Fuchs–van de Graaf Inequalities

Quantum information theory keeps two rulers for the distance between states, and for decades it was not obvious they measured the same thing. The trace distance $D(\rho,\sigma)=\tfrac12\|\rho-\sigma\|_1$ is a statistician's ruler: by Helstrom's theorem it is the largest bias any single-shot measurement can achieve in telling $\rho$ from $\sigma$ , so a small $D$ literally means the two states are hard to distinguish. The fidelity $F(\rho,\sigma)$ is a geometer's ruler: it measures overlap, connects to purifications through Uhlmann's theorem, and behaves like the squared cosine of an angle between states. One is linear and operational; the other is quadratic and geometric. They are not equal, and in general neither determines the other exactly. The Fuchs–van de Graaf inequalities are the bridge. They sandwich the trace distance between two functions of the fidelity, $1-\sqrt{F}\le D\le\sqrt{1-F}$ , so that closeness in one measure forces closeness in the other with explicit, tight constants. This single line lets a cryptographer who has proven a fidelity bound quote a trace-distance security guarantee, and lets an experimentalist who has measured a fidelity certify a worst-case operational error. The proof is a template that recurs throughout the field: settle the pure-state case by an exact two-dimensional calculation where $D=\sqrt{1-F}$ is a Pythagorean identity, then lift to mixed states with Uhlmann's theorem and the contractivity of trace distance. Because both measures are monotone under quantum channels, the whole sandwich survives processing intact.

Key takeaways

The Fuchs–van de Graaf inequalities sandwich the trace distance $D(\rho,\sigma)$ between two functions of the fidelity $F(\rho,\sigma)$ , proving the two principal distinguishability measures are quantitatively equivalent.
In the squared-fidelity convention $F=\left(\mathrm{Tr}\sqrt{\sqrt{\rho}\,\sigma\sqrt{\rho}}\right)^2$ they read $1-\sqrt{F}\le D\le\sqrt{1-F}$ ; the root fidelity $\sqrt{F}$ is the quantity that appears symmetrically.
Both bounds are tight: the upper bound is an exact equality for every pair of pure states; the lower bound saturates at the endpoints $F\in\{0,1\}$ .
The proof combines the pure-state identity $D=\sqrt{1-F}$ with Uhlmann's theorem (fidelity as maximal purification overlap) and the contractivity of trace distance.
Both $D$ and $F$ are monotone under any CPTP channel, so the sandwich is preserved by processing — closeness certified before a channel persists after it.
They underlie composable QKD security proofs, capacity-continuity bounds, and fidelity-to-distance conversions in benchmarking, and connect to the Bures metric and Helstrom discrimination.
Key idea: Trace distance and fidelity are tied by $1-\sqrt{F}\le D\le\sqrt{1-F}$ : closeness in one measure forces closeness in the other, with explicit tight constants and equality for pure states.

Where this lives today

The Fuchs–van de Graaf sandwich is a workhorse of modern quantum information, invoked wherever one measure is proven and another is required. In quantum key distribution, security is defined by the composable trace-distance criterion — the real key must be within $\varepsilon$ in trace distance of an ideal uniform, secret key — yet entanglement-based and fidelity-of-correction arguments often deliver a fidelity bound first; the upper inequality is the standard converter, and it is why $\varepsilon$ -security budgets in QKD experiments are quoted after a square-root conversion from measured infidelities. In device benchmarking, randomized benchmarking and tomography report fidelities, while fault-tolerance thresholds are stated in diamond or trace distance; the sandwich gives the rigorous conversion that certifies a processor against a trace-distance specification from its measured fidelity. In quantum Shannon theory the inequalities feed the continuity bounds — Fannes–Audenaert and Alicki–Fannes–Winter — that let capacity and entropy estimates flow between the two pictures. They also anchor the geometry of state space through the Bures metric $d_{\mathrm{Bures}}=\sqrt{2(1-\sqrt F)}$ , whose square is the lower bound itself. A single 1999 lemma about cryptographic distinguishability now sits in the security proof of every deployed QKD system and in the acceptance test of every benchmarked quantum device.

