Simple Algorithm
Quantum Computing · Rydberg State · 28 min

Fidelity & Trace Distance

How close are two quantum states, operationally?

"99% fidelity" sounds nearly perfect, yet on its own it only guarantees that no measurement can shift any outcome probability by more than 10%. Trace distance and fidelity are the two standard yardsticks for quantum states. This lesson gives each an operational meaning and proves the inequalities that convert one into the other.

What you’ll learn

  • Prove that trace distance equals the largest difference in outcome probabilities, and that it contracts under every channel.
  • Define fidelity, compute it for qubits, and prove Uhlmann's theorem.
  • Prove the Fuchs–van de Graaf inequalities and use them to convert between the two measures.
  • Apply the bounds to many copies, to the gentle measurement lemma, and to reported gate fidelities.

Trace distance: the best possible bias

DerivationProof
  1. Tr⁡Q=Tr⁡S=12Tr⁡(Q+S)=12∥ρ−σ∥1=D(ρ,σ)\Tr Q = \Tr S = \tfrac12\Tr(Q+S) = \tfrac12\norm{\rho-\sigma}_1 = D(\rho,\sigma)

    Split ρ−σ=Q−S\rho-\sigma = Q-S into its positive and negative parts (Q,S≥0Q,S\ge0 with orthogonal supports). Since Tr⁡(ρ−σ)=0\Tr(\rho-\sigma) = 0:

  2. Tr⁡(E(ρ−σ))=Tr⁡(EQ)−Tr⁡(ES)≤Tr⁡(EQ)≤Tr⁡Q=D(ρ,σ)\Tr\big(E(\rho-\sigma)\big) = \Tr(EQ)-\Tr(ES)\le\Tr(EQ)\le\Tr Q = D(\rho,\sigma)

    For any effect 0≤E≤I0\le E\le\id:

  3. Equality holds for E=ΠQE = \Pi_Q, the projector onto the support of QQ, for which Tr⁡(ΠQS)=0\Tr(\Pi_QS) = 0. For a POVM, group the outcomes with pk>qkp_k>q_k into a single effect EE; then 12∑k∣pk−qk∣=Tr⁡(E(ρ−σ))≤D\tfrac12\sum_k\abs{p_k-q_k} = \Tr\big(E(\rho-\sigma)\big)\le D. □\square

For qubits, ρ−σ=12(r⃗−s⃗)⋅σ⃗\rho-\sigma = \tfrac12(\vec r-\vec s)\cdot\vec\sigma has eigenvalues ±12∣r⃗−s⃗∣\pm\tfrac12\abs{\vec r-\vec s}, so D(ρ,σ)=12∣r⃗−s⃗∣D(\rho,\sigma) = \tfrac12\abs{\vec r-\vec s}: half the Euclidean distance in the Bloch ball, with the optimal measurement along r⃗−s⃗\vec r-\vec s. With equal priors, the best single-shot guess between ρ\rho and σ\sigma succeeds with probability 12(1+D)\tfrac12(1+D), the Helstrom bound.

Check your understanding

Let ρ=∣0⟩ ⁣⟨0∣\rho = \dyad0 and σ=I/2\sigma = \id/2. What is D(ρ,σ)D(\rho,\sigma)?

  1. 11
  2. 1/21/\sqrt2
  3. 1/21/2
  4. 1/41/4
Show the answer

C. 1/21/2

The Bloch vectors are z^\hat z and 0⃗\vec 0, so D=12∣z^−0⃗∣=12D = \tfrac12\abs{\hat z-\vec0} = \tfrac12. Directly: ρ−σ=diag⁡(12,−12)\rho-\sigma = \diag(\tfrac12,-\tfrac12) and 12(12+12)=12\tfrac12\big(\tfrac12+\tfrac12\big) = \tfrac12.

Fidelity and Uhlmann's theorem

DerivationProof sketch
  1. ∣ψρ⟩=(ρ⊗V)∣Γ⟩,∣ψσ⟩=(σ⊗W)∣Γ⟩\ket{\psi_\rho} = (\sqrt\rho\otimes V)\ket\Gamma,\qquad \ket{\psi_\sigma} = (\sqrt\sigma\otimes W)\ket\Gamma

    With ∣Γ⟩=∑i∣i⟩∣i⟩\ket\Gamma = \sum_i\ket{i}\ket{i}, every purification on R≅AR\cong A is the canonical one up to a unitary on RR (uniqueness of purifications):

