Complementary Channels and Degradability

Every quantum channel is one half of a story. Its Stinespring dilation realizes the noisy map $\mathcal{N}:A\to B$ as a noiseless isometry $V$ that lifts the input into a larger space and then discards an environment $E$ . But the environment is not a bookkeeping fiction — it is a physical system that has learned something about the input, and keeping $E$ while discarding Bob's system $B$ defines a second, equally legitimate channel: the complementary channel $\mathcal{N}^c:A\to E$ . The pair $(\mathcal{N},\mathcal{N}^c)$ is a single coin with two faces. What Bob receives and what leaks to the environment are two views of one purification, and the entire theory of quantum capacity, private communication, and quantum key distribution lives in the tension between them. This article develops that tension into a working tool. Because the dilation is unique up to an isometry on $E$ , the complement is unique up to a partial isometry — so its output entropy $S(\mathcal{N}^c(\rho))$ is an intrinsic invariant, and the coherent information $I_c=S(\mathcal{N}(\rho))-S(\mathcal{N}^c(\rho))$ is well defined. Degradability then asks a sharp operational question: can Bob, holding only his output, run a box $\mathcal{D}$ that reproduces exactly what the environment sees, $\mathcal{D}\circ\mathcal{N}=\mathcal{N}^c$ ? When yes, the coherent information becomes additive and the quantum capacity collapses to a single-letter formula — a rarity where almost every capacity is an intractable regularized limit. When the reverse holds, $\mathcal{D}\circ\mathcal{N}^c=\mathcal{N}$ , no-cloning forces the quantum capacity to zero. Degradability is thus the fault line between the handful of channels we can solve exactly and the vast majority we cannot.

Key takeaways

Every channel $\mathcal{N}$ has a complementary channel $\mathcal{N}^c$ , sending the input to the environment $E$ of its Stinespring dilation; $\mathcal{N}$ is what Bob gets, $\mathcal{N}^c$ is what leaks to the eavesdropper.
The complement is unique up to an isometry on $E$ , so $S(\mathcal{N}^c(\rho))$ is intrinsic and the coherent information $I_c=S(\mathcal{N}(\rho))-S(\mathcal{N}^c(\rho))$ is well defined.
$\mathcal{N}$ is degradable if $\mathcal{D}\circ\mathcal{N}=\mathcal{N}^c$ for some channel $\mathcal{D}$ (Bob can simulate Eve) and antidegradable if $\mathcal{D}\circ\mathcal{N}^c=\mathcal{N}$ (Eve can simulate Bob).
For degradable channels $I_c$ is concave and additive, so the quantum capacity is the single-letter maximum $Q(\mathcal{N})=\max_\rho I_c(\rho,\mathcal{N})=P$ — no regularization — giving the exact capacities of dephasing, erasure, amplitude-damping, and pure-loss channels.
Antidegradable channels have $Q=0$ by no-cloning; amplitude damping above $\gamma=1/2$ , erasure above $p=1/2$ , and pure loss below $\eta=1/2$ are the canonical cases.
Key idea: the complementary channel $\mathcal{N}^c$ is what the environment learns; degradability ( $\mathcal{D}\circ\mathcal{N}=\mathcal{N}^c$ ) means Bob can reconstruct Eve's view, which makes the coherent information additive and gives a single-letter quantum capacity, while antidegradability ( $\mathcal{D}\circ\mathcal{N}^c=\mathcal{N}$ ) lets Eve clone Bob and forces $Q=0$ .

Where this lives today

Degradability is the first question a quantum communication engineer asks about any new channel, because the answer decides whether a capacity is a one-line formula or an open problem. In fibre-optic and free-space links the model is the bosonic pure-loss channel: degradable above transmissivity $\eta=1/2$ with exact capacity $\log_2[\eta/(1-\eta)]$ , and antidegradable below it — the 3 dB no-go that quantum communication cannot survive more than half loss, the motivation for quantum repeaters. In quantum key distribution the complement is the eavesdropper: Eve holds $E$ , and for degradable links the private capacity equals the coherent information, giving provable secret-key rates. At the frontier, the depolarizing channel — neither degradable nor antidegradable in its interesting range — is where superadditivity was first demonstrated and the exact quantum capacity is still unknown; the modern response is approximate degradability, an SDP measuring the diamond-norm distance to the degradable set and turning it into rigorous capacity bounds. Degradability still marks the boundary between the solvable and the open in quantum Shannon theory.

