Schumacher Quantum Data Compression

Classical information theory begins with a sharp question: how many bits must you write down to record the output of a random source? Shannon answered it in 1948 — the irreducible cost is the entropy $H(X)$ bits per symbol, no fewer, asymptotically no more. Schumacher's 1995 theorem is the quantum sequel, and every bit as sharp. A quantum source emits, on each use, one of a menu of pure states $|\psi_i\rangle$ with probability $p_i$ ; equivalently it hands you copies of the average density operator $\rho=\sum_i p_i|\psi_i\rangle\langle\psi_i|$ . How many qubits per symbol must a compressor store so that the original state — coherences, phases, and all its entanglement with the rest of the world — is recovered with fidelity approaching one? The answer is the von Neumann entropy $S(\rho)=-\mathrm{Tr}(\rho\log_2\rho)$ . This is the single result that converts $S(\rho)$ from an algebraic expression into a physical resource count, exactly as Shannon's theorem does for $H(X)$ . The engine of the proof is the typical subspace — the quantum lift of the classical asymptotic equipartition property. Diagonalize $\rho$ , and $\rho^{\otimes n}$ becomes an i.i.d. classical source over its eigenvalues; almost all of its weight collapses onto a subspace of dimension $2^{nS(\rho)}$ , a vanishing sliver of the full $2^n$ -dimensional Hilbert space. Rotate that sliver into $\lceil nS\rceil$ qubits, discard the rest, and the source is compressed. This article proves both achievability and converse, and shows why "lossless" means entanglement fidelity — and why non-orthogonal signals compress below their classical description cost.

Key takeaways

Schumacher's noiseless coding theorem makes the von Neumann entropy $S(\rho)=-\mathrm{Tr}(\rho\log_2\rho)$ the operational cost of quantum information: a source of i.i.d. copies of $\rho$ compresses to $S(\rho)$ qubits per symbol, and no faithful scheme uses fewer.
The mechanism is the typical subspace: for $n$ copies, almost all the weight of $\rho^{\otimes n}$ lives in a subspace of dimension $\approx 2^{nS(\rho)}$ — the quantum analogue of the classical typical set.
Compression is a coherent rotate-and-truncate — an isometry onto $\lceil nS\rceil$ qubits and its inverse, never a measurement — so superpositions survive; it works because $\mathrm{Tr}(\Pi\rho^{\otimes n})\ge1-\epsilon$ .
Faithfulness is certified by entanglement fidelity $F_e$ : the channel must act trivially even on half of a maximally entangled purification, preserving all correlations with a reference.
The rate is set by the average state $\rho$ , not the label entropy $H(p_i)$ : since $S(\rho)\le H(p_i)$ , overlapping signals compress below the classical bound — genuinely sub-classical storage.
Schumacher compression is the single-letter prototype of quantum Shannon theory, seeding the Holevo bound, channel capacities, and entanglement concentration — and it coined the word "qubit".
Key idea: Replace the classical typical set by the typical subspace of $\rho^{\otimes n}$ , whose dimension grows as $2^{nS(\rho)}$ ; project onto it, store those $nS(\rho)$ qubits, and the von Neumann entropy is the exact, optimal compression rate.

The core insight

Replace the classical typical set by the typical subspace of $\rho^{\otimes n}$ , whose dimension grows as $2^{nS(\rho)}$ ; projecting onto it and storing those $nS(\rho)$ qubits compresses the source faithfully, and the von Neumann entropy is the exact, optimal rate.

Where this lives today

Schumacher compression is the quantitative backbone of quantum resource accounting, and it has moved from theorem to bench. Its typical-subspace method is the template for every later coding theorem — the Holevo–Schumacher–Westmoreland theorem for classical capacity, the Lloyd–Shor–Devetak theorem for quantum capacity, and quantum rate-distortion theory all run on typical projectors and entanglement fidelity. In entanglement concentration the same counting says $n$ copies of a partially entangled pure state distill to $\approx nS(\rho_A)$ Bell pairs, making $S$ the exchange rate for entanglement. Practically, $S(\rho)$ is the sizing rule for quantum memories and repeaters and the qubit budget in distributed-computing resource estimates. The theorem has also been demonstrated directly: NMR experiments (Mitsumori and co-workers, 2003) implemented it on nuclear spins, and Rozema and collaborators (Phys. Rev. Lett. 113, 160504, 2014) compressed a photonic qubit ensemble and recovered it with high fidelity, confirming the $S(\rho)$ rate on hardware. Open frontiers include finite-block-length refinements — where smooth min- and max-entropies replace $S(\rho)$ and quantify the $O(1/\sqrt n)$ overhead at realistic $n$ — and universal (blind) compressors that hit the rate without knowing $\rho$ .

