Typical Subspaces and Quantum Asymptotic Equipartition

A memoryless quantum source emits the same state $\rho$ independently $n$ times, handing you the enormous joint state $\rho^{\otimes n}$ on a Hilbert space of dimension $d^n$ . The central surprise of quantum Shannon theory is that almost none of that space is used: nearly all of the state's weight piles into a typical subspace of dimension only $\approx 2^{nS(\rho)}$ , exponentially smaller than $d^n$ unless the source is maximally mixed. That single fact — a huge space, a tiny occupied corner — is the quantum asymptotic equipartition property (AEP), the engine behind every i.i.d. coding theorem in the subject. The whole construction reduces to one idea. Diagonalise $\rho=\sum_x p(x)\,|x\rangle\langle x|$ ; its eigenvalues form a probability distribution $p$ . Because $\rho^{\otimes n}$ is diagonal in the product eigenbasis with eigenvalues $p(x^n)=\prod_i p(x_i)$ , the eigenvalues of the quantum state are literally the i.i.d. probabilities of a classical source with letter distribution $p$ . Classical typicality of strings $x^n$ selects a set of eigenvectors whose span is the quantum typical subspace, and since $H(p)=S(\rho)$ the classical dimension count $2^{nH(p)}$ transfers verbatim to $2^{nS(\rho)}$ . Two properties do all the work: high probability (the projector catches all but $\epsilon$ of the weight) and equipartition (inside the subspace every eigenvalue is nearly the uniform $2^{-nS}$ , so the projected state is nearly maximally mixed). Feed those into a coding argument and you get Schumacher's compression theorem, the operational meaning of $S(\rho)$ ; sharpen them into relative-entropy and conditional typicality and you get the Holevo–Schumacher–Westmoreland capacity and quantum Stein's lemma, with a one-shot refinement replacing $S$ by smoothed entropies that recover $S$ as $n\to\infty$ .

⚡ Key Takeaways

For an i.i.d. source $\rho^{\otimes n}$ with $\rho=\sum_x p(x)|x\rangle\langle x|$ , the eigenvalues are products $p(x^n)=\prod_i p(x_i)$ , so quantum typicality is the classical typical-set theorem applied to the eigenvalue distribution.
The typical subspace $T_n^\delta$ has dimension $\le 2^{n(S(\rho)+\delta)}$ yet carries almost all the weight, $\mathrm{Tr}\,\Pi_n^\delta\rho^{\otimes n}\ge 1-\epsilon$ , where $S(\rho)=-\mathrm{Tr}\,\rho\log_2\rho$ .
Inside it every eigenvalue is nearly uniform, $2^{-n(S+\delta)}\le p(x^n)\le 2^{-n(S-\delta)}$ — the quantum AEP, making the projected state near-maximally-mixed on $\approx 2^{nS}$ dimensions.
These two properties are the engine of Schumacher compression (optimal rate $S(\rho)$ qubits/symbol); relative-entropy typicality powers the HSW capacity and quantum Stein's lemma (exponent $D(\rho\|\sigma)$ ).
At finite block length the rates are smoothed min/max-entropies $S_{\min}^\epsilon,\,S_{\max}^\epsilon$ , converging to $S(\rho)$ with $O(1/\sqrt n)$ corrections set by the entropy variance.
Key idea: Diagonalise $\rho=\sum_x p(x)|x\rangle\langle x|$ ; the eigenvalues of $\rho^{\otimes n}$ are the i.i.d. probabilities $p(x^n)$ , so quantum typicality is classical typicality of the eigenvalue distribution — and Shannon entropy of $p$ equals von Neumann entropy of $\rho$ .

💡 The Core Insight

Diagonalise $\rho=\sum_x p(x)|x\rangle\langle x|$ ; then the eigenvalues of $\rho^{\otimes n}$ are the i.i.d. probabilities $p(x^n)$ , so quantum typicality is classical typicality of the eigenvalue distribution — and Shannon entropy of $p$ equals von Neumann entropy of $\rho$ .

The physical picture

Picture $\rho^{\otimes n}$ as a colossal diagonal matrix whose entries are the i.i.d. probabilities $p(x^n)$ . The law of large numbers acting on $-\tfrac1n\log_2 p(x^n)=\tfrac1n\sum_i[-\log_2 p(x_i)]$ forces this "empirical entropy per symbol" to concentrate on its mean $S(\rho)$ , so almost every eigenvalue that matters sits near $2^{-nS}$ and there are only about $2^{nS}$ of them. Keep those eigenvectors and you have a subspace holding $\approx 1$ of the total trace in an exponentially small sliver of the $d^n$ dimensions. The first figure sets the three coexisting scales; the second shows the concentration as $n$ grows.

