Quantum Reverse Shannon Theorem

The reverse Shannon theorem is the load-bearing tool in the modern theory of strong converses and channel simulation. Its most active descendant is one-shot and finite-block-length quantum information: the decoupling and quantum-state-splitting primitives that prove QRST also compute exact resource costs for state redistribution, quantum state merging, and the "mother/father" protocol family, and they underpin the second-order refinements needed for realistic, non-asymptotic communication. More broadly, reverse Shannon simulations bound the capacities that don't single-letterize: if a code for the unassisted classical, quantum, or private capacity can be simulated cheaply, it cannot carry more than those resources allow — yielding some of the strongest known strong-converse bounds. And the conceptual payoff endures: QRST pins an operational quantity, quantum mutual information, with a matched pair of tasks — coding and simulation — giving it the status reversible-engine arguments give entropy in thermodynamics.

Think of a quantum channel as a commodity bought and sold on an open market. The direct entanglement-assisted theorem is the "buy" side: spend $C_E$ noiseless bits (with free entangled pairs, or ebits) and you manufacture one channel's worth of communication. The reverse theorem is the "sell" side: hand over a physical channel and, with free entanglement, resell it for $C_E$ noiseless bits. Because the buy and sell prices coincide the market is frictionless: channels and bits are reversibly interconvertible at the single price $C_E$. The first figure draws that closed loop, anchored by the one instance you can compute by hand — the noiseless qubit, whose price is exactly 2 bits, read as teleportation one way and superdense coding the other.

Why entanglement is the load-bearing ingredient: a maximally entangled resource lets the sender teleport the action of the channel rather than physically build it. She applies a random unitary to her half of an entangled pair, arranges for the channel's effect to happen, and classically tells the receiver which unitary occurred — but only after using the entanglement to decouple the leftover, uncorrelated information so the classical description shrinks from a wasteful $\log d$ bits to $I(R;B)$ bits. That decoupling is the whole trick. The strong, entangled-input requirement is what makes the price exactly $I(R;B)$ and not less: a channel that reproduces only the right average output is cheap, but one that must preserve every correlation with a reference, behaving correctly even on half of a Bell state, is expensive — its worst case is the maximally entangled input, whose fingerprint is the channel's Choi state. The second figure makes the price concrete: the qubit erasure channel, whose cost slides linearly from 2 to 0 as the erasure probability climbs.

The theorem lives at the meeting point of two operational quantities — a capacity and a simulation cost — that turn out to be one number. First, the definition of that number: the entanglement-assisted classical capacity, an optimization of quantum mutual information over pure inputs shared with a reference.

where: $\mathcal{N}_{A'\to B}$ — the channel being characterized and simulated; $\phi_{RA'}$ — a pure bipartite input, with $R$ a reference held fixed; $I(R;B)=S(R)+S(B)-S(RB)$ — quantum mutual information of the output; $C_E$ — both the forward capacity and, by QRST, the reverse simulation cost. Reading: feed the channel half of a pure state and ask how much correlation with the untouched half survives; the best input sets $C_E$. The reverse theorem states that this same $C_E$ is the asymptotic classical-bit cost of simulating the channel with free entanglement — an equality of rates, not an inequality.

where: $\mathcal{M}_n$ — a protocol using free entanglement plus forward classical communication; $\mathcal{N}^{\otimes n}$ — $n$ parallel channel uses; the limit is over vanishing simulation error. Reading: no simulation beats $C_E$ per use (converse), and one achieves it (direct part). "Simulate" has to be defined carefully, and the strong criterion is the whole difficulty of the theorem: the simulation must be indistinguishable from the true channel even on inputs entangled with an arbitrary reference $R$, measured in trace distance.

where: $\psi_{RA^n}$ — an arbitrary pure input, possibly maximally entangled with $R$; $\|\cdot\|_1$ — trace norm, so $\tfrac12\|\cdot\|_1$ is trace distance; $\varepsilon_n$ — error vanishing in block length. Reading: correctness is demanded uniformly over all inputs at once, which is why the price is the entanglement-maximized $C_E$. The simplest nontrivial instance, the noiseless qubit channel, exposes the reversibility directly as a chain of resource inequalities that close on themselves.