The physical picture

Look at $\rho^{\otimes n}$ in the eigenbasis of $\rho$ . It is diagonal, and its eigenvalues are products $\lambda_{x_1}\cdots\lambda_{x_n}$ — precisely the probabilities of an i.i.d. classical source over the alphabet $\{x\}$ with distribution $\{\lambda_x\}$ . The law of large numbers does the work: for large $n$ , nearly every eigenvalue carrying appreciable weight is close to $2^{-nS}$ , so the weight concentrates on roughly $2^{nS}$ eigenvectors of nearly equal probability. Their span is the typical subspace — a low-dimensional shelf holding almost the entire state, and storing which shelf-position you occupy costs $\log_2(2^{nS})=nS$ qubits per block, i.e. $S$ qubits per symbol. The first figure follows that pipeline with real dimensions.

Figure

The Schumacher pipeline for a qubit source with S(ρ) ≈ 0.6009: of the 2¹⁰⁰ ≈ 1.3×10³⁰ dimensions of ρ⊗ⁿ, a typical subspace of only 2⁶⁰ ≈ 1.2×10¹⁸ carries weight ≥ 1−ε, so a unitary U rotates it into just 61 stored qubits instead of 100

The quantum twist over Shannon is coherence. You cannot simply measure which eigenstate occurred and record its label — that would collapse every superposition. Instead the encoder applies a unitary $U$ that coherently rotates the whole typical subspace into the first $\lceil nS\rceil$ qubits and discards the rest, a measurement-free move; because $\mathrm{Tr}(\Pi\rho^{\otimes n})\approx1$ , this barely disturbs the state. The second figure answers the theorem's other question — why the rate is set by the average state $\rho$ and not the labels $p_i$ . When two signals overlap they are partly indistinguishable and redundant, and the von Neumann entropy discounts that redundancy, giving $S(\rho)\le H(p_i)$ . Compression below the classical description length is not a paradox; it is the payoff of non-orthogonality.

Figure

Why the rate is sub-classical: for two equiprobable pure signals of overlap s (the modulus of their inner product), the average state has eigenvalues (1±s)/2, so S(ρ) = H₂((1+s)/2) falls from 1 to 0 as the states merge, always below the classical label entropy H(p) = 1 bit — at overlap s = 0.707 it is S = 0.6009 qubits

The core equations

The whole theorem is a statement about one number and one subspace. We name the rate, build the typical subspace that makes it achievable, then state the criterion for "faithful". The compression rate is the von Neumann entropy — the Shannon entropy of the eigenvalue spectrum of $\rho$ , in qubits because the logarithm is base two.

S(\rho)=-\mathrm{Tr}\big(\rho\log_2\rho\big)=-\sum_x\lambda_x\log_2\lambda_x \qquad (1)

where: $\rho=\sum_i p_i|\psi_i\rangle\langle\psi_i|$ — the average density operator of the source; $\lambda_x$ — its eigenvalues, a probability distribution ( $\lambda_x\ge0$ , $\sum_x\lambda_x=1$ ); $S(\rho)$ — the optimal qubits per source symbol. Reading it: $2^{S(\rho)}$ is the effective number of distinguishable directions the source explores, and everything below proves that count both achievable and unbeatable. Take $n$ independent uses. In the product eigenbasis $\rho^{\otimes n}$ is diagonal, its eigenvalues products — the hallmark of an i.i.d. classical source hiding inside the quantum state.

\rho^{\otimes n}=\sum_{x_1,\dots,x_n}\lambda_{x_1}\cdots\lambda_{x_n}\,|x_1\cdots x_n\rangle\langle x_1\cdots x_n| \qquad (2)

where: $|x_1\cdots x_n\rangle$ — a product eigenvector; $\lambda_{x_1}\cdots\lambda_{x_n}$ — its eigenvalue, drawn i.i.d. from $\{\lambda_x\}$ ; $n$ — the block length. Quantum compression has been reduced, exactly, to compressing a classical i.i.d. source over the eigenvalue alphabet. A sequence is typical when its per-symbol empirical entropy sits within $\delta$ of $S(\rho)$ . The span of the typical eigenvectors is the typical subspace, whose dimension is pinned within a factor $2^{n\delta}$ of $2^{nS}$ .

