Simple Algorithm
Quantum Computing · Rydberg State · 30 min

Quantum Metrology & the Heisenberg Limit

Entanglement-enhanced sensing and its limits.

Measure a phase with NN independent probes and your error falls as 1/N1/\sqrt N: the standard quantum limit. Entangle the probes and the bound becomes 1/N1/N: the Heisenberg limit. This lesson proves both bounds from the quantum Fisher information, shows how GHZ and NOON states reach the second, and explains why noise usually takes most of the advantage away.

What you’ll learn

  • Define the classical and quantum Fisher information and state the quantum Cramér–Rao bound.
  • Derive the standard quantum limit for product probes and prove the Heisenberg limit for arbitrary probes.
  • Show that GHZ and NOON states attain the Heisenberg limit, and explain the phase-ambiguity problem and its resolution.
  • Prove that uncorrelated Markovian dephasing removes the entanglement advantage in frequency estimation, and describe when error correction helps.

Fisher information and the Cramér–Rao bound

A probe state ρφ\rho_\varphi depends on an unknown phase φ\varphi. Measuring it gives outcome xx with probability p(x∣φ)p(x\mid\varphi), and you repeat ν\nu times. Any unbiased estimator obeys the Cramér–Rao bound

Δφ≥1νF(φ),F(φ)=∑x(∂φp(x∣φ))2p(x∣φ),\Delta\varphi \ge \frac{1}{\sqrt{\nu F(\varphi)}}, \qquad F(\varphi) = \sum_x \frac{\big(\partial_\varphi p(x\mid\varphi)\big)^2}{p(x\mid\varphi)},

and maximum-likelihood estimation attains it as ν→∞\nu\to\infty. FF is the classical Fisher information of the chosen measurement.

DerivationOne probe in an interferometer
  1. p0=cos⁡2φ2,p1=sin⁡2φ2,∂φp0=−∂φp1=−12sin⁡φp_0 = \cos^2\tfrac{\varphi}{2}, \quad p_1 = \sin^2\tfrac{\varphi}{2}, \quad \partial_\varphi p_0 = -\partial_\varphi p_1 = -\tfrac12\sin\varphi

    The Mach–Zehnder (or Ramsey) fringe.

  2. F=14sin⁡2φcos⁡2φ2+14sin⁡2φsin⁡2φ2=14sin⁡2φsin⁡2φ2cos⁡2φ2=1F = \frac{\tfrac14\sin^2\varphi}{\cos^2\frac{\varphi}{2}} + \frac{\tfrac14\sin^2\varphi}{\sin^2\frac{\varphi}{2}} = \frac{\tfrac14\sin^2\varphi}{\sin^2\frac{\varphi}{2}\cos^2\frac{\varphi}{2}} = 1

    Use sin⁡φ=2sin⁡φ2cos⁡φ2\sin\varphi = 2\sin\frac\varphi2\cos\frac\varphi2. One photon (or atom) carries exactly one unit of Fisher information, at every working point.

Standard quantum limit versus Heisenberg limit

DerivationProduct probes and GHZ probes
  1. G=∑k=1N12Zk,Uφ=e−iφG=⨂ke−iφZk/2G = \sum_{k=1}^{N}\tfrac12 Z_k, \qquad U_\varphi = e^{-i\varphi G} = \bigotimes_k e^{-i\varphi Z_k/2}

    Each of NN qubit probes picks up the phase independently.

  2. ∣+⟩⊗N:(ΔG)2=N⋅14  ⇒  FQ=N  ⇒  ΔφSQL=1νN\ket{+}^{\otimes N}:\quad (\Delta G)^2 = N\cdot\tfrac14 \;\Rightarrow\; F_Q = N \;\Rightarrow\; \Delta\varphi_{\mathrm{SQL}} = \frac{1}{\sqrt{\nu N}}

    Variances add for product states: the standard quantum limit (shot-noise limit).