The physical picture

Think of the Stinespring isometry $V$ as a fork in an information road. The input $\rho$ climbs a lossless ramp — an isometry preserves everything — and the road splits into two lanes: Bob's system $B$ and the environment $E$ . The channel $\mathcal{N}$ watches $B$ and throws away $E$ ; the complement $\mathcal{N}^c$ watches $E$ and throws away $B$ . Nothing is destroyed at the fork — the loss is entirely the discarding. That is why, for a pure input, Bob and Eve carry equal entropy: they are the two marginals of one pure state, whose halves have matching spectra by Schmidt decomposition. The first figure fixes this fork in mind and marks the extra arrow — a degrading map $\mathcal{D}$ from $B$ back to $E$ — that decides everything.

Figure

The Stinespring fork sends input A through an isometry V and splits it into Bob's system B and the environment E; the channel is N = Tr_E VρV† and its complement is Nᶜ = Tr_B VρV†, while a degrading map D with D∘N = Nᶜ witnesses degradability

The fork explains why one number governs the subject. The coherent information $I_c=S(\mathcal{N}(\rho))-S(\mathcal{N}^c(\rho))$ is Bob's output entropy minus Eve's — how much more the legitimate receiver knows than the eavesdropper. If Bob can post-process his output into a copy of Eve's ( $\mathcal{D}\circ\mathcal{N}=\mathcal{N}^c$ ) he is uniformly better informed, and this advantage survives running many copies in parallel — the additivity that makes the capacity computable. The reverse, where Eve reconstructs Bob, is fatal: two parties would share one quantum message, which no-cloning forbids unless it is empty. The second figure turns this into hard numbers on the workhorse qubit amplitude-damping channel, whose regimes meet exactly at $\gamma=1/2$ .

Figure

Quantum capacity of the qubit amplitude-damping channel versus damping γ: degradable for γ below 1/2 where Q is the single-letter maximum of H2((1−γ)p) − H2(γp), reaching Q = 0.506 at γ = 0.2, and antidegradable for γ above 1/2 where no-cloning forces Q = 0 past the self-complementary point γ = 1/2

The core equations

The framework begins with the Stinespring dilation, which represents a noisy channel as a lossless isometry followed by a partial trace. It is the machine that manufactures both $\mathcal{N}$ and its complement.

V:\mathcal{H}_A\to\mathcal{H}_B\otimes\mathcal{H}_E,\quad V^\dagger V=\mathbb{1}_A,\quad \mathcal{N}(\rho)=\mathrm{Tr}_E\big[V\rho V^\dagger\big] \qquad (1)

where: $V$ — the Stinespring isometry into system $\otimes$ environment; $V^\dagger V=\mathbb{1}_A$ — the isometry condition (no loss yet); $\mathcal{H}_B,\mathcal{H}_E$ — Bob's and the environment's spaces; $\mathrm{Tr}_E$ — partial trace over the environment. Reading it: all the noise is in discarding $E$ ; before that trace the evolution is reversible. Keep the environment and discard Bob, and the same isometry defines the complement — the eavesdropper's channel, mapping the input to what the world has recorded.

\mathcal{N}^c(\rho)=\mathrm{Tr}_B\big[V\rho V^\dagger\big],\qquad V=\sum_k K_k\otimes\lvert k\rangle_E \qquad (2)

where: $\mathcal{N}^c$ — the complement $A\to E$ ; $\{K_k\}$ — Kraus operators with $\sum_k K_k^\dagger K_k=\mathbb{1}$ ; $\lvert k\rangle_E$ — an orthonormal environment basis. Reading it: the environment output has entries $(\mathcal{N}^c(\rho))_{kl}=\mathrm{Tr}[K_k\rho K_l^\dagger]$ , so few Kraus operators means a small $E$ . The single scalar tying the two channels together is the coherent information: Bob's output entropy minus Eve's. On a purified input it equals $-S(B\mid R)$ , a genuine measure of the quantum correlation surviving the channel.