T_\delta^n=\Big\{(x_1,\dots,x_n):\Big|-\tfrac1n\log_2(\lambda_{x_1}\cdots\lambda_{x_n})-S(\rho)\Big|\le\delta\Big\},\quad (1-\epsilon)\,2^{n(S-\delta)}\le|T_\delta^n|\le 2^{n(S+\delta)} \qquad (3)

where: $T_\delta^n$ — the $\delta$ -typical index sequences; $\delta>0$ — entropy tolerance; $\epsilon$ — atypical weight, $\to0$ as $n\to\infty$ ; $|T_\delta^n|=\dim(\mathrm{supp}\,\Pi_\delta^n)$ — the typical-subspace dimension. There are $\approx2^{nS}$ typical sequences, each of eigenvalue near $2^{-nS}$ — nearly uniform on a subspace exponentially smaller than the full space. The typical projector selects that subspace; the high-probability theorem says it captures nearly all the state's weight.

\Pi_\delta^n=\sum_{(x_1,\dots,x_n)\in T_\delta^n}|x_1\cdots x_n\rangle\langle x_1\cdots x_n|,\qquad \mathrm{Tr}\big(\Pi_\delta^n\,\rho^{\otimes n}\big)\ge 1-\epsilon\ \ (n\ge n_0) \qquad (4)

where: $\Pi_\delta^n$ — orthogonal projector onto the span of typical eigenvectors; $n_0(\epsilon,\delta)$ — the block length beyond which the bound holds. So $\Pi_\delta^n$ acts like the identity on almost all of $\rho^{\otimes n}$ , and truncating to it loses only $O(\epsilon)$ . Faithfulness is measured not by the marginal state but by entanglement fidelity: the channel must preserve the source even when it is half of a maximally entangled purification. This is the correct "lossless" criterion because it certifies coherence and correlation, not merely statistics.

F_e(\rho,\mathcal{N})=\langle\Phi_\rho|\,(\mathrm{id}\otimes\mathcal{N})\big(|\Phi_\rho\rangle\langle\Phi_\rho|\big)\,|\Phi_\rho\rangle=\sum_k\big|\mathrm{Tr}(N_k\rho)\big|^2 \qquad (5)

where: $\mathcal{N}=\mathcal{D}_n\circ\mathcal{E}_n$ — the compression channel (encode then decode) with Kraus operators $\{N_k\}$ ; $|\Phi_\rho\rangle$ — a purification of $\rho$ on system+reference; $F_e\to1$ certifies all correlations with the reference survive. The Kraus form $\sum_k|\mathrm{Tr}(N_k\rho)|^2$ makes $F_e$ computable — a single dominant term close to $\Pi$ already forces $F_e$ near one. Putting the pieces together gives Schumacher's theorem: the rate $S(\rho)$ is at once achievable and optimal.

\inf\{R:\ F_e(\rho^{\otimes n},\mathcal{D}_n\circ\mathcal{E}_n)\xrightarrow{n\to\infty}1\}=S(\rho) \qquad (6)

where: $R$ — the rate in qubits per symbol, so $\mathcal{E}_n$ outputs $\lceil nR\rceil$ qubits; $\mathcal{E}_n,\mathcal{D}_n$ — the encoding and decoding channels. Any $R>S(\rho)$ is achievable with $F_e\to1$ (Derivation B) and any $R<S(\rho)$ forces $F_e\not\to1$ (Derivation C), so the infimum of faithful rates is exactly the von Neumann entropy.

Derivation AThe typical subspace and its dimension

The goal is to prove the two structural facts on which the theorem rests — that the typical subspace has dimension sandwiched between $(1-\epsilon)2^{n(S-\delta)}$ and $2^{n(S+\delta)}$ , and that it captures weight at least $1-\epsilon$ — from nothing but the law of large numbers applied to $-\log_2\lambda_x$ . Because $\rho^{\otimes n}$ is diagonal, its eigenvalues form an i.i.d. classical source, so the whole argument is classical asymptotic equipartition in quantum notation: compute the mean, invoke concentration, then run a two-sided counting argument.

Step 1. Take the negative log of a product eigenvalue and divide by $n$ : $-\tfrac1n\log_2(\lambda_{x_1}\cdots\lambda_{x_n})=\tfrac1n\sum_k(-\log_2\lambda_{x_k})$ , a sample mean of i.i.d. draws of the random variable $Z=-\log_2\lambda_x$ . Why: the log turns the multiplicative i.i.d. product into a sum the law of large numbers can grip.

