Additivity, Superadditivity and Capacity Regularization

Shannon left classical information theory with a gift so clean it is easy to take for granted: the capacity of a noisy channel is a single number you compute from one use of it, the maximum mutual information, and stacking channels in parallel buys you nothing you could not get by using each one independently. Capacity is additive, and additive quantities are computable. Quantum channels shatter this comfort. The two information measures that govern them — the Holevo quantity $\chi$ for sending classical bits down a quantum channel, and the coherent information $I_c$ for sending qubits — are not additive under the tensor product. Feed an entangled state across several uses of a channel and you can sometimes transmit strictly more per use than any product strategy allows. The honest definition of capacity is then not a one-shot maximum but a regularized limit $C=\lim_{n\to\infty}\frac1n\,C^{(1)}(\mathcal{N}^{\otimes n})$ over ever-larger blocks, an optimization whose dimension explodes exponentially and which, for most channels, no one knows how to compute. This article is about the gap between $C^{(1)}$ and $C$ : when it closes, when it stays open, and how badly it can blow up. Three landmarks organize the story. Hastings proved in 2009 that the classical Holevo capacity is generically superadditive, killing a decade-old conjecture. Coherent information is superadditive too, and worse: a channel can have zero one-shot quantum capacity yet positive true capacity. Most startling, Smith and Yard showed superactivation — two channels each with $Q=0$ whose tensor product has $Q>0$ , with no classical shadow whatsoever. Against this, degradable channels are the rare well-behaved class where coherent information is additive and $Q=Q^{(1)}$ single-letterizes.

Key takeaways

A capacity is additive when running channels in parallel cannot beat the sum of their individual capacities; when entangled inputs across uses do beat it, the capacity is superadditive and must be defined as a regularized limit $\lim_{n}\frac1n C(\mathcal{N}^{\otimes n})$ .
The classical (Holevo) capacity satisfies $C(\mathcal{N})=\lim_n\frac1n\chi(\mathcal{N}^{\otimes n})$ ; Hastings (2009) proved $\chi$ is generically superadditive, so the single-use $\chi(\mathcal{N})$ is only a lower bound on the true rate.
The quantum capacity $Q(\mathcal{N})=\lim_n\frac1n Q^{(1)}(\mathcal{N}^{\otimes n})$ is dramatically superadditive: coherent information can be zero for one use yet positive for many.
Superactivation — two channels with $Q(\mathcal{N}_1)=Q(\mathcal{N}_2)=0$ but $Q(\mathcal{N}_1\otimes\mathcal{N}_2)>0$ — has no classical analogue and shows capacity is a property of channel collections, not single channels.
Degradable channels are the tractable class: $I_c$ is concave and additive, so $Q=Q^{(1)}$ single-letterizes and is computable — the qubit erasure channel gives the exact $Q=\max(0,(1-2p)\log d)$ , vanishing at the symmetric point $p=½$ .
Key idea: a capacity equals its single-letter formula $C^{(1)}$ exactly when the underlying entropic functional is additive; entanglement across channel uses can break that additivity, forcing regularization and leaving most quantum capacities not known to be computable.

The core insight

A capacity is operationally a regularized limit $C=\lim_n\frac1n\,C^{(1)}(\mathcal{N}^{\otimes n})$ ; it collapses to the single-letter quantity $C^{(1)}$ exactly when the underlying entropic functional is additive. Entanglement spread across channel uses can break additivity, making most quantum capacities superadditive, non-single-letter, and not known to be computable.