The physical picture

Two pure states sit as unit vectors in Hilbert space, separated by an angle $\theta$ whose cosine is their overlap. Everything about how far apart they are can be read off a single right triangle: the overlap $\sqrt{F}=|\langle\psi|\phi\rangle|=\cos\theta$ is the leg along one state, and the trace distance $D=\sqrt{1-F}=\sin\theta$ is the leg perpendicular to it. The Pythagorean identity $\cos^2\theta+\sin^2\theta=1$ is exactly the pure-state relation $F+D^2=1$ . There is no slack here — for pure states the upper Fuchs–van de Graaf bound is an equality, and the first figure is that triangle drawn to scale at $\theta=45^\circ$ .

Figure

Pure states as unit vectors at angle θ: the overlap √F = cos θ is one leg and the trace distance D = sin θ is the perpendicular leg, so F + D² = 1 exactly. At θ = 45° this gives √F = 0.7071, F = 0.5000 and D = 0.7071 — the upper bound D = √(1−F) saturated

Mixing spoils the equality but never breaks the sandwich. When $\rho$ and $\sigma$ are general density operators, tracing out a purifying reference can only blur the two states together, so the trace distance drops at or below the pure-state chord — the upper bound $D\le\sqrt{1-F}$ holds with room to spare. At the same time no measurement can beat what the purified geometry allows, which pins $D$ above the floor $1-\sqrt{F}$ . The feasible region for the pair $(F,D)$ is therefore a lens-shaped band, pinched to a point at both ends: fidelity $F=0$ forces $D=1$ and $F=1$ forces $D=0$ . The second figure is that band, with the width read off at $F=0.5$ .

Figure

The feasible band for the pair (F, D). Every quantum state pair lands between the lower bound D = 1 − √F and the upper bound D = √(1−F); the upper edge is the pure-state locus. The band pinches to a point at F = 0 (forcing D = 1) and F = 1 (forcing D = 0). At F = 0.5 the width is 0.2929 ≤ D ≤ 0.7071

The core equations

The trace distance is half the trace norm of the difference of the two states — the operational ruler whose value is the maximal single-shot distinguishing bias. It is the object the lower bound protects and the upper bound caps.

D(\rho,\sigma)=\tfrac12\,\|\rho-\sigma\|_1=\tfrac12\,\mathrm{Tr}\,|\rho-\sigma| \qquad (1)

where: $\|A\|_1=\mathrm{Tr}\sqrt{A^\dagger A}$ — the trace norm, the sum of singular values; $|A|=\sqrt{A^\dagger A}$ — the positive part; $D\in[0,1]$ for normalized states, with $D=0$ iff $\rho=\sigma$ and $D=1$ iff their supports are orthogonal. Read it as "the worst-case statistical gap between the two states". The fidelity in the squared convention is the overlap-like quantity that runs from $0$ to $1$ . Its square root is what actually appears in the inequalities, which is the single most common source of confusion in the whole subject.

F(\rho,\sigma)=\left(\mathrm{Tr}\sqrt{\sqrt{\rho}\,\sigma\sqrt{\rho}}\right)^{2} \qquad (2)

where: $\sqrt{\rho}$ — the unique positive square root of the density operator; $F\in[0,1]$ , with $F=1\iff\rho=\sigma$ and $F=0$ iff the supports are orthogonal. For pure states this collapses to $F=|\langle\psi|\phi\rangle|^2$ ; the root fidelity is $\sqrt{F}=\mathrm{Tr}\sqrt{\sqrt{\rho}\,\sigma\sqrt{\rho}}$ . The inequalities themselves are the two-sided sandwich. They say a bound on either quantity immediately produces a bound on the other, and the constants are the best possible.

1-\sqrt{F(\rho,\sigma)}\;\le\;D(\rho,\sigma)\;\le\;\sqrt{1-F(\rho,\sigma)} \qquad (3)

where: $\sqrt{F}$ — the root fidelity, symmetric across both bounds. The right inequality is an equality for pure states; the left is an equality only at $F\in\{0,1\}$ . Because $\sqrt{1-x^2}\ge 1-x$ on $[0,1]$ with $x=\sqrt F$ , the two bounds are mutually consistent and pinch together at the endpoints. For any pair of pure states the upper bound is exact, an equality that turns the abstract sandwich into a clean Pythagorean statement and anchors the whole proof.

