coherent unitary errors approach the upper bound ( $\sim\sqrt r$ ), vastly larger for small $r$ . Why: coherent errors add in amplitude and interfere constructively across a circuit, so their worst-case cost grows like $\sqrt r$ .

Step 4. Draw the fault-tolerance moral. A gate with $r=10^{-4}$ but coherent error can have $\tfrac12\|\mathcal{E}-\mathrm{id}\|_\diamond\sim 10^{-2}$ — the threshold quantity — which is why randomized compiling and Pauli twirling convert coherent error into stochastic noise. Why: thresholds are stated in diamond norm, so fidelity alone can overstate the safety margin by two orders of magnitude. *For $d=2$ and $r=10^{-4}$ the Wallman–Flammia and Kueng et al. bounds give $5\times10^{-5}\le\tfrac12\|\mathcal{E}-\mathrm{id}\|_\diamond\le 0.0245$ , a $\sim 500\times$ spread.*

Limiting cases

Special channels collapse the intractable regularized formulas to closed forms:

Case	Result	Meaning
Noiseless qubit channel	$Q(\mathrm{id}_2)=C(\mathrm{id}_2)=1,\quad C_E(\mathrm{id}_2)=2$	One qubit or one bit per use; shared entanglement doubles $C$ to 2 bits via superdense coding.
Degradable channel	$Q(\mathcal{N})=\max_\rho I_c(\rho,\mathcal{N})$	Coherent information is additive, so the single-letter formula is exact (e.g. dephasing). No regularization.
Antidegradable channel	$Q(\mathcal{N})=0$	The environment can simulate the receiver, so no quantum information survives; e.g. depolarizing $p>p^*\approx0.2524$ , or 50% erasure.
Entanglement-breaking channel	$Q(\mathcal{N})=0,\ \ C(\mathcal{N})=\chi(\mathcal{N})$	Output always separable from any reference; $Q$ vanishes and $C$ is single-letter.
Identical channels ( $\mathcal{N}=\mathcal{M}$ )	$\\|\mathcal{N}-\mathcal{M}\\|_\diamond=0,\ p_{\mathrm{succ}}=\tfrac12$	Nothing to distinguish; the best guess is a coin flip.

Common trapQ = 0 does not mean the channel is useless for quantum communication

It is tempting to read the single-letter formulas as the capacities and conclude $Q(\mathcal{N})=0$ means the channel can never transmit quantum information. Both readings are wrong. First, $\chi(\mathcal{N})$ and single-letter $I_c(\mathcal{N})$ are only lower bounds whose regularizations equal the capacities; coherent information is strictly superadditive, so the block limit can exceed the one-shot value. Second, the zero-capacity channels are not closed under tensoring: two with $Q=0$ (one antidegradable, one PPT) can combine to give $Q>0$ — Smith–Yard superactivation. A separate trap: never substitute the unstabilized trace norm of Choi states for the diamond norm — omitting the reference underestimates the distance and destroys the operational meaning behind fault-tolerance thresholds.

Worked examples

1 · Quantum capacity of a depolarizing channel (fully worked)

A qubit sits in a depolarizing environment with $p=0.10$ . Estimate its quantum capacity via the hashing bound, and find the noise level at which quantum communication shuts off.

Step 1. Pauli distribution. With $p=0.10$ the error probabilities are $\big(1-\tfrac{3p}{4},\tfrac{p}{4},\tfrac{p}{4},\tfrac{p}{4}\big)=(0.925,\,0.025,\,0.025,\,0.025)$ .

Step 2. Environment entropy. $S(E)=H(0.925,0.025,0.025,0.025)=-0.925\log_2 0.925-3(0.025\log_2 0.025)=0.1041+0.3992=\mathbf{0.5032}$ bits.

Step 3. Coherent information. $I_c=S(B)-S(E)=1-0.5032=\mathbf{0.497}$ qubits per use, so $Q\ge 0.497$ — about half a protected qubit per use.