where: $\succeq$ — "can asymptotically simulate / is at least as powerful as"; cbit — one noiseless classical bit; ebit — one maximally entangled pair (free here); qubit — one noiseless qubit channel use. Reading: the left inequality is teleportation, the right is superdense coding; together they make one qubit reversibly worth two cbits, so $C_E(\mathrm{id}_2)=2$. Behind the capacity sits a single entropic object: quantum mutual information is a relative entropy, $I(R;B)=S(\sigma_{RB}\,\|\,\sigma_R\otimes\sigma_B)$, a distance from the uncorrelated product $\sigma_R\otimes\sigma_B$. Because relative entropy obeys strong subadditivity, $C_E$ is exactly additive — it single-letterizes, unlike the unassisted capacities, and this is the structural fact that lets the reverse rate match the forward rate.

where: $S(\rho\|\tau)=\mathrm{Tr}\,\rho(\log\rho-\log\tau)$ — quantum relative entropy; $\otimes$ — parallel channel use; $C$ — unassisted (regularized Holevo) classical capacity; $Q_E$ — entanglement-assisted quantum capacity, exactly half of $C_E$ by teleportation and superdense coding. Reading: $C_E$ is the well-behaved, additive capacity, and the anchor from which the assisted quantum rate follows by a factor of two.

How the simulation cost specializes for representative channels — all rates in bits per channel use, with entanglement free:

1 · The noiseless qubit costs exactly two bits (the teleportation instance)

Compute the simulation cost $C_E(\mathrm{id}_2)$ of a perfect one-qubit channel from equation (1), and identify the protocol that achieves it.

Step 1. Feed the channel one half of the maximally entangled input $\phi_{RA'}=\Phi^+$. Since $\mathrm{id}_2$ does nothing, the output $\sigma_{RB}$ is itself a maximally entangled two-qubit (Bell) state. Its marginals are maximally mixed, $S(R)=S(B)=\log 2 = 1$ bit, and the joint state is pure, $S(RB)=0$.

Step 2. Mutual information: $I(R;B)=S(R)+S(B)-S(RB)=1+1-0=\mathbf{2\ \text{bits}}$. No input does better (the maximally mixed marginal already saturates each term), so $C_E(\mathrm{id}_2)=\max_\phi I(R;B)=\mathbf{2}$.

Step 3. The achieving protocol is teleportation read as a simulation: with one shared ebit, sending 2 classical bits reproduces a perfect qubit channel. Superdense coding gives the reverse inequality ($1$ qubit $+\,1$ ebit $\succeq 2$ cbits), so the noiseless qubit is reversibly worth two cbits — equation (4) with both arrows. The converse confirms it: $\log L\ge nC_E=2n$, i.e. rate 2, is exactly teleportation's cost and cannot be beaten. C_E(id₂) = 2 bits, achieved by teleportation, optimal by the classical-register converse.

2 · The qubit erasure channel (fully worked)

The qubit erasure channel maps $\rho\mapsto (1-p)\rho + p\,|e\rangle\langle e|$, flagging "erased" with probability $p$ on an orthogonal flag $|e\rangle$. Find its simulation cost, evaluate it at $p=0.25$, and interpret the endpoints.

Step 1. Feed half of $\Phi^+$. With probability $1-p$ the receiver holds the intact qubit, maximally entangled with $R$ (the full 2 bits, as in Example 1); with probability $p$ he holds the flag $|e\rangle$, orthogonal to and uncorrelated with $R$ (0 bits). The flag is classically distinguishable from the qubit subspace, so the branches do not interfere and the reference–output mutual information is the weighted average.