  3. ∣GHZ⟩=12(∣0⟩⊗N+∣1⟩⊗N)  →Uφ  12(∣0⟩⊗N+eiNφ∣1⟩⊗N)\ket{\mathrm{GHZ}} = \tfrac{1}{\sqrt2}\big(\ket{0}^{\otimes N} + \ket{1}^{\otimes N}\big) \;\xrightarrow{U_\varphi}\; \tfrac{1}{\sqrt2}\big(\ket{0}^{\otimes N} + e^{iN\varphi}\ket{1}^{\otimes N}\big)

    Up to a global phase. The two branches are eigenvectors of GG with eigenvalues ±N/2\pm N/2, so the relative phase is NφN\varphi.

  4. (ΔG)2=N24  ⇒  FQ=N2  ⇒  ΔφHL=1ν N(\Delta G)^2 = \tfrac{N^2}{4} \;\Rightarrow\; F_Q = N^2 \;\Rightarrow\; \Delta\varphi_{\mathrm{HL}} = \frac{1}{\sqrt\nu\,N}

    The Heisenberg limit: a quadratic gain in Fisher information.

Predict before you look

The circuit below prepares a 3-qubit GHZ state, applies P(π/3)P(\pi/3) to each qubit, disentangles, and interferes qubit 0. What is the probability (in percent) that qubit 0 reads 11?

Show the answer

100 %

The GHZ branches acquire relative phase 3φ=π3\varphi = \pi, so qubit 0 ends in H∣−⟩=∣1⟩H\ket{-} = \ket{1}: P(1)=sin⁡2(3φ/2)=1P(1) = \sin^2(3\varphi/2) = 1. A single probe with the same φ\varphi gives only sin⁡2(π/6)=14\sin^2(\pi/6) = \tfrac14. The fringe runs three times faster: super-resolution.

Reaching the limit in practice

A GHZ state measures NφN\varphi only modulo 2π2\pi, so on its own it cannot tell φ\varphi from φ+2π/N\varphi + 2\pi/N. The fix is the same as in phase estimation: combine probes (or repeated applications) with N=1,2,4,…,2t−1N = 1, 2, 4, \dots, 2^{t-1} and read the digits of φ\varphi from coarse to fine. Higgins et al. (2007) realized Heisenberg-limited phase estimation this way with single photons passed repeatedly through the phase, without entanglement, using adaptive measurements.

With photons, the analogue of GHZ is the NOON state 12(∣N,0⟩+∣0,N⟩)\tfrac{1}{\sqrt2}(\ket{N,0} + \ket{0,N}) across the two arms, whose relative phase is NφN\varphi. Small-NN experiments showed super-resolved fringes (Mitchell, Lundeen & Steinberg 2004; Walther et al. 2004) and sensitivity beyond the standard quantum limit (Nagata et al. 2007).

StrategyScaling of Δφ\Delta\varphiComment
NN independent probes1/N1/\sqrt N (SQL)robust; the classical benchmark
GHZ / NOON1/N1/N (HL)fragile: one lost particle destroys the phase information
spin- or light-squeezed statesbetween SQL and HLrobust and practical; used in clocks and gravitational-wave detectors
sequential / phase estimation1/N1/N in the number of usesno entanglement; needs coherence over long total times

Noise and the limits of the limit

DerivationFrequency estimation under independent dephasing (Huelga et al. 1997)
  1. signal φ=ωt,coherence of each qubit∝e−γt,total time T=νt\text{signal } \varphi = \omega t, \qquad \text{coherence of each qubit} \propto e^{-\gamma t}, \qquad \text{total time } T = \nu t

    Estimate a frequency ω\omega; each run lasts tt and dephasing at rate γ\gamma shrinks the fringe contrast.

  2. product: Fω=ν Nt2e−2γt=TN t e−2γt  ≤  TN2eγ  (t∗=12γ)\text{product: } F_\omega = \nu\,N t^2 e^{-2\gamma t} = TN\,t\,e^{-2\gamma t} \;\le\; \frac{TN}{2e\gamma}\ \ (t^* = \tfrac{1}{2\gamma})

    Each qubit gives Fisher information t2C2t^2C^2 at the best working point, with contrast C=e−γtC = e^{-\gamma t}.