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

where: $S(\sigma)=-\mathrm{Tr}[\sigma\log_2\sigma]$ — von Neumann entropy in bits; $R$ — a reference purifying $\rho$ ; $S(B\mid R)=S(BR)-S(R)$ — the (possibly negative) conditional entropy. Reading it: $I_c>0$ requires $S(B\mid R)<0$ , a hallmark of entanglement. Degradability and antidegradability are the two one-way orderings between a channel and its complement, each saying a CPTP map interpolates from one to the other.

\mathcal{D}\circ\mathcal{N}=\mathcal{N}^c\ \ (\text{degradable}),\qquad \mathcal{D}'\circ\mathcal{N}^c=\mathcal{N}\ \ (\text{antidegradable}) \qquad (4)

where: $\mathcal{D}$ — a degrading map from Bob's system to the environment; $\mathcal{D}'$ — a degrading map from the environment to Bob's system; $\circ$ — composition. Reading it: these must hold exactly as maps on all inputs — matching on one state is not enough — and a generic channel satisfies neither. The quantum capacity in general is the regularized coherent information — the Lloyd–Shor–Devetak (LSD) theorem — an optimization over entangled inputs on unboundedly many copies, hence uncomputable in general.

Q(\mathcal{N})=\lim_{n\to\infty}\frac1n\max_{\rho^{(n)}} I_c\big(\rho^{(n)},\mathcal{N}^{\otimes n}\big) \qquad (5)

where: $Q(\mathcal{N})$ — quantum capacity in qubits per use; $\mathcal{N}^{\otimes n}$ — $n$ parallel copies; $\rho^{(n)}$ — a possibly entangled input across copies. Reading it: the limit is forced because $I_c$ can be superadditive — entangled inputs beat $n$ times the single-copy value. Degradability dissolves that limit, fixes the private capacity, and kills antidegradable channels.

\underbrace{Q=\max_\rho I_c(\rho,\mathcal{N})=P}_{\text{degradable}},\qquad \underbrace{Q=0}_{\text{antidegradable}} \qquad (6)

where: $P$ — the private capacity; the left identity holds when $\mathcal{N}$ is degradable, the right when antidegradable. Reading it: for degradable channels one concave maximization replaces the intractable limit, and $P$ and $Q$ coincide.

Derivation AWhy the coherent information is Bob's entropy minus Eve's

The goal is to prove the working identity $I_c(\rho,\mathcal{N})=S(\mathcal{N}(\rho))-S(\mathcal{N}^c(\rho))=-S(B\mid R)$ , which turns an abstract "information that survives the channel" into two computable entropies of the branches of the Stinespring fork. We purify the input, push it through the isometry to a global pure state, and use the most useful fact about pure states — complementary marginals share a spectrum — to relate Eve's entropy to a reference-Bob entropy, so the same quantity reads three equivalent ways.

Step 1. Purify the input: introduce a reference $R$ and pure $\lvert\phi\rangle_{RA}$ with $\rho_A=\mathrm{Tr}_R\lvert\phi\rangle\langle\phi\rvert$ . Why: purity lets us trade one marginal's entropy for another's, so we manufacture a pure state to work with.

Step 2. Apply the dilation on the $A$ leg: $\lvert\psi\rangle_{RBE}=(\mathbb{1}_R\otimes V)\lvert\phi\rangle_{RA}$ ; since $V$ is an isometry, $\lvert\psi\rangle_{RBE}$ is pure on all three parties. Why: a pure global state forces its marginals to pair up, $S(RB)=S(E)$ and $S(B)=S(RE)$ — the equalities the argument runs on.