Figure

Three scales for the qubit ρ = diag(0.9, 0.1) at n = 100: the full space has dimension 2¹⁰⁰ ≈ 1.27×10³⁰, yet the typical subspace of dimension ≈ 2⁴⁶·⁹ ≈ 1.3×10¹⁴ carries ≥ 1 − ε of the weight, each kept eigenvalue ≈ 2⁻⁴⁶·⁹

The concentration is the weak law of large numbers in operator clothing. For small $n$ the empirical entropy per symbol is spread out and many strings straddle the boundary; as $n$ grows the distribution sharpens onto the mean $S(\rho)$ and the fixed-width typical band $[S-\delta,\,S+\delta]$ scoops up a fraction approaching one. The band width $\delta$ is fixed before sending $n\to\infty$ — getting that order backwards is the classic beginner's error.

Figure

Concentration of the empirical entropy per symbol −(1/n) log₂ p(xⁿ) onto its mean S(ρ) = 0.469 bits for ρ = diag(0.9, 0.1): as n grows the distribution sharpens from the broad green curve to the narrow blue one, and the fixed typical band of half-width δ captures weight → 1

The typical subspace is "flat": restricted to it, $\rho^{\otimes n}$ is within $\epsilon$ of the maximally mixed state on $\approx 2^{nS}$ dimensions, and such a state loads reversibly into $\log_2$ (dimension) qubits — so the source carries no more quantum information per symbol than $S(\rho)$ . Sharper variants "fan out" the same concentration against different distributions: conditional typicality (which drives channel coding) and relative-entropy typicality with exponent $D(\rho\|\sigma)$ (which drives hypothesis testing).

The core equations

The starting point is spectral: tensoring a diagonal operator with itself keeps it diagonal and the eigenvalues multiply, so the joint state's eigenvalues are exactly the i.i.d. string probabilities.

\rho=\sum_{x} p(x)\,|x\rangle\langle x|,\qquad \rho^{\otimes n}=\sum_{x^n} p(x^n)\,|x^n\rangle\langle x^n|,\quad p(x^n)=\prod_{i=1}^n p(x_i) \qquad (1)

where: $\rho$ — single-letter source density operator on $\mathcal{H}$ , $\dim\mathcal{H}=d$ ; $p(x)$ — eigenvalues of $\rho$ , a probability distribution; $|x^n\rangle=|x_1\rangle\otimes\cdots\otimes|x_n\rangle$ — product eigenvector; $p(x^n)$ — eigenvalue of $\rho^{\otimes n}$ , the i.i.d. probability of the string $x^n$ . Reading: the quantum problem is now classical, at the cost of a change of basis. The bridge that makes the dimension count come out in the quantum information measure is the identity between von Neumann entropy and the Shannon entropy of the spectrum.

S(\rho)=-\mathrm{Tr}\,\rho\log_2\rho=-\sum_x p(x)\log_2 p(x)=H(p) \qquad (2)

where: $S(\rho)$ — von Neumann entropy in bits; $H(p)$ — Shannon entropy of the eigenvalue distribution. Reading: $S(\rho)$ depends on $\rho$ only through its spectrum, so equal-spectrum states have equal-dimension typical subspaces. Typicality selects the strings whose empirical entropy per symbol is within $\delta$ of the mean; the projector onto their eigenvectors is the typical projector.

T_n^\delta=\Big\{x^n: \Big|-\tfrac1n\log_2 p(x^n)-S(\rho)\Big|\le \delta\Big\},\qquad \Pi_n^\delta=\sum_{x^n\in T_n^\delta}|x^n\rangle\langle x^n| \qquad (3)

where: $T_n^\delta$ — the $\delta$ -typical set of eigenstring labels; $\Pi_n^\delta$ — orthogonal projector onto the typical subspace; $\delta$ — tolerance fixed before $n\to\infty$ . Reading: $\Pi_n^\delta$ is a quantum projector whose definition is dictated by a classical membership test on eigenvalues. Rearranging the membership inequality gives the equipartition bound: inside the subspace, every eigenvalue is squeezed into a narrow window around $2^{-nS}$ .

\forall x^n\in T_n^\delta:\quad 2^{-n(S(\rho)+\delta)}\le p(x^n)\le 2^{-n(S(\rho)-\delta)} \qquad (4)

where: $p(x^n)$ — an eigenvalue inside the typical subspace, nearly uniform. Reading: this near-uniformity is the "equi-partition" that makes the projected state look maximally mixed. The high-probability property is the payoff of the law of large numbers: the projector catches essentially all of the weight once $n$ is large.