Where this lives today

The additivity question sets the working frontier of quantum Shannon theory. Because $C$ and $Q$ are regularized limits whose optimizations grow exponentially in block length, no algorithm computes them for a general channel, and this drives an active hunt for single-letter upper bounds that are additive even when the capacity is not — the Rains information, the $D_{\max}$ -based "approximate degradability" bounds of Sutter–Scholz–Winter–Renner, and the max-Rains information. These sandwich the true capacity between the achievable coherent-information rate below and a computable additive bound above, and for many channels of practical interest (thermal-attenuator bosonic channels in fibre and free-space links, dephasing memories) the sandwich is now tight enough to quote capacities to engineering precision. Superactivation, meanwhile, has become a network principle: a link useless on its own can be essential in combination, so capacity must be reasoned about at the level of the whole graph of channels — now baked into quantum-repeater and quantum-internet analyses.

The physical picture

Picture a capacity as a bookkeeping question: how much information survives $n$ uses of a channel, divided by $n$ , as $n\to\infty$ . Classically the answer is boring in the best way — the optimal strategy uses each copy independently and identically, so the per-use information is just the single-use value repeated. Additivity is the statement that i.i.d. is optimal. Quantum mechanics hands you a resource classical coders never had: you may feed a single entangled state across the inputs of several channel copies at once. Sometimes that entangled input produces a joint output with lower total entropy — more order at the receiver — than any product input can, and the information per use rises above the single-use optimum. That is superadditivity, and the first figure contrasts the two strategies and names the excess.

Figure

Two ways to drive two uses of a channel. The product strategy feeds independent inputs and achieves rate 2χ(N). The entangled strategy feeds one correlated input across both uses and can achieve χ(N⊗²) which Hastings proved is sometimes strictly larger — the shaded gap is the superadditive excess with no classical analogue

The coherent information $I_c$ is the quantum-capacity analogue of this picture. It equals the entropy Bob receives minus the entropy that leaks to the environment through the complementary channel $\mathcal{N}^c$ . Because both maps act on the same correlated multi-use input, an entangled code can route the message into the channel while diverting decoherence into environment correlations Eve cannot exploit. When $I_c$ is zero for a single use but positive across many, the channel quietly accumulates capacity no one-shot measurement reveals. The second figure plots this for two workhorse channels and marks the noise levels where their one-shot quantum capacity dies.

Figure

One-shot coherent-information capacity Q⁽¹⁾ versus noise. The erasure channel (blue) falls linearly as 1−2p and hits zero at the symmetric point p = 0.5. The depolarizing channel (amber) follows 1 − H₂(p) − p·log₂3 and hits zero near p = 0.189, below which entangled multi-use codes can still push the true Q above zero

The core equations

The whole subject hinges on two entropic functionals and what the tensor product does to them. We define each single-use quantity, state the regularized capacity it produces, and record when the two coincide. Everything below is bits or qubits per channel use, with $\log\equiv\log_2$ . The classical capacity is governed by the Holevo quantity, an optimization over input ensembles measuring how much classical information a receiver can extract from the channel's output states.

\chi(\mathcal{N})=\max_{\{p_i,\rho_i\}}\Big[S\Big(\mathcal{N}\big(\textstyle\sum_i p_i\rho_i\big)\Big)-\sum_i p_i\,S\big(\mathcal{N}(\rho_i)\big)\Big] \qquad (1)

where: $\{p_i,\rho_i\}$ — an input ensemble of states $\rho_i$ sent with probabilities $p_i$ ; $\mathcal{N}$ — the quantum channel, a CPTP map; $S(\cdot)$ — von Neumann entropy of the output; $\chi$ — the Holevo information, equal to the one-shot product-state classical capacity. It is the average output entropy minus the average of the individual output entropies — the "spread" the receiver can resolve. By the Holevo–Schumacher–Westmoreland (HSW) theorem $\chi(\mathcal{N})$ is achievable with product-state inputs, but entangled inputs across a block can raise it, so the true classical capacity is the regularized limit.

