Fidelity and the Uhlmann Theorem

How close are two quantum states? The question sounds elementary, but it sits underneath almost everything quantitative in quantum information: certifying that a lab prepared the state it intended, proving that a noisy channel can still carry a qubit faithfully, and bounding what an eavesdropper can learn. Two states can be compared in two complementary ways. The trace distance answers an operational question — how well can a single measurement tell them apart? — and it is a genuine metric. The fidelity answers the dual question — how much do they overlap, and could they secretly be two faces of the same pure state? It is not itself a metric, yet it generates one, multiplies over tensor products, and behaves beautifully under processing. The defining formula $F(\rho,\sigma)=\big(\mathrm{Tr}\,|\sqrt{\rho}\sqrt{\sigma}|\big)^2$ looks forbidding, a nested square root of operators. Its meaning is unlocked by a single idea — purification. Every mixed state on a system $A$ is the shadow, under partial trace, of a pure state on a larger system $AR$ , and Uhlmann's theorem says $\sqrt{F}$ is exactly the largest overlap $|\langle\psi|\phi\rangle|$ achievable between a purification of $\rho$ and a purification of $\sigma$ sharing the reference $R$ . An awkward operator expression becomes a transparent geometric one: fidelity is the best alignment you can arrange by rotating the unobserved environment. From that variational picture nearly every property of fidelity falls out for free — symmetry, the bound $0\le F\le1$ , multiplicativity, joint concavity, and monotonicity under channels — each reducing to facts about inner products plus the freedom in the environment. Uhlmann's theorem is not merely a formula; it is the structural spine of the whole theory of quantum distinguishability.

Key takeaways

The fidelity $F(\rho,\sigma)=\big(\mathrm{Tr}\sqrt{\sqrt{\rho}\,\sigma\sqrt{\rho}}\big)^2$ scores the closeness of two states: it equals $1$ exactly when $\rho=\sigma$ and $0$ only when their supports are orthogonal.
Uhlmann's theorem recasts it variationally, $\sqrt{F(\rho,\sigma)}=\max_{|\psi\rangle,|\phi\rangle}|\langle\psi|\phi\rangle|$ , turning an operator square root into the maximal overlap of purifications.
For pure states $F=|\langle\psi|\phi\rangle|^2$ , for one-pure-one-mixed $F=\langle\psi|\sigma|\psi\rangle$ , and for commuting (classical) states $F=\big(\sum_i\sqrt{p_iq_i}\big)^2$ , the squared Bhattacharyya coefficient.
Fidelity is symmetric, multiplicative over tensor products, jointly concave, and monotone under any channel: $F(\mathcal{N}\rho,\mathcal{N}\sigma)\ge F(\rho,\sigma)$ — processing can only increase overlap.
The Fuchs–van de Graaf inequalities $1-\sqrt{F}\le\tfrac12\|\rho-\sigma\|_1\le\sqrt{1-F}$ sandwich fidelity and trace distance, making the two operationally interchangeable.
Fidelity generates the Bures metric and Bures angle $\arccos\sqrt{F}$ , the natural Riemannian geometry on state space whose curvature is the quantum Fisher information.
Key idea: fidelity is the maximal overlap between purifications, $\sqrt{F(\rho,\sigma)}=\max|\langle\psi|\phi\rangle|$ . This Uhlmann form turns operator square roots into geometry, and almost every property of fidelity follows from properties of inner products plus the freedom to rotate the purifying environment.

The core insight

Fidelity is the maximal overlap between purifications: $\sqrt{F(\rho,\sigma)}=\max|\langle\psi|\phi\rangle|$ . This Uhlmann form turns operator square roots into geometry, and almost every property of fidelity follows from properties of inner products plus the freedom to choose the purifying environment.

Where this lives today

Fidelity is the number quoted whenever a quantum device is characterized. On today's superconducting and trapped-ion processors, single-qubit gate fidelities exceed $99.9\%$ and the best two-qubit gates now cross the $99.9\%$ mark, figures obtained not by full tomography but by randomized benchmarking, which extracts the average gate fidelity from the decay rate of random Clifford sequences. Direct fidelity estimation and cross-entropy benchmarking — the metric Google used to claim quantum advantage with the 53-qubit Sycamore chip in 2019 — are both fidelity-like scores that certify a many-qubit output without exponential tomographic cost. In quantum error correction, the entanglement fidelity of the logical qubit must approach one as the code distance grows, and the surface-code threshold near a $1\%$ physical error rate is stated as a fidelity condition. In composable security proofs, the Fuchs–van de Graaf inequalities convert a fidelity guarantee on the shared state into a bound on any eavesdropper's distinguishing advantage. And in quantum metrology, the Bures metric derived from fidelity fixes the quantum Fisher information and hence the ultimate precision of atomic clocks and gravitational-wave interferometers — the same $F$ , read as a geometry rather than a score.