Step 2. Its expectation is $\mathbb{E}[Z]=-\sum_x\lambda_x\log_2\lambda_x=S(\rho)$ . By the weak law of large numbers the sample mean concentrates: $\Pr\big[|\,\tfrac1n\sum_k Z_k-S(\rho)\,|\le\delta\big]\to1$ , which is the statement that $\Pr[(x_1,\dots,x_n)\in T_\delta^n]\ge1-\epsilon$ for $n\ge n_0$ . Why: the mean log-eigenvalue is the entropy, so typical sequences have eigenvalue near $2^{-nS}$ — the equipartition.

Step 3. Typicality bounds each eigenvalue: $2^{-n(S+\delta)}\le\lambda_{x_1}\cdots\lambda_{x_n}\le2^{-n(S-\delta)}$ . Since total probability is $\le1$ , $1\ge\sum_{T_\delta^n}\lambda\ge|T_\delta^n|\,2^{-n(S+\delta)}$ , hence $|T_\delta^n|\le2^{n(S+\delta)}$ . Why: eigenvalues each $\ge2^{-n(S+\delta)}$ summing to $\le1$ number at most $2^{n(S+\delta)}$ — pigeonhole.

Step 4. For the lower bound, the typical set carries weight $\ge1-\epsilon$ while each eigenvalue is $\le2^{-n(S-\delta)}$ : $1-\epsilon\le\sum_{T_\delta^n}\lambda\le|T_\delta^n|\,2^{-n(S-\delta)}$ , hence $|T_\delta^n|\ge(1-\epsilon)2^{n(S-\delta)}$ , and $\mathrm{Tr}(\Pi_\delta^n\rho^{\otimes n})=\sum_{T_\delta^n}\lambda\ge1-\epsilon$ . Why: read the other way the same inequalities force the count large — dimension genuinely $\approx2^{nS}$ . *The key identity is $\mathbb{E}[-\log_2\lambda_x]=S(\rho)$ ; the dimension bounds $(1-\epsilon)2^{n(S-\delta)}\le|T_\delta^n|\le2^{n(S+\delta)}$ and the high-probability result $\mathrm{Tr}(\Pi\rho^{\otimes n})\ge1-\epsilon$ follow directly from the steps above.*

Go deeper · Derivation B · 1 minAchievability of the rate S plus delta

Derivation BAchievability of the rate S plus delta

The goal is to exhibit an explicit encoder and decoder that store $\lceil n(S+\delta)\rceil$ qubits per block and prove their entanglement fidelity exceeds $1-2\epsilon$ , so that as $n\to\infty$ and $\delta\to0^+$ the rate descends to $S(\rho)$ with fidelity climbing to one. The construction is the coherent rotate-and-truncate: project onto the typical subspace, map it unitarily into a small register, invert on decoding. The one non-trivial estimate is that the composite channel's leading Kraus operator is essentially the typical projector, whose overlap with $\rho^{\otimes n}$ the high-probability theorem already controls.

Step 1. Encode. Let $U$ be a unitary carrying the $\le2^{n(S+\delta)}$ -dimensional typical subspace onto the state space of $m=\lceil n(S+\delta)\rceil$ qubits. Define $\mathcal{E}_n(\sigma)=U\big(\Pi\sigma\Pi+|0\rangle\langle a|\sigma|a\rangle\langle0|\big)U^\dagger$ , where the second term dumps all atypical weight into a fixed state $|0\rangle$ . Why: projecting first keeps the register only $m$ qubits wide; the failure term keeps the map trace-preserving without storing what was discarded.

Step 2. Decode by re-embedding: $\mathcal{D}_n(\tau)=U^\dagger\tau U$ , which maps the $m$ stored qubits back into the original $d^n$ -dimensional space. On the typical subspace the composite $\mathcal{D}_n\circ\mathcal{E}_n$ acts, to leading order, as $\sigma\mapsto\Pi\sigma\Pi$ . Why: decoding is the inverse rotation, so the only loss is the weight outside $\Pi$ , replaced by $|0\rangle$ .

Step 3. Read off the dominant Kraus operator. Writing $\mathcal{D}_n\circ\mathcal{E}_n(\sigma)=\sum_k N_k\sigma N_k^\dagger$ , the leading term is $N_0\approx\Pi$ , so $F_e\ge|\mathrm{Tr}(N_0\rho^{\otimes n})|^2\ge\big(\mathrm{Tr}(\Pi\rho^{\otimes n})\big)^2$ . Why: keeping only the projector term can only lower $F_e$ (all Kraus terms are non-negative), so this is a valid lower bound.