C(\mathcal{N})=\lim_{n\to\infty}\frac1n\,\chi\big(\mathcal{N}^{\otimes n}\big)=\sup_n\frac1n\,\chi\big(\mathcal{N}^{\otimes n}\big) \qquad (2)

where: $\mathcal{N}^{\otimes n}$ — $n$ parallel uses treated as one channel; $C(\mathcal{N})$ — the true classical capacity, bits per use. Limit equals supremum because $\chi(\mathcal{N}^{\otimes n})$ is superadditive in $n$ (Derivation A). The quantum capacity is governed instead by the coherent information, the difference between what reaches Bob and what leaks to the environment.

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

where: $\rho$ — the input state; $\mathcal{N}^c$ — the complementary channel to the environment, read off any Stinespring dilation; $S(\mathcal{N}^c(\rho))$ — entropy leaked to Eve. When the environment learns less than Bob, $I_c>0$ . The Lloyd–Shor–Devetak (LSD) theorem makes the maximized coherent information achievable, and — because $I_c$ is superadditive — the true quantum capacity is again a regularized limit.

Q(\mathcal{N})=\lim_{n\to\infty}\frac1n\,Q^{(1)}\big(\mathcal{N}^{\otimes n}\big),\qquad Q^{(1)}(\mathcal{N})=\max_{\rho}\,I_c(\rho,\mathcal{N}) \qquad (4)

where: $Q^{(1)}$ — the single-letter coherent-information capacity, only a lower bound in general; $Q$ — the true quantum capacity, qubits per use. Unlike $\chi$ , $Q^{(1)}$ is not even concave, so it can vanish while $Q>0$ . For the special class of degradable channels the regularization collapses and the capacity single-letterizes exactly.

Q(\mathcal{N})=Q^{(1)}(\mathcal{N})=\max_{\rho}\,I_c(\rho,\mathcal{N})\qquad(\mathcal{N}\ \text{degradable}) \qquad (5)

where: degradable means there exists a CPTP map $\mathcal{D}$ with $\mathcal{N}^c=\mathcal{D}\circ\mathcal{N}$ — Bob can simulate Eve. Then $I_c$ is concave in $\rho$ and additive over tensor products, so a product input is optimal at every block length. One capacity is additive for every channel: the entanglement-assisted classical capacity, the cleanest object in the theory.

C_E(\mathcal{N})=\max_{\phi_{AA'}}\,I(A;B)_{\omega},\qquad \omega_{AB}=(\mathrm{id}_A\otimes\mathcal{N})(\phi_{AA'}) \qquad (6)

where: $\phi_{AA'}$ — a pure input with reference system $A$ ; $I(A;B)$ — the quantum mutual information of the output; $C_E$ — entanglement-assisted classical capacity. Because $I(A;B)$ is additive, $C_E=C_E^{(1)}$ needs no regularization. Finally, the additivity question is sharpest in the minimum output entropy, which Shor proved equivalent to Holevo additivity.

S_{\min}(\mathcal{N})=\min_{\rho}\,S\big(\mathcal{N}(\rho)\big),\qquad S_{\min}(\mathcal{N}\otimes\mathcal{M})\overset{?}{=}S_{\min}(\mathcal{N})+S_{\min}(\mathcal{M}) \qquad (7)

where: $S_{\min}$ — the least output entropy over all pure inputs. Additivity of $S_{\min}$ is equivalent to additivity of $\chi$ and of the entanglement of formation $E_F$ ; Hastings showed all three fail.

Derivation AWhy capacity is a regularized limit and never a smaller one

The goal is to show why the operational capacity is $C=\lim_n\frac1n\chi(\mathcal{N}^{\otimes n})$ rather than the tempting one-shot value $\chi(\mathcal{N})$ , and to prove that this limit exists, equals a supremum, and can only lie at or above the single-use rate. The argument rests on a super-additivity inequality that always holds plus Fekete's lemma, which turns super-additivity of a sequence into convergence of its normalized version.