Step 3. Read off the reduced states: $\mathrm{Tr}_{RE}\lvert\psi\rangle\langle\psi\rvert=\mathcal{N}(\rho)$ and $\mathrm{Tr}_{RB}\lvert\psi\rangle\langle\psi\rvert=\mathcal{N}^c(\rho)$ , so $S(B)=S(\mathcal{N}(\rho))$ and $S(E)=S(\mathcal{N}^c(\rho))$ . Why: this is just the two channels as the two partial traces of the same $V\rho V^\dagger$ , applied inside the purified state.

Step 4. Combine with $S(RB)=S(E)$ : $I_c\equiv S(B)-S(RB)=S(B)-S(E)=S(\mathcal{N}(\rho))-S(\mathcal{N}^c(\rho))$ , and $S(B)-S(RB)=-S(B\mid R)$ on the pure state, so all three forms coincide. Why: $I_c=-S(B\mid R)$ shows positive coherent information means negative conditional entropy — a purely quantum feature — while $S(B)-S(E)$ makes the complementary symmetry manifest. *Note: on any pure input the state $V\lvert\phi\rangle$ is pure on $BE$ , so $S(B)=S(E)$ and $I_c=0$ — a useful sanity check on the sign convention.*

Go deeper · Derivation B · 1 minAntidegradable channels cannot send quantum information

Derivation BAntidegradable channels cannot send quantum information

The goal is to show that antidegradability, $\mathcal{D}'\circ\mathcal{N}^c=\mathcal{N}$ , forces the coherent information to be non-positive on every input and hence the quantum capacity to zero — the cleanest no-cloning argument in quantum Shannon theory. The idea: if the environment can be processed back into Bob's output, any quantum information Bob could reliably recover would exist at both $B$ and $E$ , i.e. it would have been cloned. We make this rigorous through the data-processing inequality for the coherent information plus the antisymmetry of $I_c$ under swapping Bob and the environment.

Step 1. Write the coherent information on the purified state as $I_c(\rho,\mathcal{N})=S(B)-S(E)$ . Why: casting $I_c$ as a difference of branch entropies lets the complement enter symmetrically.

Step 2. The complement has $I_c(\rho,\mathcal{N}^c)=S(E)-S(B)=-I_c(\rho,\mathcal{N})$ , since $\mathcal{N}^c$ exchanges which branch is "output" and which is "environment". Why: this exact antisymmetry is the lever — a bound in one direction becomes a sign-flipped bound for the complement.

Step 3. Apply antidegradability: since $\mathcal{N}=\mathcal{D}'\circ\mathcal{N}^c$ is a further processing of $\mathcal{N}^c$ , monotonicity gives $I_c(\rho,\mathcal{N})\le I_c(\rho,\mathcal{N}^c)$ . Why: data processing forbids manufacturing coherent information by acting on the output, so the extra map $\mathcal{D}'$ can only reduce it.

Step 4. Chain: $I_c(\rho,\mathcal{N})\le -I_c(\rho,\mathcal{N})$ , hence $I_c(\rho,\mathcal{N})\le 0$ for every input. Why: a never-positive single-letter quantity cannot be rescued by the limit.

Step 5. Regularize. Antidegradability of $\mathcal{N}$ implies it for $\mathcal{N}^{\otimes n}$ (use $\mathcal{D}'^{\otimes n}$ ), so every term of the LSD limit is $\le 0$ and $Q=0$ . Why: this promotes the single-letter bound to the true capacity with no additivity needed. *Note: amplitude damping is self-complementary at $\gamma=1/2$ , where $\mathcal{N}=\mathcal{N}^c$ makes it trivially antidegradable with $Q=0$ , consistent with the second figure.*

Go deeper · Derivation C · 1 minDegradability collapses the quantum capacity to a single letter

Derivation CDegradability collapses the quantum capacity to a single letter

The goal is to prove that for degradable channels the coherent information is additive, so the regularized LSD formula (5) collapses to the single-letter maximum $Q(\mathcal{N})=\max_\rho I_c(\rho,\mathcal{N})$ of equation (6) — what makes a whole family (dephasing, erasure, amplitude damping below threshold, pure loss above threshold) exactly solvable. The technical heart is that for two degradable channels in parallel each environment output is a local processing of the corresponding Bob output, arranging the four systems into exactly the pattern strong subadditivity constrains.