  2. ⟨ψρ∣ψσ⟩=⟨Γ∣(ρσ⊗V†W)∣Γ⟩=Tr⁡(ρσ U),U:=(V†W)T\braket{\psi_\rho|\psi_\sigma} = \bra\Gamma\big(\sqrt\rho\sqrt\sigma\otimes V^\dagger W\big)\ket\Gamma = \Tr\big(\sqrt\rho\sqrt\sigma\,U\big),\qquad U := (V^\dagger W)^{\mathsf T}

    Use the transpose trick ⟨Γ∣(A⊗B)∣Γ⟩=Tr⁡(ABT)\bra\Gamma(A\otimes B)\ket\Gamma = \Tr(AB^{\mathsf T}):

  3. max⁡U unitary∣Tr⁡(ρσ U)∣=∥ρσ∥1=F(ρ,σ)□\max_{U\ \text{unitary}}\abs{\Tr\big(\sqrt\rho\sqrt\sigma\,U\big)} = \norm{\sqrt\rho\sqrt\sigma}_1 = \sqrt{F(\rho,\sigma)}\quad\square

    As VV and WW range over unitaries, so does UU. For AA with polar decomposition A=W0∣A∣A = W_0\abs{A}, Hölder's inequality gives ∣Tr⁡(AU)∣≤∥A∥1∥U∥∞=∥A∥1\abs{\Tr(AU)}\le\norm{A}_1\norm{U}_\infty = \norm{A}_1, with equality at U=W0†U = W_0^\dagger. Hence

Think first The operator ρ σρ\sqrt\rho\,\sigma\sqrt\rho is not symmetric in ρ\rho and σ\sigma. Why is F(ρ,σ)=F(σ,ρ)F(\rho,\sigma) = F(\sigma,\rho) anyway?

Because ∥A∥1=∥A†∥1\norm{A}_1 = \norm{A^\dagger}_1 and (ρσ)†=σρ(\sqrt\rho\sqrt\sigma)^\dagger = \sqrt\sigma\sqrt\rho. Uhlmann's theorem makes it obvious as well: the overlap ∣⟨ψρ∣ψσ⟩∣\abs{\braket{\psi_\rho|\psi_\sigma}} is symmetric.

Other properties follow just as quickly: F=1F = 1 iff ρ=σ\rho = \sigma; FF is multiplicative, F(ρ1⊗ρ2,σ1⊗σ2)=F(ρ1,σ1) F(ρ2,σ2)F(\rho_1\otimes\rho_2,\sigma_1\otimes\sigma_2) = F(\rho_1,\sigma_1)\,F(\rho_2,\sigma_2); and FF can only increase under a channel Φ\Phi. For the last, push optimal purifications through a Stinespring isometry of Φ\Phi: they become purifications of Φ(ρ)\Phi(\rho) and Φ(σ)\Phi(\sigma) with the same overlap, so Uhlmann gives F(Φ(ρ),Φ(σ))≥F(ρ,σ)F\big(\Phi(\rho),\Phi(\sigma)\big)\ge F(\rho,\sigma).

F(ρ,σ)=Tr⁡(ρσ)+2det⁡ρ det⁡σ=12(1+r⃗⋅s⃗+(1−∣r⃗∣2)(1−∣s⃗∣2))F(\rho,\sigma) = \Tr(\rho\sigma)+2\sqrt{\det\rho\,\det\sigma} = \tfrac12\Big(1+\vec r\cdot\vec s+\sqrt{\big(1-\abs{\vec r}^2\big)\big(1-\abs{\vec s}^2\big)}\Big)
The qubit formula. If M=ρ σρM = \sqrt\rho\,\sigma\sqrt\rho has eigenvalues μ1,μ2\mu_1,\mu_2, then (μ1+μ2)2=Tr⁡M+2det⁡M\big(\sqrt{\mu_1}+\sqrt{\mu_2}\big)^2 = \Tr M+2\sqrt{\det M}, with Tr⁡M=Tr⁡(ρσ)\Tr M = \Tr(\rho\sigma) and det⁡M=det⁡ρ det⁡σ\det M = \det\rho\,\det\sigma.