Step 4. Threshold. Solving $H(1-\tfrac{3p}{4},\tfrac{p}{4},\tfrac{p}{4},\tfrac{p}{4})=1$ gives $p^*\approx\mathbf{0.2524}$ ; for $p>p^*$ the channel is antidegradable and $Q=0$ . The true $Q$ actually exceeds the hashing bound in a small window just below $p^*$ (superadditive degenerate codes), so $0.2524$ is a conservative shut-off. Result: $I_c(0.10)=0.497$ qubits/use, shut-off at $p^*\approx0.2524$ .

2 · Diamond distance of a high-fidelity gate

A single-qubit gate ( $d=2$ ) is benchmarked at $\bar F=0.9999$ , i.e. $r=10^{-4}$ . Bound its diamond distance to the ideal gate and interpret the spread.

Step 1. Lower bound. $\tfrac{d-1}{d}\,r=\tfrac12\cdot10^{-4}=\mathbf{5\times10^{-5}}$ , for stochastic Pauli noise.

Step 2. Upper bound. $\sqrt{d(d+1)\,r}=\sqrt{6\times10^{-4}}=\mathbf{0.0245}$ , for purely coherent (unitary) error.

Step 3. Interpret. The diamond distance lies anywhere in $[5\times10^{-5},\,0.0245]$ — a factor of $\sim\mathbf{500}$ — so a gate superb by fidelity may be two orders of magnitude worse against a threshold. This is why twirling coherent error into stochastic noise is worth the overhead. Result: $5\times10^{-5}\le\tfrac12\|\mathcal{E}-\mathrm{id}\|_\diamond\le 0.0245$ .

Physics of the portrait

Holevo information, the classical capacity, the accessible information — these trace to one person who proved, before anyone could build a qubit, the sharpest limit on how much classical data a quantum system can carry.

Origins · 1 minAlexander Holevo and how this was discovered

Alexander Holevo

born 1943, Soviet and Russian mathematical physicist · quantum Shannon theory and channel capacities

Alexander Semenovich Holevo was working at the Steklov Mathematical Institute in Moscow in the early 1970s, in near-isolation from the Western information-theory community, when he proved what became the cornerstone of quantum Shannon theory. His 1973 bound settled a question of Gordon and Levitin: how much classical information can be recovered from a quantum system prepared in one of several states? Holevo showed the accessible information can never exceed the quantity $\chi$ that now bears his name. For two decades this was a converse without a matching achievability result, until in 1997–98 Holevo, and independently Schumacher and Westmoreland, proved $\chi$ is in fact achievable per use, completing the HSW theorem. Holevo went on to build much of the rigorous theory of quantum channels and Gaussian bosonic capacities; his monograph Quantum Systems, Channels, Information is a standard reference. His career is a reminder that the deepest limits of quantum communication were mapped out mathematically decades before the hardware existed to test them.

History

1973

— Holevo bounds the classical information extractable from a quantum ensemble

1996

— Schumacher and Nielsen introduce coherent information and the quantum data-processing inequality

1997

— Holevo, Schumacher and Westmoreland fix the classical capacity; Kitaev defines the diamond norm

2005

— Devetak proves the LSD quantum-capacity theorem

2008

— Smith and Yard superactivate the quantum capacity, then Hastings breaks Holevo additivity

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, Ch. 8 & 12 · Wilde, Quantum Information Theory (2nd ed.), Ch. 13–24 · Watrous, The Theory of Quantum Information, Ch. 3 & 8 · Preskill, Lecture Notes on Quantum Computation, Ch. 10 · Kitaev, Shen & Vyalyi, Classical and Quantum Computation · Holevo, Quantum Systems, Channels, Information. Key results: the depolarizing coherent information $I_c=1-H(1-3p/4,p/4,p/4,p/4)$ gives $I_c(0)=1$ , $I_c(0.10)=0.497$ , and zero-crossing $p^*\approx0.2524$ (sharper than the round figure near $0.25$ ); the average-to-process fidelity relation is $\bar F=(dF_{\mathrm{pro}}+1)/(d+1)$ ; the diamond–fidelity bounds for $d=2$ , $r=10^{-4}$ yield the range $[5\times10^{-5},\,0.0245]$ ; and $C_E(\mathrm{id}_2)=2$ . Superactivation: Smith–Yard (2008). Holevo additivity violation: Hastings (2009). PLOB bound: Pirandola et al. (2017).