\mathrm{Tr}\big(\Pi_n^\delta\,\rho^{\otimes n}\big)=\sum_{x^n\in T_n^\delta} p(x^n)\ge 1-\epsilon\qquad\text{for }n\ge n_0(\epsilon,\delta) \qquad (5)

where: $\epsilon$ — failure weight, driven to zero as $n\to\infty$ at fixed $\delta$ ; $n_0$ — a threshold block length set by the entropy variance through Chebyshev. Reading: "high probability" and "small subspace" are the two hands of the theorem, made consistent by (6). Combining the eigenvalue window (4) with the weight bound (5) sandwiches the dimension between exponentials of the same leading rate $S(\rho)$ .

(1-\epsilon)\,2^{n(S(\rho)-\delta)}\le \dim T_n^\delta=\mathrm{Tr}\,\Pi_n^\delta\le 2^{n(S(\rho)+\delta)} \qquad (6)

where: $\dim T_n^\delta$ — rank of $\Pi_n^\delta$ , the number of typical eigenstrings. Reading: to leading exponential order the dimension is $2^{nS(\rho)}$ — the qubits Schumacher compression must keep. Because the eigenvalues are nearly uniform, the renormalised projected state is close to maximally mixed on the typical subspace — which makes reversible loading into $nS$ qubits possible.

\rho_n^{\mathrm{typ}}=\frac{\Pi_n^\delta\,\rho^{\otimes n}\,\Pi_n^\delta}{\mathrm{Tr}\big(\Pi_n^\delta\,\rho^{\otimes n}\big)}\ \approx\ \frac{\Pi_n^\delta}{\dim T_n^\delta} \qquad (7)

where: $\rho_n^{\mathrm{typ}}$ — the projected, renormalised state. Reading: a near-maximally-mixed state on $K$ dimensions is reversibly interconvertible with $\log_2 K$ mixed qubits — why $S(\rho)$ is the qubit cost. Run against a second reference state $\sigma$ , the same machinery tests the relative-typical subspace and yields the exponent that governs hypothesis testing.

\mathrm{Tr}\big(\Pi_n^\delta\,\sigma^{\otimes n}\big)\ \le\ 2^{-n(S(\rho)-\delta)}\,\mathrm{rank}\,\Pi_n^\delta\ \le\ 2^{-n[D(\rho\|\sigma)-2\delta]} \qquad (8)

where: $\sigma$ — reference (null-hypothesis) state; $D(\rho\|\sigma)=\mathrm{Tr}\,\rho(\log_2\rho-\log_2\sigma)$ — quantum relative entropy, the optimal error exponent. Reading: the type-II error of the typical-subspace test decays at rate $D(\rho\|\sigma)$ — this is quantum Stein's lemma. Finally the one-shot generalisation: at finite $n$ the rates are smoothed min- and max-entropies, and the quantum AEP says both collapse to $S(\rho)$ in the i.i.d. limit.

\lim_{\epsilon\to0}\lim_{n\to\infty}\tfrac1n S_{\max}^{\epsilon}(\rho^{\otimes n})=\lim_{\epsilon\to0}\lim_{n\to\infty}\tfrac1n S_{\min}^{\epsilon}(\rho^{\otimes n})=S(\rho) \qquad (9)

where: $S_{\min}^\epsilon,\,S_{\max}^\epsilon$ — smoothed min- and max-entropies bounding randomness/security and compression/decoupling rates. Reading: von Neumann entropy is the asymptotic face of a family of one-shot entropies, and the AEP unifies them.

Derivation AReduction of quantum typicality to the classical asymptotic equipartition property

The goal is to show that the quantum typical-subspace theorem contains no new probability theory: it is the classical weak law of large numbers applied to the random variable $-\log_2 p(X_i)$ drawn from the eigenvalue distribution $p$ , re-read as a statement about the operator $\Pi_n^\delta\rho^{\otimes n}$ . We identify the random variable, take its empirical mean, invoke Chebyshev's inequality for an explicit block-length threshold, and translate the probability into a trace — proving the high-probability property (5) and pinning down $n_0(\epsilon,\delta)$ .

Step 1. Define $Y_i=-\log_2 p(X_i)$ , where $X_i$ are i.i.d. with law $p$ (the eigenvalues of $\rho$ ). Then $\mathbb{E}[Y_i]=H(p)=S(\rho)$ and the variance $V(\rho)=\sum_x p(x)\log_2^2 p(x)-S(\rho)^2$ is finite. Why: the surprisal $-\log_2 p$ is the quantity whose average is the entropy, so concentration of its empirical mean is concentration of the eigenvalue exponent.