The physical picture

Picture a mixed state as a pure state that has been hidden. The full state $|\psi\rangle_{AR}$ lives on the system $A$ together with a reference $R$ you cannot touch; tracing out $R$ leaves the mixed $\rho$ you hold, and the lost information is not gone but entangled into the environment. Two mixed states $\rho$ and $\sigma$ may look quite different on $A$ alone, yet their hidden halves can be rotated to line up. Uhlmann's theorem says the fidelity is precisely how well you can make them line up: the best overlap of any two purifications. The knob you turn is a unitary on the reference $R$ , and optimizing over it is the variational maximum. The first figure makes this concrete — two purifications separated by the Bures angle, and the reference rotation that closes it.

Figure

Purification and Uhlmann's theorem: ρ and σ lift to pure states in the enlarged space A⊗R, and rotating the reference R by a unitary U closes the angle θ between them until the purification overlap reaches its maximum cos θ = √F, so the Bures angle is θ = arccos √F

Fidelity is a similarity score running from $0$ (orthogonal supports, perfectly distinguishable) up to $1$ (identical). Unlike the trace distance, which grows with separation, fidelity is large when states are close — it behaves like $\cos^2$ of an angle. That angle, the Bures angle $\arccos\sqrt{F}$ , is a genuine metric, and its curvature near coincidence is the quantum Fisher information that bounds all of metrology. Replacing the density operators by commuting classical distributions collapses fidelity to the Bhattacharyya overlap $\big(\sum_i\sqrt{p_iq_i}\big)^2$ , so it is the natural noncommutative lift of a familiar statistical overlap. The second figure works one concrete pair of qubits all the way to numbers.

Figure

A worked qubit pair on the Bloch disk: ρ has Bloch vector of length 0.6 along z, σ length 0.8 along x. Jozsa's single-qubit formula gives F = 0.74 so √F = 0.860, while the trace distance T = 0.500 sits right against the Fuchs–van de Graaf upper bound 0.140 ≤ 0.500 ≤ 0.510

The core equations

The starting point is the definition, and its first surprise is that the asymmetric-looking expression is secretly symmetric. Reading it as the squared trace norm of the single operator $\sqrt{\rho}\sqrt{\sigma}$ makes both the symmetry and the range $[0,1]$ transparent.

F(\rho,\sigma)=\left(\mathrm{Tr}\,\sqrt{\sqrt{\rho}\,\sigma\,\sqrt{\rho}}\right)^2=\big\|\sqrt{\rho}\sqrt{\sigma}\big\|_1^{\,2} \qquad (1)

where:

symbol	meaning
$\rho,\sigma$	density operators on a Hilbert space $\mathcal{H}$ (positive, unit trace)
$\sqrt{\rho}$	the unique positive operator square root of $\rho$
$\\|X\\|_1$	trace norm $\mathrm{Tr}\sqrt{X^\dagger X}$ , the sum of singular values
$F$	fidelity, valued in $[0,1]$ ; the "squared" convention, so $F=\|\langle\psi\|\phi\rangle\|^2$ for pure states

Reading it: $F$ is the squared trace norm of $\sqrt{\rho}\sqrt{\sigma}$ . Since $X$ and $X^\dagger$ share singular values, and $(\sqrt{\rho}\sqrt{\sigma})^\dagger=\sqrt{\sigma}\sqrt{\rho}$ , the value is unchanged under $\rho\leftrightarrow\sigma$ — the definition is symmetric despite appearances. The whole theory pivots on Uhlmann's theorem, which trades the operator square root for a geometric overlap. It says root-fidelity is the best inner product achievable between purifications, optimized over the unitary freedom in the reference system.

\sqrt{F(\rho,\sigma)}=\max_{|\psi\rangle,\,|\phi\rangle}\big|\langle\psi|\phi\rangle\big| \qquad (2)

where:

symbol	meaning
$\|\psi\rangle_{AR}$	any purification of $\rho$ , so $\mathrm{Tr}_R\|\psi\rangle\langle\psi\|=\rho$
$\|\phi\rangle_{AR}$	any purification of $\sigma$ on the same total space $\mathcal{H}_A\otimes\mathcal{H}_R$
$\max$	over all purifications; equivalently over unitaries on the reference $R$

Reading it: because all purifications of a fixed state differ only by a unitary on $R$ , it is enough to hold one purification fixed and rotate the other. The maximization is over a single unitary, and it is always attained. Two special cases are the ones a physicist reaches for daily. When one argument is pure, the operator square root collapses and the fidelity becomes a plain expectation value — the probability that $\sigma$ passes a yes/no test for $|\psi\rangle$ .