Step 1. HSW achievability fixes the floor: with product codewords over a length- $n$ block, the rate $\chi(\mathcal{N})$ is achievable with vanishing error and no product scheme does better, so $C(\mathcal{N})\ge\chi(\mathcal{N})$ . Why: this identifies $\chi$ as an honest achievable rate, so any larger regularized value is a genuine improvement, not an artifact.

Step 2. Product strategies are a special case of joint ones: restricting $\mathcal{N}^{\otimes n}$ to product inputs reproduces $n\,\chi(\mathcal{N})$ , so entangled inputs can only help — $\chi(\mathcal{N}^{\otimes(m+n)})\ge\chi(\mathcal{N}^{\otimes m})+\chi(\mathcal{N}^{\otimes n})$ . Why: this is the super-additivity of $a_n=\chi(\mathcal{N}^{\otimes n})$ — free, requiring no hard theorem.

Step 3. Fekete's lemma says that for a super-additive sequence, $\lim_n a_n/n$ exists and equals $\sup_n a_n/n$ . Applying it gives $C=\lim_n\frac1n\chi(\mathcal{N}^{\otimes n})=\sup_n\frac1n\chi(\mathcal{N}^{\otimes n})\ge\chi(\mathcal{N})$ . Why: this is what guarantees the capacity is well-defined at all, and pins it at or above the single-use rate — the gap, if any, is the price of non-additivity.

Step 4. Whether the gap is zero is exactly the additivity question. If $\chi(\mathcal{N}^{\otimes n})=n\chi(\mathcal{N})$ for all $n$ , then $C=\chi(\mathcal{N})$ and the channel is single-letter; Hastings produced channels where $\chi(\mathcal{N}^{\otimes2})>2\chi(\mathcal{N})$ , so the strict gap is real. Why: it converts a structural dichotomy into the boundary between computable and (so far) uncomputable capacities.

Go deeper · Derivation B · 1 minThe erasure channel and why degradability makes it single-letter

Derivation BThe erasure channel and why degradability makes it single-letter

The goal is to compute the quantum capacity of the qubit erasure channel exactly, obtaining $Q=\max(0,(1-2p)\log d)$ , and to see how degradability for $p<½$ forces the regularization to collapse. Erasure is the ideal teaching case: its complement is another erasure channel, the Bob–Eve symmetry is transparent, and the coherent information optimizes in one line.

Step 1. The channel replaces its input with an orthogonal flag $|e\rangle$ with probability $p$ : $\mathcal{N}_{\mathrm{er}}(\rho)=(1-p)\rho+p\,|e\rangle\langle e|$ . Because the receiver sees the flag, erasures are heralded. The environment gets the input exactly when Bob does not, so $\mathcal{N}^c$ is an erasure channel with parameter $1-p$ . Why: identifying the complement as the "swapped" erasure channel exposes the Bob–Eve symmetry that controls the capacity.

Step 2. For $p<½$ Bob is the better-informed party, and one can post-process Bob's output to reproduce Eve's, so $\mathcal{N}^c=\mathcal{D}\circ\mathcal{N}$ — the channel is degradable. For $p>½$ the roles flip: the channel is antidegradable, Eve holds a copy at least as good as Bob's, so $Q=0$ . Why: degradability lets us stop at $Q^{(1)}$ ; antidegradability is a clean no-go that kills capacity outright.

Step 3. The coherent information is $I_c(\rho,\mathcal{N}_{\mathrm{er}})=(1-p)S(\rho)-p\,S(\rho)=(1-2p)S(\rho)$ , since output and leaked entropy are both proportional to $S(\rho)$ in an all-or-nothing channel. The maximally mixed input gives $Q^{(1)}=(1-2p)\log d$ . Why: the linearity in $S(\rho)$ is special to erasure and makes the optimization trivial and the answer exact.