Step 4. Close with the high-probability theorem: $F_e\ge\big(\mathrm{Tr}(\Pi\rho^{\otimes n})\big)^2\ge(1-\epsilon)^2\ge1-2\epsilon$ , at rate $\tfrac1n\lceil n(S+\delta)\rceil\to S(\rho)+\delta$ . Letting $\epsilon,\delta\to0$ (allowed for $n$ large) drives $F_e\to1$ at rate $\to S(\rho)$ . Why: $(1-\epsilon)^2\ge1-2\epsilon$ converts "the projector nearly is the identity" into "the code is nearly perfect". *The chain $F_e\ge(\mathrm{Tr}\,\Pi\rho^{\otimes n})^2\ge(1-\epsilon)^2\ge1-2\epsilon$ confirms the fidelity bound, and the rate satisfies $\tfrac1n\lceil n(S+\delta)\rceil\to S(\rho)+\delta\to S(\rho)$ as $\delta\to0$ .*

Go deeper · Derivation C · 1 minThe converse — no faithful scheme beats S

Derivation CThe converse — no faithful scheme beats S

The goal is to prove the matching lower bound: any code compressing into fewer than $2^{n(S-\delta)}$ dimensions must have entanglement fidelity that fails to reach one, so $S(\rho)$ is a hard floor and, with achievability, the exact rate. The strategy is a variational bound — the fidelity of any rate- $R$ code is capped by the weight of $\rho^{\otimes n}$ catchable by any rank- $2^{nR}$ projector — plus the fact that the best such projector keeps the largest eigenvalues, which typicality has forced to be exponentially small once $R<S$ .

Step 1. The image of any rate- $R$ encoder lies in a subspace of dimension $\le2^{nR}$ ; let $Q$ project onto the decoder's output support, so $\mathrm{rank}\,Q\le2^{nR}$ . A standard fidelity bound gives $F_e(\rho^{\otimes n},\mathcal{D}_n\circ\mathcal{E}_n)\le\mathrm{Tr}(Q\,\rho^{\otimes n})$ . Why: information that never entered the $2^{nR}$ -dimensional bottleneck cannot be recovered, so the overlap caps the fidelity.

Step 2. Maximize the overlap over all rank- $2^{nR}$ projectors. By the Ky Fan / eigenvalue extremal principle the optimum keeps the $2^{nR}$ largest eigenvalues of $\rho^{\otimes n}$ : $\max_{\mathrm{rank}\,Q\le2^{nR}}\mathrm{Tr}(Q\rho^{\otimes n})=\sum_{\text{largest }2^{nR}}\lambda$ . Why: no subspace beats the top eigenvectors — the tightest possible converse, independent of the code.

Step 3. Split that sum into an atypical part (weight $\le\epsilon$ ) and a typical part whose eigenvalues are each $\le2^{-n(S-\delta)}$ : $\sum_{\text{largest }2^{nR}}\lambda\le\epsilon+2^{nR}\cdot2^{-n(S-\delta)}=\epsilon+2^{-n(S-\delta-R)}$ . Why: typicality flattens the spectrum, so $2^{nR}$ eigenvalues each $\le2^{-n(S-\delta)}$ collect at most the count times the cap.

Step 4. Choose $R<S-\delta$ ; then the exponent $-n(S-\delta-R)\to-\infty$ , so the typical term vanishes and $\mathrm{Tr}(Q\rho^{\otimes n})\to0$ , forcing $F_e\to0$ . Hence faithful compression requires $R\ge S(\rho)$ . Combined with Derivation B, the optimal rate is exactly $S(\rho)$ . Why: below the entropy there are too few large eigenvalues to hold the state, so fidelity collapses — the theorem is tight from both sides. *The Ky Fan optimum satisfies $\sum_{\text{top }2^{nR}}\lambda\le\epsilon+2^{-n(S-\delta-R)}\to0$ for $R<S-\delta$ , confirming that fidelity collapses below the entropy rate.*

Limiting cases

The rate $S(\rho)$ slides between no-cost and full-cost storage as the source moves from pure to maximally mixed, collapsing to the classical Shannon rate exactly when the signals become distinguishable.