Step 1. Dilate both channels: $\lvert\psi\rangle_{R\,B_1E_1\,B_2E_2}=(\mathbb{1}_R\otimes V_1\otimes V_2)\lvert\phi\rangle_{RA_1A_2}$ , so $I_c=S(B_1B_2)-S(E_1E_2)$ . Why: we need the two-copy coherent information in branch-entropy form before bounding it by single-copy values.

Step 2. Use degradability to realize each environment locally, $E_i=\mathcal{D}_i(B_i)$ with $\mathcal{D}_i:B_i\to E_i$ , so $E_i$ comes from $B_i$ alone. Why: locality makes the map $(B_1,B_2)\to(E_1,E_2)$ factorize — the hypothesis strong subadditivity needs.

Step 3. Apply strong subadditivity: $S(B_1B_2)-S(E_1E_2)\le [S(B_1)-S(E_1)]+[S(B_2)-S(E_2)]$ once the $E_i$ are local functions of the $B_i$ . Why: SSA is the one inequality strong enough to control a four-party entropy difference, and degradability is what lets us invoke it.

Step 4. Translate back: $I_c(\rho_{12},\mathcal{N}_1\otimes\mathcal{N}_2)\le I_c(\rho_1,\mathcal{N}_1)+I_c(\rho_2,\mathcal{N}_2)$ , and the reverse holds trivially for product inputs, so $\max I_c$ is exactly additive. Why: additivity both ways pins the two-copy optimum to the sum of one-copy optima, removing the regularization.

Step 5. Iterate to all $n$ : $Q=\max_\rho I_c(\rho,\mathcal{N})$ , and degradability also makes $I_c$ concave in $\rho$ , so this is one well-behaved concave program. Why: concavity guarantees no spurious local maxima, so the closed form is genuinely computable. *Note: the dephasing channel gives $Q=1-H_2(p)$ , additive for every $p$ , consistent with this argument.*

Limiting cases

The value of degradability is a short list of channels with exactly known capacities. The table collects the canonical qubit and bosonic examples and their degradable/antidegradable thresholds.

Case	Result	Meaning
Dephasing, flip prob. $p$	degradable for all $p$ ; $Q=1-H_2(p)$	phase noise never lets Eve out-learn Bob; $Q=0$ only at $p=1/2$
Amplitude damping, damping $\gamma$	degradable iff $\gamma\le\tfrac12$ , antidegradable iff $\gamma\ge\tfrac12$	self-complementary at $\gamma=1/2$ , so $Q=0$ there and above
Erasure, erasure prob. $p$	degradable for $p\le\tfrac12$ ; $Q=\max(0,1-2p)\log_2 d$	complement is erasure with prob. $1-p$ ; antidegradable and $Q=0$ for $p\ge1/2$
Depolarizing, $0<p<\tfrac34$	neither degradable nor antidegradable (generic)	coherent information is superadditive; exact $Q$ still open
Bosonic pure loss, transmissivity $\eta$	degradable for $\eta\ge\tfrac12$ ; $Q=\max\!\big(0,\log_2\tfrac{\eta}{1-\eta}\big)$	beam-splitter complement has transmissivity $1-\eta$ ; hard cutoff at $\eta=1/2$ (3 dB)

Common trapDegradability is a property of the channel, but the complement you test it against must be a legitimate one

Three traps sit here. First, degradability is intrinsic to $\mathcal{N}$ — independent of the Kraus or Stinespring representation — but you must test it against a valid complement, since $\mathcal{N}^c$ is only defined up to an isometry on $E$ ; wrong representatives give a wrong verdict. Second, $\mathcal{D}\circ\mathcal{N}=\mathcal{N}^c$ must hold exactly as CPTP maps on every input, not on one convenient state — state-specific matching is the weaker notion of approximate degradability, which yields bounds, not equalities. Third, additivity of $I_c$ for degradable channels does not license additivity of the private or classical capacity, and a generic channel is neither degradable nor antidegradable, so neither closed-form argument applies — the honest default is "capacity unknown".