Step 4. Degradability makes $I_c$ concave and additive, so $\frac1n Q^{(1)}(\mathcal{N}^{\otimes n})=Q^{(1)}$ for all $n$ and $Q=Q^{(1)}=\max(0,(1-2p)\log d)$ . For a qubit, $Q=\max(0,1-2p)$ , vanishing at $p=½$ . Why: degradability turns an infinite-dimensional regularization into a one-line formula, matching Bennett–DiVincenzo–Smolin.

Go deeper · Derivation C · 1 minSuperactivation — two dead channels that jointly carry qubits

Derivation CSuperactivation — two dead channels that jointly carry qubits

The goal is to explain how two channels each with strictly zero quantum capacity combine into a product with positive quantum capacity — the Smith–Yard phenomenon — and why it is impossible classically. The construction pairs a symmetric channel that supplies "assistance" with a private channel that hides a secret but cannot transmit qubits alone; each supplies what the other lacks.

Step 1. Let $\mathcal{N}_1$ be a $50\%$ erasure channel. It is symmetric and antidegradable, so $Q(\mathcal{N}_1)=0$ : Eve's copy is as good as Bob's and nothing can be sent privately. Yet a symmetric channel can still assist, sharing correlations without leaking a net advantage to the environment. Why: a $Q=0$ channel is not inert — symmetric side-channels carry a resource coherent information alone does not see.

Step 2. Let $\mathcal{N}_2$ be a PPT (positive-partial-transpose) channel built from bound entanglement. It has $Q(\mathcal{N}_2)=0$ — PPT channels cannot distill entanglement — yet a strictly positive private capacity $P(\mathcal{N}_2)>0$ : it can hide a shared secret key even though it cannot transmit coherent quantum information. Why: the gap between $P>0$ and $Q=0$ is the reservoir that the assisting channel will unlock.

Step 3. Smith and Yard prove $Q(\mathcal{N}_1\otimes\mathcal{N}_2)\ge\tfrac12\,P(\mathcal{N}_2)$ : the symmetric channel $\mathcal{N}_1$ lets the private capacity of $\mathcal{N}_2$ be converted into genuine quantum capacity of the pair. Why: this inequality links a classical-like resource $(P)$ on one factor to a quantum resource $(Q)$ on the product.

Step 4. Since $P(\mathcal{N}_2)>0$ , the joint capacity is strictly positive although both factors have $Q=0$ . Classically this cannot happen. Why: $Q$ is not merely non-additive but radically non-single-letter — capacity is a property of the whole collection of channels.

Limiting cases

Where additivity holds, and where it spectacularly fails:

Case	Result	Meaning
Classical Shannon channel	$C=\max_{p(x)}I(X;Y)$	Fully additive; the single-letter mutual-information formula is exact, regularization unnecessary.
Entanglement-assisted capacity	$C_E=\max_{\phi}I(A;B)$	Quantum mutual information is additive for every channel; $C_E$ is single-letter — the cleanest quantum capacity.
Degradable channel (quantum)	$Q=\max_{\rho}I_c(\rho,\mathcal{N})$	$I_c$ concave and additive; $Q=Q^{(1)}$ , computable. Covers dephasing, erasure $p<½$ , damping $\gamma<½$ .
Generic channel (classical)	$C=\lim_n\frac1n\chi(\mathcal{N}^{\otimes n})$	Holevo information superadditive (Hastings 2009); only the regularized limit is operationally correct.
Superactivation (quantum)	$Q(\mathcal{N}_1)=Q(\mathcal{N}_2)=0,\ Q(\mathcal{N}_1\otimes\mathcal{N}_2)>0$	Smith–Yard 2008: zero-capacity channels combine to a positive capacity — uniquely quantum.