Step 2. Because $p(x^n)=\prod_i p(x_i)$ , the log of the product is a sum, so the empirical entropy per symbol is the sample mean, $-\tfrac1n\log_2 p(X^n)=\tfrac1n\sum_{i=1}^n Y_i$ . Why: independence turns a product of probabilities into a sum of surprisals the LLN can act on.

Step 3. Chebyshev on the sample mean gives $\Pr\!\big[\,\big|\tfrac1n\sum_i Y_i-S(\rho)\big|>\delta\,\big]\le V(\rho)/(n\delta^2)\to0$ at fixed $\delta$ . Setting the right side to $\epsilon$ gives $n_0(\epsilon,\delta)=V(\rho)/(\epsilon\delta^2)$ . Why: this forces the order of limits — $\delta$ fixed while $n$ grows, or $V/(n\delta^2)$ never shrinks.

Step 4. Translate to operators. The event " $X^n\in T_n^\delta$ " has probability $\sum_{x^n\in T_n^\delta}p(x^n)=\mathrm{Tr}(\Pi_n^\delta\rho^{\otimes n})$ , so $\mathrm{Tr}(\Pi_n^\delta\rho^{\otimes n})\ge 1-V(\rho)/(n\delta^2)\ge 1-\epsilon$ — equation (5). Why: the projector's expectation value is a classical probability in disguise, so a classical bound is automatically an operator statement. *Verified: standard AEP (Cover & Thomas Ch. 3) fused with $S(\rho)=H(p)$ ; for $\rho=\mathrm{diag}(0.9,0.1)$ one finds $V(\rho)=0.904$ bits², matching the worked example.*

Derivation BSchumacher compression rate from the typical subspace

The goal is to convert the two typical-subspace properties into a coding theorem: that $S(\rho)$ qubits per symbol suffice to compress the i.i.d. source $\rho$ with vanishing error, and that no smaller rate can. We build an encoder that measures the typical projector and isometrically loads the surviving part into few qubits, bound its fidelity with the gentle-measurement lemma, then prove a matching converse from the eigenvalue upper bound — pinching the optimal rate to exactly the von Neumann entropy.

Step 1. Setup and encoder. The source emits $\rho^{\otimes n}$ . The encoder measures $\{\Pi_n^\delta,\,I-\Pi_n^\delta\}$ ; on the "typical" outcome it applies an isometry taking the $\dim T_n^\delta\le 2^{n(S+\delta)}$ -dimensional typical subspace into a register of $\lceil n(S+\delta)\rceil$ qubits. Why: by equation (6) the subspace fits in $n(S+\delta)$ qubits, capping the rate at $S+\delta$ before any fidelity analysis.

Step 2. Decoder and fidelity. The decoder embeds the register back through the inverse isometry. By the gentle-measurement lemma, a measurement whose favourable outcome has probability $\ge 1-\epsilon$ disturbs the state by $O(\sqrt\epsilon)$ in trace distance, so $F\big(\rho^{\otimes n},\,\mathcal D\circ\mathcal E\big)\ge 1-\epsilon'$ with $\epsilon'\to0$ . Why: projecting onto a high-probability subspace barely perturbs the state, so discarding the atypical part is almost lossless.

Step 3. Achievability. The qubits transmitted number $\log_2\dim T_n^\delta\le n(S+\delta)$ , so the rate is $\le S+\delta$ ; taking $n\to\infty$ then $\delta\to0$ achieves rate $S(\rho)$ with fidelity $\to1$ . Why: the dimension bound says how few qubits you keep; the fidelity bound says you lose almost nothing by keeping only those.

Step 4. Converse. Any scheme keeping only a rank- $2^{nR}$ projector $\Lambda$ with $R<S-\delta$ has overlap $\mathrm{Tr}(\Lambda\rho^{\otimes n})\le 2^{nR}\cdot 2^{-n(S-\delta)}+\epsilon\to0$ , by the equipartition upper bound plus the negligible atypical remainder — so fidelity cannot stay high and $R\ge S(\rho)$ is necessary. Why: equipartition caps how much weight any small subspace can capture, a lower bound on the rate.