Step 2. Hence $C_E = (1-p)\cdot 2 + p\cdot 0 = \mathbf{2(1-p)\ \text{bits}}$, the value plotted in the second figure and the known entanglement-assisted capacity of the erasure channel (Bennett–DiVincenzo–Smolin).

Step 3. At $p=0.25$: $C_E = 2(1-0.25)=2(0.75)=\mathbf{1.5\ \text{bits per use}}$. So simulating $n$ uses of the $25\%$-erasure channel on entangled inputs costs asymptotically $1.5\,n$ noiseless classical bits, with free entanglement. Endpoints: $p=0$ recovers $\mathrm{id}_2$ and cost 2 (Example 1); $p=1$ gives cost 0, because the output is always the fixed flag $|e\rangle$, carries nothing about the input, and Bob prepares it locally — a sanity check that the strong criterion charges only for correlation the channel genuinely preserves. C_E = 2(1−p); at p = 0.25 the cost is 1.5 bits/use.

The reverse Shannon theorem is inseparable from the person who framed both its question and much of the machinery that answered it: channel simulation as the dual of coding, the resource-inequality bookkeeping that makes ebits and cbits fungible, and the two protocols — teleportation and superdense coding — that anchor the loop all trace to one research program.

References. M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information (10th Anniv. ed.), Ch. 11–12 · M. M. Wilde, Quantum Information Theory (2nd ed.), Ch. 20–21 (entanglement-assisted capacity) and Ch. 24 (reverse Shannon) · J. Preskill, Lecture Notes on Quantum Computation, Ch. 10 · J. Watrous, The Theory of Quantum Information, Ch. 8 · A. Kitaev, A. Shen, M. Vyalyi, Classical and Quantum Computation (background on channels and entropy) · C. H. Bennett, P. W. Shor, J. A. Smolin, A. V. Thapliyal, "Entanglement-Assisted Capacity of a Quantum Channel and the Reverse Shannon Theorem," IEEE Trans. Inf. Theory 48 (2002) 2637 · C. H. Bennett, I. Devetak, A. W. Harrow, P. W. Shor, A. Winter, "The Quantum Reverse Shannon Theorem and Resource Tradeoffs for Simulating Quantum Channels," IEEE Trans. Inf. Theory 60 (2014) 2926 · M. Berta, M. Christandl, R. Renner, "The Quantum Reverse Shannon Theorem Based on One-Shot Information Theory," Commun. Math. Phys. 306 (2011) 579.

Provenance. Capacity formula $C_E=\max I(R;B)$ and its additivity: BSST (2002), Wilde Ch. 20–21. Strong QRST and its proof structure (decoupling, quantum state splitting, one-shot smoothing): BDHSW (2014) and Berta–Christandl–Renner (2011). A verifier should recompute: (i) $C_E(\mathrm{id}_2)=2$, depolarizing $C_E=0$, erasure $C_E=2(1-p)$ from equation (1); (ii) additivity of $C_E$ from strong subadditivity; (iii) the surviving-rate identity $I(R;B)=S(R)+S(B)-S(E)$ using the purity $S(RB)=S(E)$ of the Stinespring dilation — note this corrects the migrated draft's $S(B)-S(E|R)$, which instead equals $S(R)$; (iv) the converse $\log L\ge nC_E-o(n)$ from data processing plus the classical-register bound $I(R^n;M|T_B)\le\log L$ (factor 1, checked against the teleportation cost 2 for id₂). Cross-references: Nielsen & Chuang Ch. 11–12, Preskill Ch. 10, Watrous Ch. 8.

Prerequisites. Quantum Channels and CPTP Maps · Stinespring Dilation Theorem · Quantum Conditional Entropy and Mutual Information · Von Neumann Entropy · Strong Subadditivity of Quantum Entropy · Data-Processing Inequality. Leads to. Holevo Bound · Complementary Channels and Degradability · Choi–Jamiołkowski Isomorphism.

Ask the tutor: Ask me to walk through why simulating a perfect qubit costs exactly two classical bits — and how teleportation and superdense coding close that into a reversible loop.