Fuchs–van de Graaf: converting between the two

DerivationProof
  1. D(ψ,ϕ)=1−∣⟨ψ∣ϕ⟩∣2=1−F(ψ,ϕ)D(\psi,\phi) = \sqrt{1-\abs{\braket{\psi|\phi}}^2} = \sqrt{1-F(\psi,\phi)}

    Pure states. In the plane spanned by ∣ψ⟩\ket\psi and ∣ϕ⟩\ket\phi, the traceless operator ∣ψ⟩ ⁣⟨ψ∣−∣ϕ⟩ ⁣⟨ϕ∣\dyad\psi-\dyad\phi has determinant −(1−∣⟨ψ∣ϕ⟩∣2)-\big(1-\abs{\braket{\psi|\phi}}^2\big), hence eigenvalues ±1−F\pm\sqrt{1-F}:

  2. D(ρ,σ)≤D(ψρ,ψσ)=1−F(ρ,σ)D(\rho,\sigma)\le D(\psi_\rho,\psi_\sigma) = \sqrt{1-F(\rho,\sigma)}

    Upper bound. By Uhlmann, pick purifications with ∣⟨ψρ∣ψσ⟩∣2=F(ρ,σ)\abs{\braket{\psi_\rho|\psi_\sigma}}^2 = F(\rho,\sigma). The partial trace is a channel, so contractivity gives

  3. 1−F=12∑k(pk−qk)2≤12∑k∣pk−qk∣(pk+qk)=12∑k∣pk−qk∣≤D□1-\sqrt F = \tfrac12\sum_k\big(\sqrt{p_k}-\sqrt{q_k}\big)^2\le\tfrac12\sum_k\abs{\sqrt{p_k}-\sqrt{q_k}}\big(\sqrt{p_k}+\sqrt{q_k}\big) = \tfrac12\sum_k\abs{p_k-q_k}\le D\quad\square

    Lower bound. Fuchs and Caves showed that F=min⁡∑kpkqk\sqrt F = \min\sum_k\sqrt{p_kq_k} over all POVMs, with pk=Tr⁡(Ekρ)p_k = \Tr(E_k\rho) and qk=Tr⁡(Ekσ)q_k = \Tr(E_k\sigma). For a minimizing POVM, using ∣p−q∣≤p+q\abs{\sqrt p-\sqrt q}\le\sqrt p+\sqrt q:

Predict before you look

A lab reports fidelity F=0.99F = 0.99 (squared convention) between its state and a pure target. What is the largest trace distance consistent with this?

Show the answer

0.1

D≤1−F=0.01=0.1D\le\sqrt{1-F} = \sqrt{0.01} = 0.1, attained when the lab's state is also pure. The lower bound 1−0.99≈0.0051-\sqrt{0.99}\approx0.005 is twenty times smaller: fidelity and trace distance can differ by a square root.

PropertyTrace distance DDFidelity FF
Operational meaningBest single-shot bias: Psucc=12(1+D)P_{\text{succ}} = \tfrac12(1+D) for equal priorsBest overlap of purifications; F\sqrt F is the smallest classical (Bhattacharyya) overlap over measurements
Under channelsContracts: D(Φρ,Φσ)≤D(ρ,σ)D(\Phi\rho,\Phi\sigma)\le D(\rho,\sigma)Grows: F(Φρ,Φσ)≥F(ρ,σ)F(\Phi\rho,\Phi\sigma)\ge F(\rho,\sigma)
Tensor powersOnly bounds, e.g. D(ρ⊗n,σ⊗n)≤nD(ρ,σ)D(\rho^{\otimes n},\sigma^{\otimes n})\le nD(\rho,\sigma)Multiplicative: F(ρ⊗n,σ⊗n)=F(ρ,σ)nF(\rho^{\otimes n},\sigma^{\otimes n}) = F(\rho,\sigma)^n
Qubits12∣r⃗−s⃗∣\tfrac12\abs{\vec r-\vec s}12(1+r⃗⋅s⃗+(1−∣r⃗∣2)(1−∣s⃗∣2))\tfrac12\big(1+\vec r\cdot\vec s+\sqrt{(1-\abs{\vec r}^2)(1-\abs{\vec s}^2)}\big)

Multiplicativity makes fidelity the tool of choice for many copies. Combined with Fuchs–van de Graaf, it brackets the nn-copy trace distance:

1−F(ρ,σ)n/2 ≤ D(ρ⊗n,σ⊗n) ≤ 1−F(ρ,σ)n1-F(\rho,\sigma)^{n/2}\ \le\ D\big(\rho^{\otimes n},\sigma^{\otimes n}\big)\ \le\ \sqrt{1-F(\rho,\sigma)^n}
Distinguishability approaches 1 exponentially fast in nn whenever F<1F<1. The exact exponent of the optimal error probability is the quantum Chernoff bound (see State Discrimination & the Helstrom Bound).

Using the bounds

An outcome that is almost certain barely disturbs the state, and ε\sqrt\varepsilon is the right scaling. This lemma powers sequential decoding in quantum Shannon theory and Aaronson's shadow tomography; Weak Measurement & the Zeno Effect develops it further.