F(|\psi\rangle\langle\psi|,\sigma)=\langle\psi|\sigma|\psi\rangle, \qquad F(|\psi\rangle\langle\psi|,|\phi\rangle\langle\phi|)=|\langle\psi|\phi\rangle|^2 \qquad (3)

where: $|\psi\rangle$ is a target pure state and $\langle\psi|\sigma|\psi\rangle$ is the acceptance probability of the projective test $\{|\psi\rangle\langle\psi|,\,\mathbb{1}-|\psi\rangle\langle\psi|\}$ . This is the single most-used formula in experimental benchmarking. Fidelity multiplies over independent systems and, crucially, never decreases under processing. Multiplicativity makes it well behaved for many copies; monotonicity — the data-processing inequality — makes it a resource monotone.

F(\rho_1\otimes\rho_2,\;\sigma_1\otimes\sigma_2)=F(\rho_1,\sigma_1)\,F(\rho_2,\sigma_2), \qquad F\big(\mathcal{N}(\rho),\mathcal{N}(\sigma)\big)\ge F(\rho,\sigma) \qquad (4)

where: $\otimes$ is the tensor product of independent systems and $\mathcal{N}$ is any CPTP map (quantum channel). A channel can only make two states harder to tell apart, hence more overlapping — fidelity goes up, never down. Fidelity and trace distance are two windows on the same closeness, and the Fuchs–van de Graaf inequalities lock them together. A guarantee on one is automatically a guarantee on the other.

1-\sqrt{F(\rho,\sigma)}\;\le\;\tfrac12\|\rho-\sigma\|_1\;\le\;\sqrt{1-F(\rho,\sigma)} \qquad (5)

where: $\tfrac12\|\rho-\sigma\|_1=T(\rho,\sigma)$ is the trace distance, the maximal distinguishing bias in one measurement, and $\sqrt{F}$ is the root fidelity. For pure states the right inequality becomes an equality. A fidelity bound $F\ge1-\epsilon$ caps the trace distance at $\sqrt{\epsilon}$ . Finally, fidelity has a metric geometry: it generates the Bures distance, a true metric on density operators, and the Bures angle, which obeys the triangle inequality and reduces to the Fubini–Study angle on pure states.

D_B(\rho,\sigma)=\sqrt{2\big(1-\sqrt{F}\big)}, \qquad \theta_B(\rho,\sigma)=\arccos\sqrt{F(\rho,\sigma)} \qquad (6)

where: $D_B$ is the Bures distance and $\theta_B$ the Bures angle. Expanding $F$ to second order along a smooth family $\rho_\theta$ gives $D_B^2=\tfrac14\,\mathcal{F}_Q\,\mathrm{d}\theta^2$ , with $\mathcal{F}_Q$ the quantum Fisher information — the geometric origin of the quantum Cramér–Rao bound.

Derivation AUhlmann's theorem from the polar decomposition

The goal is to prove that the maximal overlap of purifications equals root-fidelity, $\sqrt{F(\rho,\sigma)}=\max_{|\phi\rangle}|\langle\psi|\phi\rangle|$ , by fixing one purification and constructing the optimal partner explicitly. The engine is one clean identity: for any operator $X$ , the maximum of $|\mathrm{Tr}(XU)|$ over unitaries $U$ equals the trace norm $\|X\|_1$ , with the optimizer read off from the polar decomposition of $X$ . We work on $\mathcal{H}_A\otimes\mathcal{H}_R$ with $\dim\mathcal{H}_R=\dim\mathcal{H}_A$ and write purifications as vectorized operators, so that inner products become traces.

Step 1. Vectorize: to an operator $A$ on $\mathcal{H}_A$ associate the vector $|A\rangle\!\rangle=\sum_{ij}A_{ij}|i\rangle_A|j\rangle_R$ , which satisfies $\mathrm{Tr}_R|A\rangle\!\rangle\langle\!\langle A|=AA^\dagger$ . Hence $|\sqrt{\rho}\rangle\!\rangle$ is a purification of $\rho$ because $\sqrt{\rho}\sqrt{\rho}^\dagger=\rho$ , and the general purification of $\rho$ is $|\sqrt{\rho}\,U\rangle\!\rangle$ for a unitary $U$ . Why: vectorization turns the abstract "choose a purification" into the concrete "choose a unitary $U$ ", which is what we will optimize over.