Case	Result	Meaning
Pure source	$\rho=\vert\psi\rangle\langle\psi\vert\Rightarrow S(\rho)=0$	A known pure state carries no information; it compresses to zero qubits — the decoder just re-prepares $\vert\psi\rangle$ .
Maximally mixed	$\rho=I/d\Rightarrow S(\rho)=\log_2 d$	Every direction is equally likely; no redundancy, and the rate equals the full $\log_2 d$ qubits per symbol.
Orthogonal signals	$\langle\psi_i\vert\psi_j\rangle=\delta_{ij}\Rightarrow S(\rho)=H(p_i)$	Distinguishable states behave classically; the quantum rate collapses to the Shannon rate of the labels.
Non-orthogonal signals	$S(\rho)<H(\{p_i\})$	Overlap peaks the eigenvalue spectrum below the label entropy, permitting genuinely sub-classical compression.
Finite block length	$F_e\ge1-2\epsilon$ , not $F_e=1$	At finite $n$ the code is only near-perfect; exact reversibility is an asymptotic ( $n\to\infty$ ) statement.

Common trapSchumacher compression stores H(pᵢ) qubits, and is exactly lossless at every block length

Both halves are wrong. The rate is set by the von Neumann entropy of the average state, $S(\rho)$ , not the Shannon entropy $H(p_i)$ of the labels you prepared — and whenever the signals overlap, $S(\rho)$ is strictly smaller, so you store fewer qubits than any classical description of the labels needs. A source of $|0\rangle$ and $|+\rangle$ has $H(p)=1$ bit but $S(\rho)\approx0.60$ qubits, because non-orthogonal states are partly indistinguishable and $S(\rho)$ counts only distinguishable content. The second error is subtler: compression is lossless only asymptotically. At finite $n$ the entanglement fidelity is $\ge1-2\epsilon$ , close to but not equal to one, and the scheme fails on the atypical weight $\le\epsilon$ . It is faithful as $n\to\infty$ , not reversible on any single block — "noiseless" in the same asymptotic sense as Shannon's.

Worked examples

1 · Sub-classical compression of a non-orthogonal source (fully worked)

A source emits $|0\rangle$ and $|+\rangle=\tfrac1{\sqrt2}(|0\rangle+|1\rangle)$ with equal probability $\tfrac12$ . Find the classical label entropy, the average state $\rho$ , its eigenvalues, the von Neumann entropy, and the asymptotic number of qubits per symbol. How does it beat the classical bound?

Step 1. Classical labels. Two equiprobable symbols give $H(p)=\mathbf{1\ bit}$ . A classical recorder that stored which state was prepared would need one bit per use.

Step 2. Average state. With $|+\rangle\langle+|=\tfrac12\big(\begin{smallmatrix}1&1\\1&1\end{smallmatrix}\big)$ , $\rho=\tfrac12(|0\rangle\langle0|+|+\rangle\langle+|)=\big(\begin{smallmatrix}3/4&1/4\\1/4&1/4\end{smallmatrix}\big)$ .

Step 3. Eigenvalues. Trace $=1$ , determinant $=\tfrac{3}{16}-\tfrac1{16}=\tfrac18$ , so $\lambda_\pm=\tfrac12\big(1\pm\sqrt{1-\tfrac12}\big)=\tfrac12\big(1\pm\tfrac1{\sqrt2}\big)$ , numerically $\lambda_+=\mathbf{0.85355}$ , $\lambda_-=\mathbf{0.14645}$ . The overlap $\langle0|+\rangle=\tfrac1{\sqrt2}$ appears directly, as $(1\pm s)/2$ .

Step 4. Entropy. $S(\rho)=-0.85355\log_2 0.85355-0.14645\log_2 0.14645=\mathbf{0.6009\ qubits}$ , equivalently the binary entropy $H_2(0.85355)$ .

Step 5. Interpretation. The source compresses to $S(\rho)\approx0.60$ qubits per symbol asymptotically — strictly below the classical $H(p)=1$ bit — because $|0\rangle$ and $|+\rangle$ are not orthogonal, so partly redundant, and $S(\rho)$ charges only for the distinguishable part. Fewer than one qubit per symbol for a two-state source is impossible classically, routine quantumly. Answer: H(p) = 1 bit, but S(ρ) ≈ 0.6009 qubits per symbol.

2 · Sizing the typical subspace and the register

Take the same qubit source, eigenvalues $\{0.85355,0.14645\}$ , and $n=100$ copies. How many dimensions does the full space have, how large is the typical subspace, how many qubits are stored, and what fraction of the space does the typical subspace occupy?