D(\rho,\sigma)=\sqrt{1-F(\rho,\sigma)}\qquad\text{for pure }\rho,\sigma \qquad (4)

where: here $F=|\langle\psi|\phi\rangle|^2$ , so $D=\sqrt{1-|\langle\psi|\phi\rangle|^2}=\sin\theta$ with $\cos\theta=|\langle\psi|\phi\rangle|$ . This is the identity drawn in the first figure and the seed of Derivation 1. The operational meaning of the trace distance — the reason the sandwich is worth having — is Helstrom's variational formula and the discrimination probability it fixes.

D(\rho,\sigma)=\max_{0\le P\le I}\mathrm{Tr}\!\left[P(\rho-\sigma)\right],\qquad p_{\text{success}}=\tfrac12+\tfrac12 D(\rho,\sigma) \qquad (5)

where: $P$ — a measurement operator (POVM element) with $0\le P\le I$ ; $p_{\text{success}}$ — the best single-shot probability of correctly identifying an equally likely $\rho$ or $\sigma$ . The optimal $P$ projects onto the positive part of $\rho-\sigma$ , so $D=\mathrm{Tr}\,Q_+$ where $\rho-\sigma=Q_+-Q_-$ . Both measures are monotone under quantum channels, which is what lets the sandwich travel through any physical process without loosening.

D(\mathcal{E}(\rho),\mathcal{E}(\sigma))\le D(\rho,\sigma),\qquad F(\mathcal{E}(\rho),\mathcal{E}(\sigma))\ge F(\rho,\sigma) \qquad (6)

where: $\mathcal{E}$ — any completely positive trace-preserving (CPTP) channel. Trace distance contracts and fidelity grows, so $\sqrt{1-F}$ and $1-\sqrt{F}$ both contract: closeness certified before the channel is certified after it. Contractivity of $D$ under the partial trace is precisely the tool that lifts the pure-state identity to mixed states.

Derivation ATrace distance of two pure states equals the fidelity chord

The goal is to prove the exact equality $D=\sqrt{1-F}$ for a pair of pure states by direct computation, establishing that the upper Fuchs–van de Graaf bound is saturated on all of pure-state space. The whole calculation lives in the at-most-two-dimensional subspace the two states span, where the difference of their projectors is a traceless Hermitian $2\times 2$ matrix whose eigenvalues can be written down by inspection. This pure-state identity is the seed from which both mixed-state bounds are grown, so it must be airtight and its endpoints must check.

Step 1. Write $\rho=|\psi\rangle\langle\psi|$ and $\sigma=|\phi\rangle\langle\phi|$ and set $f=|\langle\psi|\phi\rangle|=\sqrt F$ . The two vectors span a subspace of dimension at most two; pick an orthonormal basis $\{|0\rangle,|1\rangle\}$ for it. Why: trace distance depends only on the operator $\rho-\sigma$ , whose entire action is confined to this span, so the problem is genuinely two-dimensional no matter how large the Hilbert space is.

Step 2. Absorb phases so that $|\psi\rangle=|0\rangle$ and $|\phi\rangle=f|0\rangle+\sqrt{1-f^2}\,|1\rangle$ with $f\in[0,1]$ . Why: a global phase on $|\phi\rangle$ leaves both $\rho$ and $\sigma$ unchanged, so this parametrization is fully general and reduces the data to the single number $f$ .

Step 3. Form the difference. In this basis $\rho-\sigma=\begin{pmatrix}1-f^2 & -f\sqrt{1-f^2}\\ -f\sqrt{1-f^2} & -(1-f^2)\end{pmatrix}$ , a traceless Hermitian matrix. Why: the diagonal entries are the population differences and the off-diagonal is the coherence; tracelessness is guaranteed because both states are normalized.

Step 4. Read off the eigenvalues. For a traceless Hermitian $2\times 2$ matrix the eigenvalues are $\pm\sqrt{-\det}$ , and here $\det=-(1-f^2)$ , so $\lambda_\pm=\pm\sqrt{1-f^2}$ . Why: with zero trace the two eigenvalues are equal and opposite, and their magnitude is fixed by the determinant alone — no diagonalization is needed.