Common trapAssuming quantum capacities are single-letter because the classical one is

The most seductive error is to carry Shannon's intuition across the quantum line and write $C=\chi(\mathcal{N})$ or $Q=Q^{(1)}(\mathcal{N})$ as if the single-use formula were the answer. It is only a lower bound. For non-degradable channels the gap can be strict — Hastings for $\chi$ , DiVincenzo–Shor–Smolin for $I_c$ — and the one-shot quantum value can even be zero while the true capacity is positive, so quoting $Q^{(1)}=0$ as "useless for qubits" is wrong without a degradability argument. A second, subtler trap: superadditivity does not violate the data-processing inequality. DPI holds for every fixed state and every single channel use; superadditivity is about optimizing over entangled inputs spread across uses, which beats the product-input optimum without any information ever increasing under processing.

Worked examples

1 · Erasure channel across the whole noise range (fully worked)

A qubit ( $d=2$ ) is sent down an erasure channel. Find its quantum capacity at $p=0.25$ , at the symmetric point $p=0.5$ , and at $p=0.6$ , and say which regime each is in.

Step 1. Use the exact single-letter result from Derivation B, valid because the channel is degradable for $p<½$ : $Q=\max(0,(1-2p)\log_2 d)=\max(0,1-2p)$ for a qubit.

Step 2. At $p=0.25$ : $Q=1-2(0.25)=\mathbf{0.5}$ qubits per use — degradable, so this is the true capacity, not just a bound; no entangled multi-use code can beat it.

Step 3. At $p=0.5$ : $Q=1-2(0.5)=\mathbf{0}$ . The symmetric point where Bob and Eve are on equal footing — the boundary between degradability and antidegradability, and quantum information dies precisely here.

Step 4. At $p=0.6$ the formula would give $-0.2$ , but the channel is now antidegradable, so $Q=\max(0,-0.2)=\mathbf{0}$ : Eve holds a copy at least as good as Bob's. Summary: $Q=0.5,\ 0,\ 0$ qubits at $p=0.25,\ 0.5,\ 0.6$ — linear until it clamps at zero at the symmetric point.

2 · Depolarizing channel: the one-shot threshold (fully worked)

The qubit depolarizing channel $\mathcal{N}(\rho)=(1-p)\rho+\frac p3(X\rho X+Y\rho Y+Z\rho Z)$ has one-shot coherent-information capacity $Q^{(1)}=1-H_2(p)-p\log_2 3$ at the maximally mixed input. Evaluate it at $p=0.10$ and find where it vanishes.

Step 1. Compute the binary entropy at $p=0.10$ : $H_2(0.10)=-0.10\log_2 0.10-0.90\log_2 0.90=0.3322+0.1368=\mathbf{0.4690}$ bits.

Step 2. The Pauli-spread term is $p\log_2 3=0.10\times1.58496=\mathbf{0.15850}$ bits — the entropy the three Pauli errors dump into the environment.

Step 3. Combine: $Q^{(1)}=1-0.4690-0.1585=\mathbf{0.3725}$ qubits per use — positive, so a hashing-bound code transmits at least $0.37$ qubits per use, and this is only a lower bound on the true $Q$ .

Step 4. Find the threshold by solving $1=H_2(p)+p\log_2 3$ numerically: it gives $p\approx\mathbf{0.1893}$ . Above $p\approx0.189$ the one-shot capacity is zero, yet DiVincenzo–Shor–Smolin showed entangled five-use codes make the true $Q$ positive slightly beyond this point — the cleanest concrete signature of superadditivity.

Physics of the portrait

The central figure here is not a person but capacity itself: a number everyone wants, defined by a limit almost no one can evaluate, assembled over sixty years from Shannon's clean classical original through Holevo's bound, the HSW and LSD coding theorems, and the additivity catastrophes of the 2000s.