Step 1. Full space. $\rho^{\otimes100}$ lives on $2^{100}\approx\mathbf{1.27\times10^{30}}$ dimensions — naive cost $100$ qubits per block.

Step 2. Typical dimension. Up to the $2^{n\delta}$ slack, $\dim\approx2^{nS}=2^{100\times0.6009}=2^{60.09}\approx\mathbf{1.2\times10^{18}}$ .

Step 3. Register size. The encoder rotates that subspace into $\lceil nS\rceil=\lceil60.09\rceil=\mathbf{61\ qubits}$ — rate $61/100=0.61$ qubits per symbol, converging to $0.6009$ as $n\to\infty$ .

Step 4. Occupancy. The typical subspace is a fraction $2^{60.09}/2^{100}=2^{-39.91}\approx\mathbf{9.7\times10^{-13}}$ of the full space — under one part in a trillion — yet carries weight $\ge1-\epsilon$ . An astronomically small corner of Hilbert space holds essentially the entire state, and compression just relabels that corner. Answer: 2¹⁰⁰ dims total, ~2⁶⁰ typical, 61 qubits stored, occupying ≈ 10⁻¹² of the space.

Physics of the portrait

The von Neumann entropy was written down in 1932 as an abstract functional of a density matrix. It took sixty-three years and one theorem to give it a job description — from the physicist who also named quantum information's unit of currency.

Origins · 1 minBenjamin Schumacher and how this was discovered

Benjamin Schumacher

b. 1960, American physicist · quantum information theory

Benjamin Schumacher is the physicist who turned the von Neumann entropy into an operational quantity and, in the same stroke, named the qubit. Trained at the University of Texas at Austin under John Archibald Wheeler and long a professor at Kenyon College, Schumacher asked in the early 1990s whether Shannon's noiseless coding theorem had a genuine quantum analogue — not "how many bits to describe a quantum state" but "how many quantum systems to store it faithfully". Working with Richard Jozsa, he introduced the word "qubit" around 1993 as the natural unit of compressed quantum information, crediting the coinage to a conversation with William Wootters. The 1994 Jozsa–Schumacher paper gave a first proof; Schumacher's 1995 "Quantum coding" (Phys. Rev. A 51, 2738) laid out the full theory — the typical subspace, the compression scheme, and the entanglement fidelity that certifies faithfulness. His later work with Michael Nielsen sharpened entanglement fidelity into a tool for noisy channels, seeding the coherent information at the heart of the quantum capacity theorem. Schumacher's theorem is now the opening chapter of quantum Shannon theory: every later capacity result inherits its typical-subspace method and its figure of merit. A single question about compression gave the field both its foundational theorem and its vocabulary.

History

1932

— von Neumann defines the quantum entropy S(ρ) = −Tr ρ log ρ

1948

— Shannon proves the classical noiseless source coding theorem, H(X) bits per symbol

1993

— Schumacher and Jozsa name the qubit as the unit of compressed quantum information

1994

— Jozsa and Schumacher publish a proof of the quantum noiseless coding theorem

1995

— Schumacher's Quantum coding establishes S(ρ) as the optimal compression rate

Going further

Texts. Nielsen and Chuang, Quantum Computation and Quantum Information, §12.2 (Theorems 12.4–12.5) · Preskill, Lecture Notes for Physics 219, Ch. 10 · Wilde, Quantum Information Theory, Ch. 18 (Schumacher Compression) · Watrous, The Theory of Quantum Information, Ch. 5 · Schumacher, "Quantum coding", Phys. Rev. A 51, 2738 (1995) · Jozsa and Schumacher, "A new proof of the quantum noiseless coding theorem", J. Mod. Opt. 41, 2343 (1994). The key results are: dimension bounds $(1-\epsilon)2^{n(S-\delta)}\le|T_\delta^n|\le2^{n(S+\delta)}$ from eigenvalue equipartition; high-probability theorem $\mathrm{Tr}(\Pi\rho^{\otimes n})\ge1-\epsilon$ via the law of large numbers; achievability bound $F_e\ge(\mathrm{Tr}\,\Pi\rho^{\otimes n})^2\ge1-2\epsilon$ ; and the converse via Ky Fan eigenvalue maximization. The worked example uses $\rho=\tfrac12(|0\rangle\langle0|+|+\rangle\langle+|)$ with eigenvalues $\tfrac12(1\pm\tfrac1{\sqrt2})=\{0.85355,0.14645\}$ , giving $S=H_2(0.85355)\approx0.6009$ ; for $n=100$ , $2^{nS}=2^{60.09}$ , $\lceil nS\rceil=61$ , and the typical subspace occupies $\approx9.7\times10^{-13}$ of the full space. The entanglement fidelity formula $F_e=\sum_k|\mathrm{Tr}(N_k\rho)|^2$ is the standard Kraus-operator expression.