Step 5. Conclusion. Achievability ( $R\le S+\delta$ ) and converse ( $R\ge S-\delta$ ) pinch the optimal rate to $R^\star=S(\rho)$ as $\delta\to0$ — Schumacher's noiseless coding theorem, the operational meaning of von Neumann entropy. Why: matching bounds turn an information measure into an operational resource cost. *Verified: matches Nielsen & Chuang §12.2 and Wilde Ch. 18; the counting uses only the equipartition bound (4) and the high-probability bound (5) established above.*

Derivation CQuantum Stein's lemma exponent from relative typicality

The goal is to extract the optimal error exponent for asymmetric quantum hypothesis testing between a true state $\rho$ and a null state $\sigma$ , and to show it equals the quantum relative entropy $D(\rho\|\sigma)$ . We build a test from a projector that is simultaneously high-probability for $\rho^{\otimes n}$ and small under $\sigma^{\otimes n}$ , bound the two error types, and invoke data processing for the converse. The care point: $\rho$ and $\sigma$ generally do not commute, so the joint relative-typical projector needs the Hiai–Petz and Ogawa–Nagaoka constructions, not a naive common eigenbasis.

Step 1. Test operator. Choose a projector $\Pi$ onto the joint relative-typical subspace, where simultaneously $\rho^{\otimes n}\approx 2^{-nS(\rho)}$ and $\sigma^{\otimes n}\approx 2^{-n[S(\rho)+D(\rho\|\sigma)]}$ per eigenvalue; accept $\rho$ if the outcome lands in $\Pi$ , else $\sigma$ . Why: aligning "typical for ρ" with "atypical for σ" lets one projector separate the hypotheses at the relative-entropy rate.

Step 2. Type-I error. Accepting $\sigma$ when the truth is $\rho$ has probability $\alpha_n=\mathrm{Tr}[(I-\Pi)\rho^{\otimes n}]\le\epsilon$ , since $\Pi$ is high-probability for $\rho^{\otimes n}$ . Why: the first error is the ordinary typical-subspace property applied to the true state.

Step 3. Type-II error. Accepting $\rho$ when the truth is $\sigma$ has probability $\beta_n=\mathrm{Tr}[\Pi\,\sigma^{\otimes n}]$ ; each $\sigma^{\otimes n}$ eigenvalue inside is $\le 2^{-n[S+D-\delta]}$ and $\mathrm{rank}\,\Pi\le 2^{n(S+\delta)}$ , so $\beta_n\le 2^{-n[D(\rho\|\sigma)-\delta']}$ . Why: this is equation (8) — the exponent is the gap between how σ and ρ weight the same subspace.

Step 4. Optimality. Monotonicity of quantum relative entropy under the binary measurement channel $\{\Pi,I-\Pi\}$ (data processing) shows no test beats exponent $D(\rho\|\sigma)$ at bounded type-I error, so the test is asymptotically optimal and $\lim_{n\to\infty}-\tfrac1n\log_2\beta_n^\epsilon=D(\rho\|\sigma)$ . Why: data processing turns the operational exponent into $D(\rho\|\sigma)$ , closing the loop. *Verified: Hiai–Petz (1991) achievability and Ogawa–Nagaoka (2000) strong converse; the exponent $D(\rho\|\sigma)$ matches Wilde Ch. 23 and Watrous Ch. 5.*

Limiting cases

How typicality degenerates or specialises in the corner cases:

Case	Result	Meaning
Pure source	$\rho=\vert\psi\rangle\langle\psi\vert\Rightarrow S(\rho)=0$	typical subspace is one-dimensional, spanned by $\vert\psi\rangle^{\otimes n}$ ; compression rate $0$ — a known pure state carries no quantum information
Maximally mixed	$\rho=I/d\Rightarrow S(\rho)=\log_2 d$	typical subspace fills the whole space; no compression, every eigenvalue is exactly $d^{-n}$ , so equipartition is exact at any $n$
Vanishing $\delta$ , fixed $n$	$\delta\to0^+$ before $n\to\infty$	the $1-\epsilon$ weight guarantee collapses — high probability needs $\delta$ fixed and $n\to\infty$ first; the limits do not commute
Finite block length	$\tfrac1n S_{\max}^\epsilon(\rho^{\otimes n})\to S(\rho)$	the operational rate is a smoothed max-entropy with $O(1/\sqrt n)$ corrections governed by the entropy variance $V(\rho)$
Non-commuting reference	$[\rho,\sigma]\ne0$	Stein exponent is still $D(\rho\\|\sigma)$ , but the relative-typical projector needs Hiai–Petz/Ogawa–Nagaoka, not a common eigenbasis

⚠️ The typical subspace is set by the source's eigenbasis, not by which qubits look like |0⟩

A string $x^n$ is typical because its eigenvalue probability $p(x^n)$ is near $2^{-nS}$ , not because the physical qubits are "mostly $|0\rangle$ "; the two coincide only when $\rho$ is diagonal in the computational basis. Rotate $\rho$ to $U\rho U^\dagger$ and the spectrum — hence $S(\rho)$ , hence the typical-subspace dimension — is unchanged, but the typical eigenvectors rotate with $U$ , giving a completely different subspace. Two states with the same spectrum share a typical-subspace dimension but generally not a typical subspace — typicality is a property of the source's spectral decomposition, not of any channel or measurement you later apply.