Prerequisites. Quantum Operations and CPTP Maps · Kraus Representation of Quantum Channels · Open Quantum Systems and Reduced Dynamics · Depolarizing Channel. Leads to. Quantum Error Correction (codes & thresholds) · Randomized Benchmarking · Amplitude Damping Channel · Pauli and Bit/Phase-Flip Channels.

Ask the tutor: Want to see how degenerate codes beat the depolarizing hashing bound, or how randomized compiling closes the fidelity–diamond gap on a real device?

What is the difference between the classical, private, and quantum capacities?

Three currencies over the same channel. The classical capacity $C$ is the top rate of ordinary bits; the private capacity $P$ the top rate of bits that also stay secret from the environment; the quantum capacity $Q$ the top rate of qubits — intact states, superpositions and entanglement included. They obey $Q\le P\le C$ , since protecting coherence is at least as demanding as protecting secrecy, which is at least as demanding as delivering bits. A channel can have $C>0$ while $Q=0$ — passing classical data but destroying every superposition, as entanglement-breaking channels do.

Why can't we just use the single-letter coherent information as the quantum capacity?

Because coherent information is superadditive: entangling inputs across several uses can make the per-use value strictly larger than for any single use. So $\max_\rho I_c(\rho,\mathcal{N})$ is only the hashing bound $Q\ge I_c$ , and the capacity is the regularized limit $\lim_n\tfrac1n\max I_c(\mathcal{N}^{\otimes n})$ . For degradable channels (e.g. dephasing) it is additive and the single-letter formula is exact, but in general no finite $n$ suffices — which is why $Q$ has no known closed form and is not known to be computable.

What exactly does superactivation mean, and why is it impossible classically?

Two channels, each individually useless for quantum communication ( $Q_1=Q_2=0$ ), acquire a positive joint capacity together ( $Q_{1\otimes 2}>0$ ). Smith and Yard exhibited such a pair — one antidegradable, one PPT-bound. It is possible because the zero-capacity channels are not closed under tensoring, a consequence of superadditive coherent information. Classically it cannot happen: Shannon capacities are additive, so a zero-capacity channel stays zero in any combination. Superactivation is thus one of the sharpest signs that quantum capacity is not a per-use quantity.

Why is the diamond norm the right way to compare two channels?

Because it is stabilized: it optimizes the output trace distance over inputs entangled with a protected reference of arbitrary size. Entangled probes can strictly improve your ability to tell two channels apart, so any measure fixing an unentangled input underestimates the true distinguishability. Its payoff is exact — the best one-shot probability of identifying which of two equally likely channels acted is $\tfrac12+\tfrac14\|\mathcal{N}-\mathcal{M}\|_\diamond$ — and it is the quantity fault-tolerance thresholds are stated in.

If my gate has 99.99% fidelity, isn't it good enough for fault tolerance?

Not necessarily — fidelity can flatter a gate. The error rate $r=1-\bar F$ bounds the diamond distance only within a band: $\tfrac{d-1}{d}r\le\tfrac12\|\mathcal{E}-\mathrm{id}\|_\diamond\le\sqrt{d(d+1)r}$ . Stochastic Pauli noise gives $\sim r$ , but coherent (unitary) error scales like $\sqrt r$ , which for $r=10^{-4}$ on a qubit reaches $\approx0.0245$ — some 500 times the infidelity. Since thresholds are quoted in diamond norm, such a gate may sit right at the edge. The fix is randomized compiling or Pauli twirling, which convert coherent error into stochastic noise.

How does entanglement assistance double the classical capacity of a perfect qubit?

Unassisted, one use of a noiseless qubit channel carries one classical bit, $C(\mathrm{id}_2)=1$ . If sender and receiver already share an entangled pair, the sender applies one of four Pauli operations to her half and sends it; a Bell measurement on both halves reads out two bits. This is superdense coding, the content of $C_E(\mathrm{id}_2)=2$ . The extra bit is prepaid with shared entanglement, but in the entanglement-assisted accounting it doubles the throughput — which is why $C_E$ is the largest capacity.