Prerequisites. Von Neumann Entropy · Shannon Entropy and Classical Information Measures · Quantum Channels and Completely Positive Trace-Preserving Maps · Kraus Representation of Quantum Channels. Leads to. Holevo Bound · Quantum Conditional Entropy and Mutual Information · Data-Processing Inequality · Strong Subadditivity of Quantum Entropy.

Ask the tutor: Ask me to walk through why the typical subspace of $\rho^{\otimes n}$ shrinks to $2^{nS}$ dimensions, or to compute the Schumacher rate and register size for any ensemble you name.

Is Schumacher compression the same as classical Shannon compression applied to qubits?

No — and the difference is the point. Classical compression of the labels costs $H(p_i)$ bits per symbol, the Shannon entropy of which state you prepared. Schumacher compression costs $S(\rho)$ qubits, the von Neumann entropy of the average state, and whenever the signals overlap, $S(\rho)<H(p_i)$ . The saving is purely quantum: non-orthogonal states cannot be perfectly distinguished, so recording their labels stores information you can never fully read out. Schumacher's scheme also never measures — it rotates coherently — so it preserves superpositions and entanglement that a scheme built on measured labels would destroy.

What is the typical subspace, and why is it so much smaller than the Hilbert space?

It is the span of the product eigenvectors of $\rho^{\otimes n}$ whose eigenvalue is close to $2^{-nS}$ . Its dimension is about $2^{nS}$ , while the full space has $d^n$ ; for a qubit source with $S=0.6$ and $n=100$ that is $2^{60}$ inside $2^{100}$ — under a trillionth of the space. It is small because the eigenvalues concentrate: by the law of large numbers almost all the weight sits on the $\approx2^{nS}$ sequences whose empirical entropy is near $S$ . Everything else, though far more numerous, carries negligible probability.

Why does the theorem use entanglement fidelity instead of ordinary fidelity?

Ordinary (average) fidelity only checks that the marginal output resembles the input on average. That is too weak: a channel could scramble phases or destroy entanglement with an outside system while still returning the right average state. Entanglement fidelity closes the gap by feeding the compressor one half of a maximally entangled purification of $\rho$ and demanding the joint state survive. If $F_e\to1$ , all coherences and correlations with any reference are preserved — the state is genuinely intact, not merely statistically plausible. This is why $F_e\to1$ is the correct definition of "lossless".

Does compression require actually diagonalizing ρ?

The standard scheme assumes you know $\rho$ — the encoder needs the typical projector, which depends on the eigenbasis. But knowing $\rho$ is enough; you never need the individual signal states $|\psi_i\rangle$ or their probabilities, only the average. Moreover, universal (blind) Schumacher compressors reach the rate $S(\rho)$ without knowing $\rho$ in advance — by estimating the spectrum on a sacrificial fraction of copies or projecting onto a universal typical subspace. So $S(\rho)$ is the rate even for an ignorant compressor, at vanishing overhead.

Is Schumacher compression exactly reversible, like a lossless zip file?

Not exactly, and not at finite block length. The entanglement fidelity is $\ge1-2\epsilon$ , arbitrarily close to one but not equal to it, and the scheme fails on the atypical weight $\le\epsilon$ outside the typical subspace. In the limit $n\to\infty$ the error vanishes and recovery becomes perfect — "noiseless" in the same asymptotic sense as Shannon's. For any finite block you accept a small, controllable infidelity, and one-shot refinements using smooth min/max entropies quantify how small it can be at given $n$ .

How does Schumacher compression connect to the Holevo bound?

They are two sides of quantum resource accounting. Schumacher says a source of average state $\rho$ needs $S(\rho)$ qubits to store faithfully. The Holevo bound says that from the same ensemble you can read out at most $\chi\le S(\rho)$ classical bits by any measurement, with equality only for orthogonal states. So you may move nearly $S(\rho)$ qubits yet recover far fewer classical bits — quantum information you can hold but not fully classicalize, which is exactly what makes it a distinct resource.