Worked examples

1 · Compressing a biased qubit source (fully worked)

A source emits the qubit $\rho=\mathrm{diag}(0.9,\,0.1)$ i.i.d. Find its entropy and variance, describe the typical subspace at $n=100$ , state the compression rate, and estimate the block length $n_0$ for failure weight $\epsilon=0.05$ at tolerance $\delta=0.1$ .

Step 1. Entropy. Since $\rho$ is diagonal, $S(\rho)=H(0.9)=-0.9\log_2 0.9-0.1\log_2 0.1=0.1368+0.3322=\mathbf{0.469\ \text{bits}}$ — the qubits-per-symbol cost.

Step 2. Entropy variance. With surprisals $y_0=-\log_2 0.9=0.1520$ , $y_1=-\log_2 0.1=3.3219$ , the mean is $\mathbb{E}[Y]=0.469$ (equal to $S$ ) and $\mathbb{E}[Y^2]=0.9(0.1520)^2+0.1(3.3219)^2=1.1243$ , so $V(\rho)=1.1243-0.469^2=\mathbf{0.904\ \text{bits}^2}$ ( $\sigma=0.951$ bits) — this controls the finite- $n$ corrections.

Step 3. Typical subspace at $n=100$ . The full space has dimension $2^{100}\approx 1.27\times10^{30}$ , but the typical subspace has dimension $\approx 2^{nS}=2^{46.9}\approx\mathbf{1.3\times10^{14}}$ — an occupied fraction $2^{-53.1}\approx 10^{-16}$ , so sixteen orders of magnitude of the space is empty.

Step 4. Compression rate. Schumacher's theorem stores the source in $S(\rho)=\mathbf{0.469}$ qubits per symbol — about $47$ qubits per block of $100$ , fidelity $\to1$ , and no scheme does better. Typical strings have $\approx 90$ zeros and $\approx 10$ ones.

Step 5. Block-length estimate. Chebyshev gives $n_0=V(\rho)/(\epsilon\delta^2)=0.904/(0.05\times 0.01)=\mathbf{1808}$ : this crude bound guarantees $\epsilon=0.05$ at $\delta=0.1$ once $n\gtrsim 1800$ , and the scaling $n_0\propto V/(\epsilon\delta^2)$ is the takeaway. S = 0.469 bits, V = 0.904 bits², typical dim ≈ 2⁴⁶·⁹, rate = 0.469 qubits/symbol, n₀ ≈ 1808.

2 · Entanglement concentration from partially entangled pairs

Alice and Bob share $n$ copies of $|\psi\rangle_{AB}=\sqrt{0.85}\,|00\rangle+\sqrt{0.15}\,|11\rangle$ . How many Bell pairs can they distill by local operations, and how does the typical subspace produce them?

Step 1. Reduced state. Tracing out Bob gives $\rho_A=\mathrm{diag}(0.85,\,0.15)$ , so the entropy of entanglement is $S(\rho_A)=H(0.85)=0.1993+0.4105=\mathbf{0.610\ \text{bits}}$ .

Step 2. Project onto the typical subspace. Alice measures the typical projector $\Pi_n^\delta$ of $\rho_A^{\otimes n}$ . With probability $\ge 1-\epsilon$ the state collapses into the typical subspace, whose Schmidt coefficients are all $\approx 2^{-nS(\rho_A)/2}$ by equipartition — a nearly maximally entangled state of Schmidt rank $\approx 2^{nS(\rho_A)}$ .

Step 3. Count the ebits. Schmidt rank $2^{m}$ is $m$ Bell pairs, so the distillable entanglement is $m=nS(\rho_A)$ . For $n=1000$ this yields $\approx\mathbf{610}$ near-perfect Bell pairs from $1000$ partially entangled ones, at LOCC-optimal rate. Distillable rate = S(ρ_A) = 0.610 ebits per copy; 1000 copies → ≈ 610 Bell pairs.