Origins · 1 minquantum channel capacity and how this was discovered

quantum channel capacity

the regularized limit C = limₙ (1/n) C⁽¹⁾(N⊗ⁿ) · quantum Shannon theory of noisy channels

Channel capacity was born fully formed in Claude Shannon's 1948 theory: for a classical channel it is the single-letter maximum of the mutual information, additive, so the whole edifice of coding rests on a computable number. The quantum generalization took half a century and refused to be tidy. Alexander Holevo's 1973 bound capped how much classical information a quantum ensemble can carry, and in 1996–1997 Holevo, Schumacher, and Westmoreland turned it into the HSW coding theorem, defining the product-state capacity $\chi$ . The qubit-carrying quantum capacity emerged from Schumacher and Nielsen's coherent information and was proved to equal its regularization by Lloyd, Shor, and Devetak, with Devetak's 2005 paper the rigorous capstone. Then came the surprises. In 1998 DiVincenzo, Shor, and Smolin found coherent information superadditive; in 2004 Peter Shor proved four additivity conjectures — minimum output entropy, Holevo capacity, entanglement of formation, and a strong version — stand or fall together; and in 2009 Matthew Hastings demolished all of them with a random-channel counterexample. Strangest of all, Graeme Smith and Jon Yard exhibited superactivation in 2008: two channels each incapable of sending a single qubit that together can. Capacity, it turned out, is not a property of a channel but of the entangled resources one may spread across its uses — a number defined by a limit that remains, for most channels, uncomputed.

History

1973

— Holevo proves his bound on accessible classical information in a quantum ensemble

1997

— Holevo, Schumacher and Westmoreland establish the HSW product-state classical capacity

1998

— DiVincenzo, Shor and Smolin show coherent information is superadditive for the depolarizing channel

2008

— Smith and Yard demonstrate superactivation of the quantum capacity

2009

— Hastings refutes the additivity conjectures with a random-channel counterexample

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, Ch. 12 · Wilde, Quantum Information Theory, 2nd ed., Ch. 13–24 (capacities, coherent information, additivity) · Watrous, The Theory of Quantum Information, Ch. 8 · Preskill, Lecture Notes on Quantum Computation, Ch. 10 · Hayashi, Quantum Information Theory: Mathematical Foundation, Ch. 9.

Key papers. M. B. Hastings, Nature Physics 5, 255 (2009), superadditivity of communication capacity · G. Smith & J. Yard, Science 321, 1812 (2008), quantum channel superactivation · P. W. Shor, Comm. Math. Phys. 246, 453 (2004), equivalence of additivity conjectures · I. Devetak, IEEE Trans. Inf. Theory 51, 44 (2005), quantum capacity as regularized coherent information · D. DiVincenzo, P. Shor & J. Smolin, Phys. Rev. A 57, 830 (1998), superadditive coherent information.

Provenance. A verifier should reproduce: (1) the erasure capacity $Q=\max(0,(1-2p)\log d)$ from $I_c=(1-2p)S(\rho)$ and degradability for $p<½$ (Bennett–DiVincenzo–Smolin; Wilde Ch. 24); (2) the depolarizing one-shot value $Q^{(1)}=1-H_2(p)-p\log_2 3$ , with $H_2(0.10)=0.4690$ , $0.10\log_2 3=0.15850$ , $Q^{(1)}(0.10)=0.3725$ , and threshold $p\approx0.1893$ ; (3) the Fekete/super-additivity argument giving $C=\sup_n\frac1n\chi(\mathcal{N}^{\otimes n})$ ; (4) Shor's equivalence of the four additivity conjectures and Hastings's refutation; (5) the Smith–Yard bound $Q(\mathcal{N}_1\otimes\mathcal{N}_2)\ge\tfrac12 P(\mathcal{N}_2)$ ; (6) additivity of $C_E$ (Bennett–Shor–Smolin–Thapliyal). All quantities are textbook-standard, with $\log\equiv\log_2$ throughout; no nonstandard constants are introduced.

Prerequisites. Von Neumann Entropy · Quantum Channels and CPTP Maps · Holevo Bound · Complementary Channels and Degradability · Strong Subadditivity of Quantum Entropy. Leads to. Complementary Channels and Degradability · Data-Processing Inequality · Quantum Relative Entropy.