Step 5. Take half the trace norm: $D=\tfrac12(|\lambda_+|+|\lambda_-|)=\sqrt{1-f^2}=\sqrt{1-F}$ . This is an exact equality, not a bound, so the right inequality saturates for every pure pair. Why: the trace norm is the sum of the absolute eigenvalues, and halving it converts the equal-and-opposite spectrum directly into the chord $\sin\theta$ . *As a check: orthogonal states $f=0$ give $D=1$ ; identical states $f=1$ give $D=0$ ; at $f=\cos45^\circ$ , $D=\sin45^\circ=0.7071$ with $F=0.5$ , matching the first figure.*

Go deeper · Derivation B · 1 minThe upper bound from Uhlmann's theorem and contractivity

Derivation BThe upper bound from Uhlmann's theorem and contractivity

The goal is to lift the pure-state equality to arbitrary mixed states and prove $D(\rho,\sigma)\le\sqrt{1-F(\rho,\sigma)}$ . The strategy is the recurring pattern of the field: represent the fidelity as a maximal overlap of purifications through Uhlmann's theorem, apply the exact pure-state identity to those purifications, and then use the fact that discarding the purifying system is a quantum channel under which trace distance can only shrink. The inequality is therefore pure-state geometry followed by one application of monotonicity, and it is tight exactly when the inputs are already pure.

Step 1. Invoke Uhlmann's theorem: there exist purifications $|\psi\rangle,|\phi\rangle$ in an enlarged space $\mathcal H\otimes\mathcal H'$ with $|\langle\psi|\phi\rangle|^2=F(\rho,\sigma)$ , and this maximal overlap is achieved. Why: Uhlmann's theorem is precisely the statement that mixed-state fidelity is the best pure-state overlap available after purifying, which is what makes the pure-state identity applicable.

Step 2. Apply Derivation 1 to the purifications: $D(|\psi\rangle\langle\psi|,|\phi\rangle\langle\phi|)=\sqrt{1-|\langle\psi|\phi\rangle|^2}=\sqrt{1-F(\rho,\sigma)}$ . Why: the purifications are themselves pure states, so their trace distance is fixed exactly by the chord identity, with no slack.

Step 3. Recognize the partial trace as a channel: $\rho=\mathrm{Tr}_{\mathcal H'}|\psi\rangle\langle\psi|$ and $\sigma=\mathrm{Tr}_{\mathcal H'}|\phi\rangle\langle\phi|$ , and $\mathrm{Tr}_{\mathcal H'}$ is CPTP. Why: recovering the original states from their purifications is literally an act of discarding the reference, and discarding is the most physical channel there is.

Step 4. Apply contractivity of trace distance, equation (6): $D(\rho,\sigma)\le D(|\psi\rangle\langle\psi|,|\phi\rangle\langle\phi|)=\sqrt{1-F(\rho,\sigma)}$ . Why: a channel can only make two states harder to tell apart, so tracing out the reference cannot increase the distance beyond the pure-state chord.

Step 5. Conclude the upper bound holds for all $\rho,\sigma$ , with equality when both are pure because then the purifications are the states themselves. Why: the only step that introduces slack is Step 4, and it introduces none when there is nothing to trace out. *As a check: for pure inputs Steps 2 and 4 coincide and the inequality becomes the exact identity $D=\sqrt{1-F}$ ; the mixed example in Worked example 3 gives $D=0.7<\sqrt{1-F}=0.7071$ , strictly inside the bound.*

Go deeper · Derivation C · 2 minThe lower bound from the root fidelity

Derivation CThe lower bound from the root fidelity

The goal is to establish the left half of the sandwich, $1-\sqrt{F(\rho,\sigma)}\le D(\rho,\sigma)$ , so that a small trace distance guarantees a large fidelity — equivalently, that a low fidelity forces the two states far apart. (It is the upper bound, not this one, that turns a large fidelity into a small trace distance.) Purification cannot carry this half: contractivity makes $D(\rho,\sigma)$ at most the distance between purifications, which is the wrong direction for a floor. The correct route is measurement — the root fidelity is the smallest Bhattacharyya overlap any measurement can produce, the resulting classical distributions obey an elementary $\ell_1$ -versus-Bhattacharyya inequality, and the trace distance dominates the classical distance of every measurement. The floor is set by that worst-case overlap and is saturated only at the endpoints.