Physics of the portrait

Typical subspaces are the tool with which Benjamin Schumacher gave von Neumann entropy an operational meaning and, in the same stroke, named the fundamental unit of quantum information.

Benjamin Schumacher

coined the term 'qubit' in 1995 · quantum data compression · quantum Shannon theory

Benjamin Schumacher is the American physicist who founded quantum information theory as a quantitative subject in the mid-1990s. Building on a 1994 collaboration with Richard Jozsa, his 1995 paper "Quantum coding" did two lasting things at once. First, it proved the quantum analogue of Shannon's noiseless coding theorem: an i.i.d. source $\rho$ can be faithfully compressed to $S(\rho)$ qubits per symbol and no fewer, giving the von Neumann entropy — until then a formal quantity from statistical mechanics — a concrete operational meaning as the incompressible quantum information content of a source. The device that made the proof work is the typical subspace, the direct quantum lift of Shannon's typical set. Second, the same paper introduced the word "qubit" for the unit in which that information is measured, a coinage that immediately became universal. The method proved so flexible that it now underlies essentially every i.i.d. coding theorem in quantum Shannon theory — the Holevo–Schumacher–Westmoreland channel capacity, entanglement concentration, and quantum hypothesis testing all run on the same concentration-of-measure engine — with the later smoothed-entropy programme of Renner, Tomamichel and others recovering his asymptotic result as the quantum AEP.

⏳ History

1948

— Shannon proves the classical AEP and the noiseless coding theorem

1994

— Jozsa and Schumacher formulate faithful quantum data compression

1995

— Schumacher names the 'qubit' and proves the typical-subspace compression theorem

2000

— Ogawa and Nagaoka establish the strong converse of quantum Stein's lemma

2009

— Tomamichel, Colbeck and Renner prove the fully smooth quantum AEP

Where this lives today

Typical subspaces are the load-bearing lemma under most of modern quantum information. Every proof that an i.i.d. resource has a clean asymptotic rate — data compression, channel capacities, entanglement distillation, randomness extraction, state merging — is at bottom a typical-subspace count, and the rates those theorems produce are dimensions of typical subspaces. The frontier has moved to the finite regime: real devices process modest block lengths, so $S(\rho)$ is replaced by smoothed min- and max-entropies with $O(1/\sqrt n)$ corrections set by the entropy variance $V(\rho)$ — the same $V(\rho)=0.904$ bits² from the worked example. This one-shot framework (Renner, Tomamichel, Datta and others) makes quantum key distribution provably secure against finitely many signals and underpins the composable security proofs used in fielded QKD systems, while quantum Stein's lemma gives relative entropy the operational meaning that anchors resource theories of coherence and thermodynamics.

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §12.2 (Schumacher compression and typical subspaces) · Wilde, Quantum Information Theory, 2nd ed., Ch. 15 (Typical Subspaces) and Ch. 18 (Schumacher compression) · Preskill, Lecture Notes on Quantum Computation, Ch. 10 (quantum Shannon theory) · Watrous, The Theory of Quantum Information, Ch. 5 (quantum entropy and source coding) · Cover & Thomas, Elements of Information Theory, Ch. 3 (classical AEP and typical sets) · Tomamichel, Quantum Information Processing with Finite Resources (smoothed entropies and the quantum AEP).

Provenance. A verifier should recompute: (1) $\rho^{\otimes n}=\sum_{x^n}p(x^n)|x^n\rangle\langle x^n|$ and $S(\rho)=H(p)$ (N&C §11.3); (2) the equipartition window by exponentiating the typical-set definition; (3) the dimension sandwich $(1-\epsilon)2^{n(S-\delta)}\le\dim T_n^\delta\le 2^{n(S+\delta)}$ ; (4) the Chebyshev estimate $\Pr[\text{atypical}]\le V(\rho)/(n\delta^2)$ ; (5) Schumacher rate $=S(\rho)$ achievability and converse (N&C §12.2, Wilde Ch. 18); (6) the Stein exponent $D(\rho\|\sigma)$ and Holevo quantity $\chi=S(\bar\rho)-\sum_x p_x S(\rho_x)$ (Wilde Ch. 20, 23; Watrous Ch. 5). Worked numbers: $S(0.9)=0.469$ bits, $V=0.904$ bits², typical dim $\approx 2^{46.9}$ at $n=100$ , $n_0=1808$ at $(\epsilon,\delta)=(0.05,0.1)$ , $S(0.85)=0.610$ ebits.

Prerequisites. Shannon Entropy and Classical Information Measures · Von Neumann Entropy · Quantum Relative Entropy. Leads to. Holevo Bound · Quantum Channels and Completely Positive Trace-Preserving Maps · Data-Processing Inequality.