Ask the tutor: Ask me to walk through exactly why degradability makes coherent information additive, or to show the Smith–Yard superactivation construction step by step.

What is the difference between additive and superadditive, in one sentence each?

A capacity is additive when the best you can do with two channels in parallel is exactly the sum of what each does alone, so the single-use formula is the whole answer. It is superadditive when a jointly-designed, typically entangled, input across the uses transmits strictly more than that sum, so the one-shot formula undercounts and you must optimize over larger and larger blocks. Classical Shannon capacity is always additive; the quantum Holevo and coherent-information capacities are generically superadditive, which is why their true values are defined by a regularized limit rather than a formula you evaluate once.

Why is regularization such a problem — can't we just compute the limit?

Evaluating $\lim_n\frac1n C^{(1)}(\mathcal{N}^{\otimes n})$ requires optimizing an entropic functional over inputs to $\mathcal{N}^{\otimes n}$ , whose dimension grows exponentially in $n$ . There is no known bound on how large $n$ must be before $\frac1n C^{(1)}(\mathcal{N}^{\otimes n})$ is within a given tolerance of the limit, so truncating at any finite $n$ gives only a lower bound with no error control. This is why the field invests so heavily in additive upper bounds — quantities that single-letterize by construction and sandwich the capacity from above — since matching an achievable lower bound to a computable upper bound is the only general route to pinning a capacity down.

How can a channel have zero one-shot capacity but positive true capacity?

The single-letter quantity $Q^{(1)}=\max_\rho I_c(\rho,\mathcal{N})$ is neither concave nor additive, so its being zero for one use says nothing definitive about many uses. A channel can leak as much to the environment as it delivers to Bob on any single input — giving $I_c\le0$ — yet on an entangled input spread across several uses, code degeneracy makes several error patterns act identically, reducing the effective entropy fed to the environment below the naive sum. The averaged coherent information of $\mathcal{N}^{\otimes n}$ then turns positive even though $Q^{(1)}(\mathcal{N})=0$ . Superactivation is the extreme version, where even the individual channels have exactly zero capacity.

Does superadditivity break the data-processing inequality?

No, and conflating the two is a common error. The data-processing inequality concerns a fixed state and a single channel: information cannot increase when you process a system, and it holds for every channel use in a coding scheme. Superadditivity is a different kind of statement — about the optimization over inputs. When you may choose an entangled input correlated across many uses, the best such input can outperform the best product input, and the per-use information rises. No individual quantity increases under processing; a larger optimization domain simply yields a larger optimum. DPI is a "for all states" fact; superadditivity is a "there exists a better input" fact.

Which quantum capacity is the well-behaved one, and why does it matter?

The entanglement-assisted classical capacity $C_E$ is the cleanest object in the theory: it equals a single-letter maximum of the quantum mutual information $I(A;B)$ , and because mutual information is additive under tensor products, $C_E=C_E^{(1)}$ for every channel with no regularization. This matters because $C_E$ is genuinely computable — a finite convex optimization — so it serves as a benchmark and, since $C\le C_E$ and $Q\le\tfrac12 C_E$ , as a computable upper bound on the harder capacities. When shared entanglement is free (as in many quantum-network settings), $C_E$ is also the operationally relevant number, part of why entanglement distribution is treated as basic infrastructure.

Is degradability common, or are most channels non-degradable?

Degradable channels are a special class, but they include several of the most physically important channels: the erasure channel below $p=½$ , all dephasing channels, and amplitude-damping channels below damping $\tfrac12$ . For these the quantum capacity single-letterizes and is exactly computable, which is why they dominate textbook examples and resource accounting in memories and repeaters. Most generic channels — the depolarizing channel above threshold being the canonical example — are neither degradable nor antidegradable, and their true quantum capacity remains unknown, bounded but not determined. The theory of approximate degradability extends the tractable region: a channel close to degradable in diamond norm can be bounded tightly even when not exactly single-letter.