Worked examples

1 · Quantum capacity of the dephasing channel (fully worked)

The qubit dephasing channel $\mathcal{N}_p(\rho)=(1-p)\rho+p\,Z\rho Z$ applies a phase flip with probability $p$ . Find its complementary channel, confirm degradability, and compute the quantum capacity, evaluating it at $p=0.1$ .

Step 1. Kraus and isometry. Take $K_0=\sqrt{1-p}\,\mathbb{1}$ , $K_1=\sqrt{p}\,Z$ , so $V=\sqrt{1-p}\,\mathbb{1}\otimes\lvert0\rangle_E+\sqrt{p}\,Z\otimes\lvert1\rangle_E$ ; since $(1-p)\mathbb{1}+pZ^2=\mathbb{1}$ , $V$ is a valid isometry with a two-dimensional environment.

Step 2. The complementary channel. Its matrix elements $(\mathcal{N}^c(\rho))_{kl}=\mathrm{Tr}[K_k\rho K_l^\dagger]$ have fixed diagonal $\{1-p,\,p\}$ and off-diagonal $\sqrt{p(1-p)}\,\mathrm{Tr}[\rho Z]$ : the environment sees a state with fixed populations whose coherence tracks only $\langle Z\rangle$ — a classical-like readout of the input's diagonal.

Step 3. Degradability. A fixed dephasing-type map $\mathcal{D}$ carries Bob's output to Eve's for every $p$ , so $\mathcal{N}_p$ is degradable for all $p$ and equation (6) applies: $Q=\max_\rho I_c(\rho,\mathcal{N}_p)$ .

Step 4. Maximize. By symmetry the optimum is $\rho=\mathbb{1}/2$ : Bob's output is still $\mathbb{1}/2$ so $S(\mathcal{N}_p(\rho))=1$ bit, while the environment output has spectrum $\{1-p,p\}$ so $S(\mathcal{N}_p^c(\rho))=H_2(p)$ with $H_2(p)=-p\log_2 p-(1-p)\log_2(1-p)$ . Therefore

Q(\mathcal{N}_p)=1-H_2(p).

Step 5. Numbers and sanity. At $p=0$ or $p=1$ the map is unitary and $Q=1$ ; at $p=1/2$ , $H_2(1/2)=1$ so $Q=0$ . At $p=0.1$ , $H_2(0.1)=0.4690$ , giving $Q(\mathcal{N}_{0.1})=0.531$ qubits per use, and by degradability the private capacity equals it, $P=Q=0.531$ .

2 · Amplitude damping: the degradable-to-dead transition

The amplitude-damping channel with damping $\gamma$ has $K_0=\lvert0\rangle\langle0\rvert+\sqrt{1-\gamma}\,\lvert1\rangle\langle1\rvert$ , $K_1=\sqrt{\gamma}\,\lvert0\rangle\langle1\rvert$ , and coherent information $I_c(p)=H_2((1-\gamma)p)-H_2(\gamma p)$ at input population $p=\langle1\rvert\rho\lvert1\rangle$ . Locate the degradable region and evaluate $Q$ at $\gamma=0.2$ and $\gamma=0.5$ .

Step 1. Degradable region. The complement of amplitude damping $\gamma$ is amplitude damping with damping $1-\gamma$ , so $\mathcal{N}_\gamma$ is degradable when $\gamma\le 1/2$ and antidegradable for $\gamma\ge1/2$ ; only in the degradable region does $Q=\max_p I_c(p)$ hold.

Step 2. Evaluate at $\gamma=0.2$ (degradable). Maximizing $I_c(p)=H_2(0.8p)-H_2(0.2p)$ over $p\in[0,1]$ gives the optimum near $p^\star=0.449$ , where $Q(0.2)=0.506$ qubits per use — the amber point on the second figure (the symmetric input $p=1/2$ already gives $0.502$ ).