Ask the tutor: Ask me to walk through the Chebyshev block-length estimate, the gentle-measurement step in Schumacher's proof, or how conditional typicality assembles the HSW capacity — with the numbers filled in.

Why does the von Neumann entropy, not the Hilbert-space dimension, set the compression rate?

Because almost all of the Hilbert space is empty. The full space has dimension $d^n$ , suggesting $\log_2 d$ qubits per symbol, but the source only ever visits the typical subspace of dimension $\approx 2^{nS(\rho)}$ , and $S(\rho)\le\log_2 d$ with equality only for the maximally mixed state. The gap $\log_2 d-S(\rho)$ per symbol is redundancy you can throw away. Equipartition makes the visited part nearly flat, so it loads reversibly into $nS(\rho)$ qubits. The dimension bounds information content; the entropy is the content.

Does the typical subspace depend on the tolerance δ and the block length n?

Yes to both, by design. For fixed $n$ , enlarging $\delta$ widens the typical band, so $T_n^\delta$ captures more weight but also more dimensions; shrinking $\delta$ does the reverse. The theorem works only because you fix $\delta$ first, then send $n\to\infty$ : the captured weight rises to $1$ while the dimension's leading rate stays pinned at $S(\rho)$ , since $\pm\delta$ shifts the exponent only to $S\pm\delta$ , negligibly close to $S$ per symbol. Reverse the order — $\delta\to0$ at fixed $n$ — and the high-probability guarantee evaporates.

What is the difference between typical, conditional, and relative-entropy typicality?

They are the same concentration idea read against different distributions. Ordinary typicality asks that $\rho^{\otimes n}$ have its eigenvalue exponent near $S(\rho)$ — this drives compression. Conditional typicality fixes a classical record $x^n$ and asks the conditioned state $\rho_{x^n}=\bigotimes_i\rho_{x_i}$ to be typical, which is what channel coding needs since a codeword is a specific input string. Relative-entropy typicality compares $\rho$ to a reference $\sigma$ and lands the exponent on $D(\rho\|\sigma)$ for hypothesis testing. All three are the law of large numbers applied to a surprisal variable; only the underlying probability changes.

Is a typical string the same as a most-probable string?

No, and the distinction is the crux of the AEP. The single most probable string for $\rho=\mathrm{diag}(0.9,0.1)$ is the all-zeros string, with probability $0.9^n$ — but it is atypical: its empirical entropy per symbol is $-\log_2 0.9=0.152$ bits, far from $S=0.469$ . Typical strings have about $0.9n$ zeros and $0.1n$ ones, each with the middling probability $\approx 2^{-nS}$ ; although each is far less likely than the all-zeros string, there are so many of them ( $\approx 2^{nS}$ ) that they collectively carry almost all the weight. Typicality matches the frequency profile of the source, not maximal probability.

How does this connect to the Holevo bound and channel capacity?

The Holevo–Schumacher–Westmoreland theorem stacks two layers of typical subspaces. To send classical data through a channel $\mathcal N$ , Bob decodes with a POVM built from the typical projector of the average output $\bar\rho=\sum_x p_x\rho_x$ and the conditional typical projectors of each codeword. The average output's subspace has dimension $\approx 2^{nS(\bar\rho)}$ ; each codeword occupies $\approx 2^{n\sum_x p_x S(\rho_x)}$ . Dividing counts how many nearly orthogonal codewords fit, giving the achievable rate $\chi=S(\bar\rho)-\sum_x p_x S(\rho_x)$ — the Holevo quantity. So the Holevo bound's tightness is a typical-subspace count, one level up from compression.

What replaces S(ρ) when I only have a few copies rather than n → ∞?

Smoothed entropies. At finite $n$ the operational rates become the smoothed max-entropy $S_{\max}^\epsilon$ (compression, decoupling) and the smoothed min-entropy $S_{\min}^\epsilon$ (extractable randomness, security), each optimised over states $\epsilon$ -close to $\rho^{\otimes n}$ ; they bound the dimension of an approximate typical subspace when a hard concentration statement is unavailable. The quantum AEP is the theorem that $\tfrac1n S_{\max/\min}^\epsilon(\rho^{\otimes n})\to S(\rho)$ (as $n\to\infty$ then $\epsilon\to0$ ), with a correction of order $\sqrt{V(\rho)/n}$ . This finite-resource picture is what makes real quantum-cryptographic security proofs, which never have infinitely many signals, rigorous.