Step 2. Compute the overlap. The vectorized inner product is $\langle\!\langle A|B\rangle\!\rangle=\mathrm{Tr}(A^\dagger B)$ , so taking the fixed purification $|\psi\rangle=|\sqrt{\rho}\rangle\!\rangle$ of $\rho$ and the free purification $|\phi\rangle=|\sqrt{\sigma}\,U\rangle\!\rangle$ of $\sigma$ gives $\langle\psi|\phi\rangle=\mathrm{Tr}(\sqrt{\rho}\,\sqrt{\sigma}\,U)$ . Why: the whole overlap is now a single trace containing one adjustable unitary — the problem is reduced to a standard optimization.

Step 3. Maximize over the reference unitary using $\max_U|\mathrm{Tr}(XU)|=\|X\|_1$ , attained at $U=W^\dagger$ from the polar decomposition $X=|X^\dagger|\,W$ . With $X=\sqrt{\rho}\sqrt{\sigma}$ this yields $\max_U|\mathrm{Tr}(\sqrt{\rho}\sqrt{\sigma}\,U)|=\|\sqrt{\rho}\sqrt{\sigma}\|_1$ . Why: the trace-norm identity is exactly the statement that aligning the phases of every singular value — which the reference unitary is free to do — sums their magnitudes.

Step 4. Identify with the definition: $\|\sqrt{\rho}\sqrt{\sigma}\|_1=\mathrm{Tr}\sqrt{\sqrt{\rho}\,\sigma\sqrt{\rho}}=\sqrt{F(\rho,\sigma)}$ . The maximal purification overlap therefore equals $\sqrt{F}$ , which is Uhlmann's theorem. Why: this closes the loop — the geometric quantity we maximized is literally the trace-norm form of the fidelity we started from. *Verified: each step uses a standard identity (vectorization, $\langle\!\langle A|B\rangle\!\rangle=\mathrm{Tr}(A^\dagger B)$ , $\max_U|\mathrm{Tr}XU|=\|X\|_1$ , polar decomposition); cross-check against Watrous Ch. 3 and Wilde Ch. 9.*

Go deeper · Derivation B · 1 minMonotonicity under channels from Stinespring dilation

Derivation BMonotonicity under channels from Stinespring dilation

The goal is to show that no quantum channel can decrease fidelity, $F(\mathcal{N}(\rho),\mathcal{N}(\sigma))\ge F(\rho,\sigma)$ , and to see why the inequality runs that way. The proof is almost pure structure: it combines Uhlmann's variational form with the Stinespring dilation, which realizes any channel as an isometry into a larger space followed by discarding an environment. Isometries preserve inner products exactly, and discarding a subsystem only enlarges the set of available purifications — so the Uhlmann maximum can only grow.

Step 1. Dilate the channel. By Stinespring, write $\mathcal{N}(\cdot)=\mathrm{Tr}_E\big(V\,\cdot\,V^\dagger\big)$ for an isometry $V:\mathcal{H}_A\to\mathcal{H}_B\otimes\mathcal{H}_E$ into the output tensored with an environment $E$ . Why: this replaces the abstract CPTP map by two elementary moves — an inner-product-preserving embedding and a partial trace — each of which we can track through Uhlmann's theorem.

Step 2. Push purifications through $V$ . Let $|\psi\rangle,|\phi\rangle$ be Uhlmann-optimal purifications of $\rho,\sigma$ with $\langle\psi|\phi\rangle=\sqrt{F(\rho,\sigma)}$ . Then $V|\psi\rangle$ purifies $\mathcal{N}(\rho)$ and $V|\phi\rangle$ purifies $\mathcal{N}(\sigma)$ , and because $V^\dagger V=\mathbb{1}$ the overlap is unchanged: $\langle\psi|V^\dagger V|\phi\rangle=\langle\psi|\phi\rangle=\sqrt{F(\rho,\sigma)}$ . Why: an isometry rigidly preserves geometry, so it carries the optimal alignment of the inputs into a valid alignment of the outputs with no loss of overlap.