Step 1. Use the Fuchs–Caves variational form of the fidelity, $\sqrt{F(\rho,\sigma)}=\min_{\{M_i\}}\sum_i\sqrt{p_iq_i}$ with $p_i=\mathrm{Tr}(M_i\rho)$ and $q_i=\mathrm{Tr}(M_i\sigma)$ , and fix a POVM $\{M_i^\star\}$ that attains the minimum. Why: this is the one step that turns mixed-state fidelity into a statement about ordinary probability distributions, where an elementary inequality can bite.

Step 2. For those two distributions, $\tfrac12\sum_i|p_i-q_i|\ \ge\ \tfrac12\sum_i(\sqrt{p_i}-\sqrt{q_i})^2=1-\sum_i\sqrt{p_iq_i}$ , using $|p-q|=|\sqrt p-\sqrt q|(\sqrt p+\sqrt q)\ge(\sqrt p-\sqrt q)^2$ termwise and $\sum_ip_i=\sum_iq_i=1$ . Why: this is the whole content of the lower bound at the classical level — the $\ell_1$ distance dominates the squared Hellinger distance because the sum of two square roots is never smaller than their difference.

Step 3. Apply monotonicity in the direction that helps: a measurement is a channel, so $D(\rho,\sigma)\ge\tfrac12\sum_i|p_i-q_i|$ for the POVM of Step 1. Chaining with Step 2, $D(\rho,\sigma)\ge 1-\sum_i\sqrt{p_iq_i}=1-\sqrt{F(\rho,\sigma)}$ . Why: a floor on $D$ has to come from a measurement that is already good at distinguishing, and the fidelity-attaining one is exactly that — whereas purification runs the opposite way and can only ever yield a ceiling.

Step 4. Conclude $1-\sqrt F\le D$ for all $\rho,\sigma$ , the pure-state case included. Why: every step held for arbitrary density operators, so the floor is a property of the state pair itself and not of any particular purification.

Step 5. Check tightness. For pure states the bound reads $1-\cos\theta\le\sin\theta$ , an equality only at $\theta=0$ (identical states) and $\theta=\pi/2$ (orthogonal states); it is otherwise strict. Why: away from the endpoints the linear floor undershoots the true chord, which is exactly why the lower bound is generally looser than the upper one. *As a check: $F=0$ gives $1\le D$ , which with the upper bound forces $D=1$ ; $F=1$ gives $0\le D$ ; the overlap- $1/\sqrt2$ pure pair of Worked example 1 gives floor $0.2929$ against the true $D=0.7071$ .*

Limiting cases

The sandwich $1-\sqrt F\le D\le\sqrt{1-F}$ specializes cleanly in the regimes that matter most:

Case	Result	Meaning
Identical states	$\rho=\sigma\Rightarrow F=1,\ D=0$	Both bounds collapse to $0$ ; the inequalities are trivially saturated.
Orthogonal supports	$F=0\Rightarrow 1\le D\le 1\Rightarrow D=1$	The upper bound forces $D=1$ : perfectly distinguishable states are maximally far in both measures.
Pure states	$D=\sqrt{1-F}=\sqrt{1-\|\langle\psi\|\phi\rangle\|^2}$	The right inequality is an exact equality for every pure pair; the band's upper edge is the pure-state locus.
Nearly identical	$F=1-\epsilon\Rightarrow \tfrac{\epsilon}{2}\lesssim D\le\sqrt{\epsilon}$	Since $1-\sqrt F\approx\epsilon/2$ for small $\epsilon$ , the trace distance can be quadratically smaller than the chordal bound $\sqrt\epsilon$ .
After a channel $\mathcal E$	$D\downarrow,\ F\uparrow$ , so both bounds shrink	Processing preserves the sandwich; certified closeness survives any CPTP map.

Common trapReading a fidelity number without checking the convention

The single most common error is convention confusion. This article uses the squared fidelity $F=\left(\mathrm{Tr}\sqrt{\sqrt\rho\,\sigma\sqrt\rho}\right)^2$ , so that $F=|\langle\psi|\phi\rangle|^2$ for pure states and the inequalities read $1-\sqrt F\le D\le\sqrt{1-F}$ . Many texts — including some editions of Nielsen and Chuang — instead call $\widetilde F=\mathrm{Tr}\sqrt{\sqrt\rho\,\sigma\sqrt\rho}$ the "fidelity", which is the root fidelity here. In that convention the very same inequalities become $1-\widetilde F\le D\le\sqrt{1-\widetilde F^2}$ . The two look different but are identical statements with $\widetilde F=\sqrt F$ . Before plugging any measured number into a bound, confirm whether the quoted quantity is fidelity or root fidelity — a squared-versus-unsquared slip turns a $10^{-6}$ infidelity into a $10^{-3}$ distance error.