Step 3. Evaluate at $\gamma=0.5$ (threshold). Here $\gamma=1-\gamma$ , so the channel equals its own complement: $I_c(p)=H_2(0.5p)-H_2(0.5p)=0$ identically, so $Q(0.5)=0$ , pinned at zero for all $\gamma\ge1/2$ .

Step 4. The lesson. As $\gamma$ climbs from $0$ to $1/2$ the capacity falls smoothly from $1$ to $0$ ( $1,0.709,0.506,0.328,0.162,0$ at $\gamma=0,0.1,0.2,0.3,0.4,0.5$ ), and the upper half of the range is dead — the qubit sibling of the pure-loss cutoff at $\eta=1/2$ .

Physics of the portrait

The complementary channel is named not after a person but after an idea: a channel and its leakage are two projections of one reversible isometry — a lineage running from a 1955 theorem about positive maps to the modern theory of quantum capacity.

Origins · 1 mincomplementary channel and how this was discovered

complementary channel

the environment's channel obtained by tracing out Bob rather than the environment · quantum Shannon theory of degradable channels

The complementary channel is the mathematical shadow every quantum channel casts onto its environment, and its history is the history of learning to take dilations seriously. W. Forrest Stinespring proved in 1955 that every completely positive map extends to an isometry into a larger space — a result that decades later became the physicist's statement that noise is unitary interaction with an unwatched environment. Once channels were routinely dilated, the complementary (or conjugate) channel emerged as the object encoding what that environment records, developed in the 1970s by Lindblad and Holevo. The information-theoretic payoff waited for the coherent information of Schumacher and Nielsen (1996) and the Lloyd–Shor–Devetak theorem (1997–2005) that made it the quantum capacity. The decisive structural idea — degradability — is due to Devetak and Shor in 2005: by asking whether Bob can degrade his output into the environment's, they isolated exactly the channels for which the capacity is single-letter, and proved private and quantum capacities then coincide. Giovannetti and Fazio pinned the amplitude-damping threshold to $\gamma=1/2$ that year, and Cubitt, Ruskai, and Smith mapped the degradable set in 2008 — turning a theorem about abstract positive maps into the daily triage tool of quantum communication.

History

1955

— Stinespring proves every completely positive map dilates to an isometry

1996

— Schumacher and Nielsen define the coherent information

2005

— Devetak and Shor introduce degradable channels and prove single-letter additivity

2005

— Giovannetti and Fazio fix the amplitude-damping degradability threshold at γ = 1/2

2008

— Cubitt, Ruskai, and Smith map the structure of the degradable set

Going further

Texts. Wilde, Quantum Information Theory, 2nd ed., ch. 13 and 24 · Watrous, The Theory of Quantum Information, ch. 2 and 8 · Nielsen and Chuang, Quantum Computation and Quantum Information, ch. 8 and 12 · Preskill, Lecture Notes on Quantum Computation, ch. 10. Key results to keep in mind: $\mathcal{N}^c=\mathrm{Tr}_B[V\rho V^\dagger]$ is unique up to an environment isometry; $I_c=S(B)-S(E)=-S(B\mid R)$ on the purified state; the antidegradable $\Rightarrow Q=0$ chain and degradable single-letter additivity both follow from strong subadditivity; $Q(\mathcal{N}_p)=1-H_2(p)$ with $H_2(0.1)=0.4690\Rightarrow Q=0.531$ and $Q(1/2)=0$ ; amplitude-damping values $Q(0.2)=0.506$ , $Q(0.5)=0$ . The thresholds $\gamma=1/2$ , $p=1/2$ (erasure), $\eta=1/2$ (pure loss) and the structure of the degradable set are standard.

Prerequisites. Stinespring Dilation Theorem · Quantum Channels and CPTP Maps · Kraus Representation of Quantum Channels · Von Neumann Entropy · Data-Processing Inequality. Leads to. Quantum Joint, Conditional Entropy and Mutual Information · Strong Subadditivity of Quantum Entropy · Holevo Bound.