Step 3. Maximize over the larger set. The states $V|\psi\rangle,V|\phi\rangle$ are just two of the purifications of $\mathcal{N}(\rho),\mathcal{N}(\sigma)$ ; the full set includes every reference rotation, so the Uhlmann maximum over it is at least the overlap we already have: $\sqrt{F(\mathcal{N}\rho,\mathcal{N}\sigma)}\ge\sqrt{F(\rho,\sigma)}$ . Why: adding admissible candidates to a maximization can only raise the optimum — this monotonic-enlargement of the purification set is the entire content of data processing for fidelity. *Verified: the argument needs only $V^\dagger V=\mathbb{1}$ and Uhlmann's theorem; it reproduces the standard data-processing inequality (Nielsen & Chuang §9.2.3, Wilde Ch. 9).*

Go deeper · Derivation C · 1 minFidelity and trace distance the Fuchs–van de Graaf upper bound

Derivation CFidelity and trace distance the Fuchs–van de Graaf upper bound

The goal is to derive the more frequently used half of the Fuchs–van de Graaf sandwich, $\tfrac12\|\rho-\sigma\|_1\le\sqrt{1-F(\rho,\sigma)}$ , converting a fidelity guarantee into a distinguishability guarantee. The strategy is to lift both states to Uhlmann-optimal purifications, use the exact pure-state trace-distance identity there, then contract back down with the partial trace, which can only shrink trace distance. Because the pure-state relation is an equality and the contraction an inequality, the two combine into the stated bound.

Step 1. Choose Uhlmann-optimal purifications $|\psi\rangle,|\phi\rangle$ of $\rho,\sigma$ , and fix their global phase so that $\langle\psi|\phi\rangle=\sqrt{F(\rho,\sigma)}\ge0$ is real and nonnegative. Why: Uhlmann guarantees such purifications exist with overlap exactly $\sqrt{F}$ , giving us pure states whose distance we can compute in closed form.

Step 2. Use the exact pure-state identity $\tfrac12\big\||\psi\rangle\langle\psi|-|\phi\rangle\langle\phi|\big\|_1=\sqrt{1-|\langle\psi|\phi\rangle|^2}$ , a two-dimensional computation in the span of $|\psi\rangle,|\phi\rangle$ . Why: for two pure states the operator $|\psi\rangle\langle\psi|-|\phi\rangle\langle\phi|$ has eigenvalues $\pm\sqrt{1-|\langle\psi|\phi\rangle|^2}$ , so its trace norm is known exactly — no inequality yet.

Step 3. Contract with the partial trace. Since $\mathrm{Tr}_R$ is CPTP and trace distance is monotone under channels, $\tfrac12\|\rho-\sigma\|_1=\tfrac12\|\mathrm{Tr}_R(|\psi\rangle\langle\psi|-|\phi\rangle\langle\phi|)\|_1\le\tfrac12\||\psi\rangle\langle\psi|-|\phi\rangle\langle\phi|\|_1$ . Why: discarding the reference cannot make states easier to distinguish, so the mixed-state distance is bounded by the pure-state distance we just evaluated.

Step 4. Combine Steps 2 and 3: $\tfrac12\|\rho-\sigma\|_1\le\sqrt{1-|\langle\psi|\phi\rangle|^2}=\sqrt{1-F(\rho,\sigma)}$ . The companion lower bound $1-\sqrt{F}\le T$ follows dually from the measurement (minimization) form of the fidelity. Why: substituting the Uhlmann overlap $|\langle\psi|\phi\rangle|=\sqrt{F}$ delivers the bound in terms of $F$ alone, which is the form used in security and error-correction proofs. *Verified: reproduces Fuchs & van de Graaf (1999); the pure-state identity and trace-distance monotonicity are textbook (Nielsen & Chuang §9.2.3).*

Limiting cases

Fidelity simplifies cleanly wherever the structure of the states does. Each row below is a sanity check the general formula must reproduce.

Case	Result	Meaning
Both states pure	$F=\vert\langle\psi\vert\phi\rangle\vert^2$	the operator square root collapses to the squared overlap; the Uhlmann maximization is trivial because pure states have essentially unique purifications up to phase
One state pure	$F(\vert\psi\rangle\langle\psi\vert,\sigma)=\langle\psi\vert\sigma\vert\psi\rangle$	fidelity is the expectation of $\sigma$ in $\vert\psi\rangle$ — the pass probability of a projective test, the workhorse of benchmarking
Commuting (classical)	$F=\big(\sum_i\sqrt{p_iq_i}\big)^2$	simultaneously diagonalizable states give the squared Bhattacharyya coefficient; the Bures angle becomes the classical Hellinger distance
Orthogonal supports	$\mathrm{supp}\,\rho\perp\mathrm{supp}\,\sigma\Rightarrow F=0$	perfectly distinguishable states, trace distance one; this is the only way to reach $F=0$
Near coincidence	$F(\rho,\rho+\mathrm{d}\rho)=1-\tfrac14\,\mathrm{d}s_B^2+\cdots$	the second-order expansion defines the Bures metric, whose coefficient is the quantum Fisher information bounding parameter estimation