Worked examples

1 · Two pure states with a known overlap (fully worked)

Two pure states have overlap $|\langle\psi|\phi\rangle|=1/\sqrt2$ . Find $F$ , find $D$ , and check where the pair sits inside the sandwich.

Step 1. The squared-convention fidelity is the squared overlap: $F=|\langle\psi|\phi\rangle|^2=(1/\sqrt2)^2=\mathbf{0.5}$ .

Step 2. For pure states the upper bound is exact (equation 4), so $D=\sqrt{1-F}=\sqrt{0.5}=1/\sqrt2\approx\mathbf{0.7071}$ . Geometrically the states sit at $\theta=45^\circ$ , with $\cos\theta=\sqrt F$ and $\sin\theta=D$ — exactly the first figure.

Step 3. Check the lower bound: $1-\sqrt F=1-\sqrt{0.5}=1-0.7071=\mathbf{0.2929}$ . So the sandwich reads $0.2929\le 0.7071\le 0.7071$ : the upper bound is saturated (pure states) while the lower bound is strict and loose, undershooting the true distance by a factor of about $2.4$ .

Step 4. Convert to a discrimination probability with equation (5): $p_{\text{success}}=\tfrac12+\tfrac12 D=0.5+0.3536=\mathbf{0.8536}$ . Even these fairly overlapping states can be told apart correctly about $85\%$ of the time in a single optimal measurement.

2 · From a fidelity guarantee to a security bound

A QKD protocol certifies fidelity $F\ge 1-10^{-6}$ to an ideal state. What trace-distance security can you claim, and how loose is the complementary bound?

Step 1. Apply the upper bound: $D\le\sqrt{1-F}\le\sqrt{10^{-6}}=\mathbf{10^{-3}}$ . The protocol is within trace distance $0.001$ of ideal — a genuine composable security statement, since trace distance is the operational figure that bounds any distinguishing advantage an adversary can have.

Step 2. Apply the lower bound for contrast: $D\ge 1-\sqrt F=1-\sqrt{1-10^{-6}}\approx\mathbf{5\times10^{-7}}$ . This says $D$ is at least about $5\times10^{-7}$ , three orders of magnitude below the upper estimate.

Step 3. Read the lesson: the security-relevant direction is the upper bound, and it costs a square root — an infidelity of $10^{-6}$ only guarantees a distance of $10^{-3}$ . To certify a trace distance of $10^{-6}$ you must drive the infidelity down to $10^{-12}$ . The usable guarantee is $D\le 10^{-3}$ .

3 · A mixed pair where the upper bound is strict

Let $\rho=\mathrm{diag}(0.9,0.1)$ and $\sigma=\mathrm{diag}(0.2,0.8)$ in the same basis. Compute $D$ and $F$ and place the pair in the band.

Step 1. The states commute (both diagonal), so trace distance is the classical statistical distance: $D=\tfrac12\big(|0.9-0.2|+|0.1-0.8|\big)=\tfrac12(0.7+0.7)=\mathbf{0.7}$ .

Step 2. For commuting states the root fidelity is the Bhattacharyya overlap: $\sqrt F=\sqrt{0.9\cdot0.2}+\sqrt{0.1\cdot0.8}=\sqrt{0.18}+\sqrt{0.08}=0.4243+0.2828=0.7071$ , so $F=\mathbf{0.5000}$ .

Step 3. Evaluate the bounds: upper $\sqrt{1-F}=\sqrt{0.5}=0.7071$ , lower $1-\sqrt F=1-0.7071=0.2929$ . The sandwich reads $0.2929\le 0.7\le 0.7071$ .

Step 4. Interpret: unlike the pure pair of example 1, here $D=0.7$ sits strictly below the upper edge $0.7071$ . These states are mixed, so the pure-state equality no longer holds — the mixing slack is small but real, and it is exactly the room the upper bound leaves once states stop being pure. $D=0.7$ , strictly inside the band with $F=0.5$ .