Ask the tutor: Name any channel — dephasing, amplitude damping, or your own Kraus set — and we will build its complement, point the degrading arrow, and read off whether its quantum capacity is a one-line formula or an open problem.

Why is the complementary channel well defined if the Stinespring dilation is not unique?

Two dilations of the same channel differ only by an isometry $W$ on the environment, $V'=(\mathbb{1}_B\otimes W)V$ . Tracing out Bob then gives $\mathcal{N}^{c\prime}(\rho)=W\mathcal{N}^c(\rho)W^\dagger$ — the candidate complements are related by a fixed partial isometry on $E$ . Every isometry-invariant quantity is therefore unambiguous, including the one thing we actually use, the environment's output entropy $S(\mathcal{N}^c(\rho))$ , since entropy is unitarily invariant. Degradability is invariant too, because composing with $W$ can be absorbed into the degrading map. So although the complement is defined only up to an environment isometry, everything the theory computes with it is fixed.

What is the difference between the complementary channel and simply the noise?

The complement is not "the error" — it is a bona fide channel mapping the same input $\rho$ to a legitimate output, the reduced state of the environment. Where $\mathcal{N}$ answers "what does Bob hold?", $\mathcal{N}^c$ answers "what does the rest of the universe hold?"; both come from the single isometry $V$ by tracing out the other system. It matters because of information conservation: the isometry loses nothing, so any information missing from Bob went to $E$ , and $\mathcal{N}^c$ says exactly what and how much — which is why the eavesdropper in a security proof is modelled by the complementary channel.

How can I quickly tell whether a channel is degradable?

Write down a Stinespring isometry and read off $\mathcal{N}^c$ , then ask whether a CPTP map $\mathcal{D}$ satisfies $\mathcal{D}\circ\mathcal{N}=\mathcal{N}^c$ ; for low-dimensional channels this is a small linear-algebra check, and for named channels the complement is another member of the same family, making it transparent. A shortcut: if $\mathcal{N}$ and $\mathcal{N}^c$ swap a parameter across $1/2$ (damping $\gamma\leftrightarrow1-\gamma$ ), degradability holds on the less noisy side. When neither $\mathcal{D}$ nor $\mathcal{D}'$ exists exactly, the channel is generic and you fall back on approximate degradability for bounds.

Why does antidegradability force the quantum capacity to be exactly zero?

Because it would otherwise let you clone. If $\mathcal{N}$ were antidegradable and had $Q>0$ , Bob could reliably recover transmitted states, while the environment — which by antidegradability can be processed into a copy of Bob's output — could independently recover the same states, giving two faithful copies of an unknown state in violation of no-cloning. The formal version shows the coherent information is non-positive on every input, and a never-positive single-letter quantity regularizes to zero, so $Q=0$ exactly — which is why amplitude damping above $\gamma=1/2$ , erasure above $p=1/2$ , and pure loss below $\eta=1/2$ transmit no quantum information at all.

Does additivity of the coherent information also make the private and classical capacities single-letter?

No, and conflating these is a common error. Degradability makes the coherent information additive, single-letterizing $Q$ , and for degradable channels the private capacity $P$ coincides with $Q$ so it too is single-letter. But the classical capacity is governed by the Holevo quantity, a different functional that can be superadditive even when the coherent information is additive. And the additivity argument needs degrading maps for both factors of a tensor product — one degradable factor is not enough. Keep the three capacities separate: degradability is a statement about $Q$ (and, as a bonus, $P=Q$ ), not a universal additivity theorem.

If almost no channel is exactly degradable, why does the concept matter?

Two reasons. First, the degradable channels model much of real hardware: dephasing is pure phase noise, amplitude damping is energy relaxation ( $T_1$ decay), erasure is heralded photon loss, and pure loss is optical fibre — so the solvable set is the backbone of practical analysis, not a curiosity. Second, degradability is the reference point for everything else: approximate degradability measures how far a channel sits from the degradable set and turns that distance into rigorous capacity bounds by SDP — currently the best handle on hard cases like the depolarizing channel. It matters not because every channel is degradable, but because degradability is the axis along which all channels are measured.