Common trapTwo conventions: F versus √F

The literature is split on what to call "fidelity." Nielsen & Chuang define $F=\mathrm{Tr}\sqrt{\sqrt{\rho}\sigma\sqrt{\rho}}$ (no square), so their pure-state fidelity is $|\langle\psi|\phi\rangle|$ ; Wilde, Watrous, and much of the quantum-Shannon literature use the squared version $F=|\langle\psi|\phi\rangle|^2$ adopted here. Both agree that $F=1$ iff $\rho=\sigma$ , but every quantitative formula mixing them differs by a square. Before quoting a Fuchs–van de Graaf bound, a Bures-angle formula, or a threshold like " $F\ge0.99$ ", check which convention the source uses — $0.99$ and $0.9801=(0.99)^2$ describe genuinely different states. A silent convention mismatch is the single most common error in reading the fidelity literature.

Worked examples

1 · Two qubits on the Bloch disk (fully worked)

Let $\rho$ have Bloch vector $\vec r=(0,0,0.6)$ and $\sigma$ have $\vec s=(0.8,0,0)$ . Compute the fidelity, the trace distance, and check the Fuchs–van de Graaf inequalities.

Step 1. For a single qubit, Jozsa's closed form (squared convention) is $F=\mathrm{Tr}(\rho\sigma)+2\sqrt{\det\rho\,\det\sigma}$ , with $\mathrm{Tr}(\rho\sigma)=\tfrac12(1+\vec r\cdot\vec s)$ and $\det\rho=\tfrac14(1-|\vec r|^2)$ . Here $\vec r\cdot\vec s=0$ (orthogonal Bloch vectors), so $\mathrm{Tr}(\rho\sigma)=\tfrac12$ .

Step 2. The determinant term: $\det\rho=\tfrac14(1-0.36)=0.16$ and $\det\sigma=\tfrac14(1-0.64)=0.09$ , so $2\sqrt{0.16\cdot0.09}=2\sqrt{0.0144}=2(0.12)=0.24$ . Adding, $F=0.50+0.24=\mathbf{0.74}$ , hence $\sqrt{F}=\mathbf{0.860}$ .

Step 3. Trace distance for qubits is half the Euclidean distance of Bloch vectors: $T=\tfrac12|\vec r-\vec s|$ . Here $\vec r-\vec s=(-0.8,0,0.6)$ , so $|\vec r-\vec s|=\sqrt{0.64+0.36}=1.0$ and $T=\mathbf{0.500}$ .

Step 4. Check Fuchs–van de Graaf: $1-\sqrt{F}=1-0.860=0.140$ and $\sqrt{1-F}=\sqrt{0.26}=0.510$ . The bracket reads $0.140\le0.500\le0.510$ — satisfied, and notably the upper bound is nearly tight, which is typical when one state is close to pure (here $|\vec s|=0.8$ ). Result: $F=0.74$ , $\sqrt{F}=0.860$ , $T=0.500$ , and $0.140\le0.500\le0.510$ .

2 · A pure target versus a depolarized copy

A device is meant to output $|\psi\rangle$ but instead produces $\sigma=(1-p)\,|\psi\rangle\langle\psi|+p\,\tfrac{\mathbb{1}}{2}$ , a depolarized qubit with $p=0.2$ . What is the fidelity to the target?

Step 1. One argument is pure, so use $F=\langle\psi|\sigma|\psi\rangle$ directly — no operator square root is needed.

Step 2. Evaluate term by term: $\langle\psi|(1-p)|\psi\rangle\langle\psi|\,|\psi\rangle=(1-p)$ , and $\langle\psi|\,p\tfrac{\mathbb{1}}{2}\,|\psi\rangle=\tfrac{p}{2}$ . Summing, $F=(1-p)+\tfrac{p}{2}=1-\tfrac{p}{2}$ .

Step 3. With $p=0.2$ , $F=1-0.1=\mathbf{0.90}$ . The Fuchs–van de Graaf lower bound then guarantees the trace distance is at least $1-\sqrt{0.90}=1-0.9487=\mathbf{0.051}$ and at most $\sqrt{0.10}=\mathbf{0.316}$ . Result: a 20% depolarizing admixture costs exactly half of $p$ in fidelity, $F=0.90$ .