Physics of the portrait

The inequalities carry two names because they were proven together, in a 1999 paper about the distinguishability measures a quantum cryptographer actually needs.

Origins · 1 minChristopher Fuchs and how this was discovered

Christopher Fuchs

b. 1964 · American physicist · quantum information and the foundations of QBism

Christopher A. Fuchs earned his PhD at the University of New Mexico in 1996 under Carlton Caves, writing a thesis on distinguishability and information gain in quantum theory — the subject that would make his name. The state-distance inequalities appeared in a 1999 IEEE Transactions on Information Theory paper, "Cryptographic distinguishability measures for quantum-mechanical states," written with the cryptographer Jeroen van de Graaf. Their concern was practical: quantum cryptography states its security in terms of how well an eavesdropper could distinguish real from ideal states, and different parts of the analysis naturally produce either a fidelity or a trace distance, so a rigorous two-way conversion between the two was needed. The bounds $1-\sqrt F\le D\le\sqrt{1-F}$ supplied exactly that dictionary, with tight constants, and they were quickly absorbed into the standard toolkit — appearing a year later in Nielsen and Chuang's textbook and becoming a routine step in essentially every composable-security proof since. Fuchs later became one of the founders of QBism, the interpretation that reads the quantum state as an agent's personal degree of belief; the same instinct that treats quantum states as objects of inference runs straight through the distinguishability inequalities that bear his and van de Graaf's names.

History

1969

— Helstrom formulates optimal quantum state discrimination and the trace-distance bound

1976

— Uhlmann defines the transition probability, the mixed-state fidelity

1994

— Jozsa recasts fidelity in the modern density-operator form

1999

— Fuchs and van de Graaf prove the inequalities linking fidelity and trace distance

2000

— Nielsen and Chuang give the textbook proof and popularize the bounds

Going further

Texts. Fuchs & van de Graaf, Cryptographic distinguishability measures for quantum-mechanical states, IEEE Trans. Inf. Theory 45, 1216 (1999) · Nielsen & Chuang, Quantum Computation and Quantum Information, §9.2 · Watrous, The Theory of Quantum Information, Ch. 3 · Wilde, Quantum Information Theory, 2nd ed., Ch. 9 · Preskill, Lecture Notes for Physics 219, Ch. 10 · Kitaev, Shen & Vyalyi, Classical and Quantum Computation, §11. The key results to keep in mind: the $2\times2$ eigenvalue computation gives $D=\sqrt{1-f^2}$ for pure states; the elementary inequality $\sqrt{1-x^2}\ge 1-x$ on $[0,1]$ underpins the lower bound; overlap $1/\sqrt2$ gives $F=0.5$ , $D=0.7071$ , $p_{\text{success}}=0.8536$ ; $F\ge1-10^{-6}$ gives $D\le10^{-3}$ ; and the mixed pair $\mathrm{diag}(0.9,0.1)$ versus $\mathrm{diag}(0.2,0.8)$ gives $D=0.7$ , $F=0.5$ , strictly inside the band. The squared-fidelity convention $F=\left(\mathrm{Tr}\sqrt{\sqrt\rho\,\sigma\sqrt\rho}\right)^2$ is used throughout, as distinct from the root-fidelity alternative.

Prerequisites. Trace Distance · Fidelity and Uhlmann's Theorem · Data-Processing Inequality · Quantum Channels and CPTP Maps. Leads to. Holevo Bound · Stinespring Dilation Theorem · Quantum Relative Entropy.

Ask the tutor: Ask me to prove either bound line by line, or to compute $D$ and $F$ and place any qubit pair you name inside the sandwich.

If trace distance and fidelity bound each other, why keep both measures?

Because they answer different questions and are easy to compute in different situations. Trace distance is the operational one: by Helstrom's theorem it is the maximal single-shot distinguishing bias, so it is the natural language for security and worst-case error. Fidelity is the geometric one: it connects to purifications through Uhlmann's theorem, is multiplicative over tensor products, and is often far easier to compute or measure — average gate fidelity, for instance, is what randomized benchmarking returns directly. The Fuchs–van de Graaf inequalities let you work in whichever measure is convenient for a given step and then convert, with tight constants, to whichever measure the final statement demands. Keeping both is not redundancy; it is having two coordinate systems for one space.