Key takeaways

  • D(ρ,σ)=12∥ρ−σ∥1D(\rho,\sigma) = \tfrac12\norm{\rho-\sigma}_1 is the largest difference in any outcome probability; it contracts under channels, and for qubits it is half the Bloch-ball distance.
  • F(ρ,σ)=∥ρσ∥12F(\rho,\sigma) = \norm{\sqrt\rho\sqrt\sigma}_1^2 is the maximal overlap of purifications (Uhlmann); it is multiplicative and grows under channels. Check which convention, FF or F\sqrt F, a paper uses.
  • Fuchs–van de Graaf: 1−F≤D≤1−F1-\sqrt F\le D\le\sqrt{1-F}, with equality on the right for pure states. A fidelity of 99% can still mean a trace distance of 10%.
  • Gentle measurement: an outcome of probability at least 1−ε1-\varepsilon disturbs the state by at most ε\sqrt\varepsilon in trace distance.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

Two pure states have ∣⟨ψ∣ϕ⟩∣2=0.75\abs{\braket{\psi|\phi}}^2 = 0.75. What is their trace distance?

  1. 0.250.25
  2. 0.50.5
  3. 0.750.75
  4. 0.1340.134
Show the answer

B. 0.50.5

For pure states D=1−F=0.25=0.5D = \sqrt{1-F} = \sqrt{0.25} = 0.5 exactly.

Question 2

The same channel Φ\Phi is applied to two states ρ\rho and σ\sigma. What is guaranteed?

  1. DD can only decrease and FF can only increase
  2. DD and FF are both unchanged
  3. DD can increase if Φ\Phi is entanglement-breaking
  4. FF can decrease if Φ\Phi is not unital
Show the answer

A. DD can only decrease and FF can only increase

Data processing: D(Φρ,Φσ)≤D(ρ,σ)D(\Phi\rho,\Phi\sigma)\le D(\rho,\sigma) and F(Φρ,Φσ)≥F(ρ,σ)F(\Phi\rho,\Phi\sigma)\ge F(\rho,\sigma) for every CPTP map. Unitaries keep both unchanged; general channels can only make states harder to tell apart.

Question 3

An outcome EE has Tr⁡(Eρ)=0.99\Tr(E\rho) = 0.99. What upper bound does the gentle measurement lemma give for the trace distance DD between ρ\rho and the normalized post-measurement state EρE/Tr⁡(Eρ)\sqrt E\rho\sqrt E/\Tr(E\rho)?

  1. 0.010.01
  2. 0.00010.0001
  3. 0.20.2
  4. 0.10.1
Show the answer

D. 0.10.1

Here ε=0.01\varepsilon = 0.01, so F(ρ,ρ′)≥0.99F(\rho,\rho')\ge0.99 and D≤0.01=0.1D\le\sqrt{0.01} = 0.1. The disturbance scales like the square root of the failure probability.

Question 4

Which property makes fidelity especially convenient for nn-copy problems?

  1. It is a metric on states
  2. It is linear in ρ\rho
  3. It is multiplicative: F(ρ⊗n,σ⊗n)=F(ρ,σ)nF(\rho^{\otimes n},\sigma^{\otimes n}) = F(\rho,\sigma)^n
  4. It equals one minus the trace distance
Show the answer

C. It is multiplicative: F(ρ⊗n,σ⊗n)=F(ρ,σ)nF(\rho^{\otimes n},\sigma^{\otimes n}) = F(\rho,\sigma)^n

Multiplicativity follows from ρ⊗ρ′=ρ⊗ρ′\sqrt{\rho\otimes\rho'} = \sqrt\rho\otimes\sqrt{\rho'} and ∥A⊗B∥1=∥A∥1∥B∥1\norm{A\otimes B}_1 = \norm{A}_1\norm{B}_1. Trace distance has no such rule, so one bounds it through FnF^n using Fuchs–van de Graaf.

Further reading

  • C. A. Fuchs & J. van de Graaf, "Cryptographic distinguishability measures for quantum-mechanical states", IEEE Transactions on Information Theory 45, 1216 (1999).
  • A. Uhlmann, "The 'transition probability' in the state space of a ∗^*-algebra", Reports on Mathematical Physics 9, 273 (1976).
  • R. Jozsa, "Fidelity for mixed quantum states", Journal of Modern Optics 41, 2315 (1994).
  • J. Watrous, The Theory of Quantum Information, Ch. 3 (Cambridge, 2018).
  • M. A. Nielsen & I. L. Chuang, Quantum Computation and Quantum Information, Ch. 9 (Cambridge, 2010).