3 · A classical (commuting) pair

Two commuting states share an eigenbasis with probability vectors $p=(0.5,0.5)$ and $q=(0.9,0.1)$ . Find the fidelity.

Step 1. Commuting states reduce fidelity to the squared Bhattacharyya coefficient, $F=\big(\sum_i\sqrt{p_iq_i}\big)^2$ . The overlap is $\sqrt{0.45}+\sqrt{0.05}=0.6708+0.2236=0.8944$ .

Step 2. Square it: $F=(0.8944)^2=\mathbf{0.80}$ (cleanly, $0.45+0.05+2\sqrt{0.0225}=0.5+0.30=0.8$ ), and the Bures angle is $\arccos0.894=26.6^\circ$ . Result: $F=0.80$ , matching the classical statistical-overlap intuition exactly.

Physics of the portrait

Fidelity is one of those quantities whose deep meaning was supplied not by the physicists who use it but by a mathematical physicist working on the transition probability in operator algebras — and it is his variational formula, not the original square-root definition, that the field actually runs on.

Origins · 1 minArmin Uhlmann and how this was discovered

Armin Uhlmann

1930–2016 · German physicist · operator algebras and quantum information geometry

Armin Uhlmann spent most of his career at the University of Leipzig, in the tradition of mathematical physics that runs through the theory of operator algebras. In the mid-1970s he was studying the notion of "transition probability" between states of a general *-algebra — a natural object in the algebraic approach to quantum theory pioneered by the Leipzig and Göttingen schools. In a 1976 paper he proved what is now called Uhlmann's theorem: the transition probability between two density operators equals the maximum, over all their purifications in a common enlarged space, of the squared overlap of the purifying vectors. At the time this was a statement about states on von Neumann algebras, phrased in a language most physicists did not read, and for over a decade it lived quietly in the mathematical-physics literature. Its explosion into the mainstream came in the 1990s, when Richard Jozsa re-derived the mixed-state fidelity from a short list of physically motivated axioms and connected it directly to Uhlmann's variational formula, and when the rise of quantum computing made "how close are these two states?" a question thousands of experimentalists had to answer. Uhlmann also gave his name to the Bures–Uhlmann metric and to the Uhlmann phase, a mixed-state generalization of the geometric Berry phase. It is a recurring pattern in quantum information: the physically decisive object was first written down as a theorem about operator algebras, and only later recognized as the thing the whole field needed.

History

1969

— Bures introduces the metric on states that fidelity will generate

1976

— Uhlmann proves the transition probability equals the maximal purification overlap

1994

— Jozsa axiomatizes mixed-state fidelity and gives the single-qubit closed form

1999

— Fuchs and van de Graaf bracket fidelity and trace distance

2008

— randomized benchmarking makes gate fidelity a routine laboratory measurement

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §9.2.2 (fidelity) and §12.6 · Wilde, Quantum Information Theory, 2nd ed., Ch. 9 (distinguishability measures) · Watrous, The Theory of Quantum Information, Ch. 3 (fidelity and the Uhlmann theorem) · Preskill, Lecture Notes for Physics 219, Ch. 10 · Bengtsson & Życzkowski, Geometry of Quantum States, Ch. 9 (Bures geometry).

Primary sources. Uhlmann, "The transition probability in the state space of a -algebra", Rep. Math. Phys. <strong>9</strong>, 273 (1976) · Jozsa, "Fidelity for mixed quantum states", J. Mod. Opt. <strong>41</strong>, 2315 (1994) · Fuchs & van de Graaf, IEEE Trans. Inf. Theory* 45, 1216 (1999).

Provenance. A verifier should recompute: (1) the equality of the two definitional forms via the singular-value identity $\|X^\dagger\|_1=\|X\|_1$ ; (2) the Uhlmann derivation through vectorization and $\max_U|\mathrm{Tr}(XU)|=\|X\|_1$ (Watrous Ch. 3); (3) the pure-state trace-distance identity $\sqrt{1-|\langle\psi|\phi\rangle|^2}$ and the Fuchs–van de Graaf bracket; (4) the worked qubit example — Jozsa's formula gives $F=0.50+0.24=0.74$ , $\sqrt{F}=0.860$ , trace distance $T=\tfrac12|\vec r-\vec s|=0.500$ , and the bracket $0.140\le0.500\le0.510$ ; (5) the classical case $\big(\sqrt{0.45}+\sqrt{0.05}\big)^2=0.80$ . Conventions follow the squared definition ( $F=|\langle\psi|\phi\rangle|^2$ ); compare against Nielsen & Chuang's unsquared $\sqrt{F}$ when matching formulas.