Which bound do I use, and when is each tight?

Use the upper bound $D\le\sqrt{1-F}$ when you have a fidelity and need to guarantee small trace distance — this is the security direction, and it is an exact equality for pure states, so it is the tighter and more useful of the two. Use the lower bound $1-\sqrt F\le D$ when you have a trace distance and want to bound the residual fidelity, or to argue that two states are not too close. The lower bound is generally loose: it is an equality only at the endpoints $F\in\{0,1\}$ and can undershoot the true distance by more than a factor of two in between, as Worked example 1 shows ( $0.2929$ against $0.7071$ ). When both states are pure the two bounds do not coincide, but the upper one becomes exact.

Why does a square root appear in the bounds at all?

It is the geometric signature of working on a sphere of states. For pure states the overlap is $\sqrt F=\cos\theta$ and the trace distance is the chordal separation $D=\sin\theta$ , so the Pythagorean identity $\cos^2\theta+\sin^2\theta=1$ becomes $F+D^2=1$ , i.e. $D=\sqrt{1-F}$ . The square root is nothing more exotic than the relationship between a chord and a cosine on a circle. Fidelity, being an overlap, is quadratic in the states while trace distance is linear, so converting between them necessarily crosses one power of a square root. The mixed-state inequalities are just how much slack mixing introduces around this exact pure-state relation — and the slack always shrinks $D$ below the chord, never above it.

Does the sandwich still hold after a noisy channel?

Yes, and that is one of its most useful features. Both measures are monotone under any CPTP map: trace distance cannot increase and fidelity cannot decrease when the same channel is applied to both states. Consequently both $\sqrt{1-F}$ and $1-\sqrt F$ can only shrink under processing, so the inequality $1-\sqrt F\le D\le\sqrt{1-F}$ is preserved and any closeness you certified before the channel remains certified afterward. This is exactly what makes the sandwich safe to use in multi-stage protocols and capacity-continuity arguments, where states pass through encoders, noisy links, and decoders — the two rulers stay locked together at every stage.

How does the lower bound relate to the Bures metric?

Directly: the Bures distance is $d_{\mathrm{Bures}}(\rho,\sigma)=\sqrt{2(1-\sqrt F)}$ , so its square is $2(1-\sqrt F)$ , twice the Fuchs–van de Graaf lower bound. The Bures metric is the Riemannian, infinitesimally-defined distance on state space whose geodesics and curvature encode quantum estimation limits through the quantum Fisher information. The lower inequality $1-\sqrt F\le D$ therefore reads $\tfrac12 d_{\mathrm{Bures}}^2\le D$ , tying the smooth geometric distance to the operational one. Near coincident states, where $F=1-\epsilon$ , the lower bound $1-\sqrt F\approx\epsilon/2$ scales linearly in the infidelity while the upper bound $\sqrt\epsilon$ scales as its square root.

What exactly makes trace distance the "operational" measure?

Helstrom's theorem. Suppose you are handed a single copy of a state that is equally likely to be $\rho$ or $\sigma$ and must guess which, using any measurement allowed by quantum mechanics. The best possible probability of guessing right is $p_{\text{success}}=\tfrac12+\tfrac12 D(\rho,\sigma)$ , achieved by measuring the sign of $\rho-\sigma$ . So $D=0$ means no better than a coin flip and $D=1$ means you always win. This is why security definitions are written in trace distance: $D\le\varepsilon$ literally caps any adversary's distinguishing advantage at $\varepsilon$ , and the Fuchs–van de Graaf upper bound is what converts a fidelity into precisely that guarantee.

Are these the only inequalities between the two measures?

They are the standard tight pair, but the story continues. The upper bound $D\le\sqrt{1-F}$ is exactly tight (an equality for pure states) and the lower bound $1-\sqrt F\le D$ is tight at the endpoints, so together they bracket state space as tightly as two single-variable bounds can. Sharper relations hold in special cases — when one state is pure, or when both are qubits. There are also parallel inequalities to other quantities: quantum Pinsker ties trace distance to relative entropy, and the Fannes–Audenaert bound ties it to entropy differences. The Fuchs–van de Graaf pair remains the first tool reached for whenever the two principal distance measures must be compared.