Prerequisites. Von Neumann Entropy · Quantum Relative Entropy · Stinespring Dilation Theorem · Quantum Channels and CPTP Maps. Leads to. Data-Processing Inequality · Holevo Bound · Quantum Channels and CPTP Maps · Complementary Channels and Degradability.

Ask the tutor: Ask me to walk through Uhlmann's theorem line by line, or to compute the fidelity, Bures angle, and Fuchs–van de Graaf bounds for any pair of states you name.

Is fidelity a distance between quantum states?

Not directly — and this trips people up. Fidelity runs the "wrong way" for a distance: it is $1$ when the states coincide and decreases as they separate, so it is a similarity, like $\cos^2$ of an angle, not a metric. What fidelity does is generate two honest metrics. The Bures distance $D_B=\sqrt{2(1-\sqrt{F})}$ and the Bures angle $\theta_B=\arccos\sqrt{F}$ both satisfy the triangle inequality and vanish only when $\rho=\sigma$ . So when you want a genuine notion of "how far apart", use the Bures angle; when you want a convenient, multiplicative closeness score with a clean operational meaning, use $F$ itself.

Why does purifying the states make fidelity easier, not harder?

It seems paradoxical to enlarge the problem — replacing two mixed states by pure states on a bigger space — yet that is exactly what simplifies it. The fidelity of pure states is trivial, just $|\langle\psi|\phi\rangle|^2$ . The only difficulty for mixed states is that each has many purifications, related by unitaries on the reference. Uhlmann's theorem says: don't fight that freedom, use it — rotate the references to align the purifications as well as possible, and the best overlap is $\sqrt{F}$ . The operator square root in the definition is the algebraic residue of having optimized the reference freedom away, so enlarging the space trades a hard operator computation for an easy geometric optimization.

Can two full-rank states ever have fidelity exactly zero?

No. Fidelity vanishes only when the supports of $\rho$ and $\sigma$ are orthogonal, and a full-rank state has support equal to the whole Hilbert space. So any two full-rank states — for example any two faithful thermal states, or any states with a maximally mixed component — necessarily have $F>0$ . Zero fidelity is reserved for perfectly distinguishable states, which requires at least one of them to be rank-deficient with support living entirely in the kernel of the other. This is the mirror image of $F=1$ , which happens only when $\rho=\sigma$ exactly. In between, $F$ interpolates smoothly, and near either endpoint the Bures angle $\arccos\sqrt{F}$ gives the honest metric distance.

Why is the measurement form of fidelity a minimum, while trace distance is a maximum?

Because the two measures ask opposite questions. Trace distance is the best single-shot distinguishing bias, so it maximizes over POVMs — you look for the measurement that separates the states most. Fidelity measures overlap, and any measurement is a channel, which by data processing can only increase overlap. So for every POVM the classical Bhattacharyya coefficient of the outcome distributions is at least $\sqrt{F}$ , and the quantum fidelity is the tightest such lower bound — the minimum over POVMs, achieved by the optimal (Helstrom-type) measurement. The duality "distance maximizes, fidelity minimizes" is a direct reflection of one being a separation and the other an overlap.

Does high fidelity guarantee small error in any later computation?

It gives a rigorous but sometimes loose guarantee. Through Fuchs–van de Graaf, $F\ge1-\epsilon$ forces the trace distance below $\sqrt{\epsilon}$ , and trace distance bounds the difference in any measurement probability, so high fidelity does cap downstream errors. But the bound can be pessimistic, and fidelity is multiplicative: $n$ copies each at fidelity $F$ have joint fidelity $F^n$ , which decays — a subtlety that matters when a protocol consumes many imperfect states. Use fidelity as a certified, composable figure of merit, not as a promise that a specific task sees no error.

How is fidelity actually measured in a real experiment?

Rarely by evaluating the operator square root, which would need full tomography of both states. When the target is a known pure state $|\psi\rangle$ , the fidelity is just $\langle\psi|\sigma|\psi\rangle$ , estimable by measuring the single projector $|\psi\rangle\langle\psi|$ — or, for stabilizer states, by sampling a few stabilizer expectation values (direct fidelity estimation). For gates, randomized benchmarking runs random Clifford sequences of growing length and reads the average gate fidelity off the exponential decay of the survival probability, robust to state-preparation and measurement errors. For large entangled outputs, cross-entropy benchmarking compares measured bitstring frequencies to the ideal distribution. Each sidesteps the square root by exploiting structure in the state or process being certified.