  3. GHZ: Fω=ν N2t2e−2Nγt=TN2 t e−2Nγt  ≤  TN2eγ  (t∗=12Nγ)\text{GHZ: } F_\omega = \nu\,N^2t^2e^{-2N\gamma t} = TN^2\,t\,e^{-2N\gamma t} \;\le\; \frac{TN}{2e\gamma}\ \ (t^* = \tfrac{1}{2N\gamma})

    The GHZ coherence decays NN times faster, e−Nγte^{-N\gamma t}, exactly cancelling its NN-fold faster phase.

  4. With optimized interrogation times the two strategies give identical precision, Δω≥2eγ/(NT)\Delta\omega \ge \sqrt{2e\gamma/(NT)}: back to 1/N1/\sqrt N. More generally, for uncorrelated Markovian noise the asymptotic gain over the SQL is at most a constant factor (Escher, de Matos Filho & Davidovich 2011; Demkowicz-Dobrzański, Kołodyński & Guţă 2012).

Check your understanding

Under independent Markovian dephasing, with the interrogation time optimized separately for each strategy, how does a GHZ state of NN atoms compare with NN unentangled atoms for frequency estimation?

  1. GHZ gives 1/N1/N scaling, a quadratic improvement
  2. GHZ is N\sqrt N times better
  3. They achieve the same precision; the faster phase accumulation is exactly offset by the NN-fold faster decoherence
  4. GHZ is useless because it decoheres instantly
Show the answer

C. They achieve the same precision; the faster phase accumulation is exactly offset by the NN-fold faster decoherence

Both optimize to Fω=TN/(2eγ)F_\omega = TN/(2e\gamma). The GHZ state accumulates phase NN times faster but loses coherence NN times faster, so its optimal run is NN times shorter and the gain disappears.

Key takeaways

  • Precision is bounded by Δφ≥1/νFQ\Delta\varphi \ge 1/\sqrt{\nu F_Q}, with FQ=4(ΔG)2F_Q = 4(\Delta G)^2 for pure states; one interferometer probe has F=1F = 1.
  • Product probes give FQ=NF_Q = N (SQL, 1/N1/\sqrt N); any probes obey FQ≤N2(λmax⁡−λmin⁡)2F_Q \le N^2(\lambda_{\max}-\lambda_{\min})^2 (Heisenberg limit, 1/N1/N), saturated by GHZ and NOON states.
  • The 2π/N2\pi/N phase ambiguity is resolved by multi-scale (phase-estimation-like) strategies, and sequential schemes reach 1/N1/N without entanglement.
  • Uncorrelated Markovian noise reduces the gain to a constant factor; error correction restores 1/N1/N only when the signal lies outside the noise's Lindblad span.

Check yourself

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

Question 1

N=100N = 100 unentangled qubit probes, a single repetition (ν=1\nu = 1), ideal measurement. What is the smallest achievable Δφ\Delta\varphi?

  1. 0.010.01
  2. 0.10.1
  3. 11
  4. 0.0010.001
Show the answer

B. 0.10.1

Product probes give FQ=N=100F_Q = N = 100, so Δφ≥1/100=0.1\Delta\varphi \ge 1/\sqrt{100} = 0.1. The Heisenberg limit with entangled probes would be 1/N=0.011/N = 0.01.

Question 2

An NN-qubit GHZ state evolves under e−iφZ/2e^{-i\varphi Z/2} on each qubit. What relative phase separates its two branches?

  1. φ\varphi
  2. N φ\sqrt N\,\varphi
  3. NφN\varphi
  4. N2φN^2\varphi
Show the answer

C. NφN\varphi

∣0⟩⊗N\ket{0}^{\otimes N} and ∣1⟩⊗N\ket{1}^{\otimes N} are eigenvectors of G=∑kZk/2G = \sum_k Z_k/2 with eigenvalues ±N/2\pm N/2, so their relative phase is NφN\varphi, which gives FQ=N2F_Q = N^2.

Question 3

Why can't a single GHZ-type measurement estimate a completely unknown φ∈[0,2π)\varphi\in[0, 2\pi) at the Heisenberg limit?

  1. GHZ states cannot be prepared for N>2N > 2
  2. The Heisenberg limit applies only to photons
  3. The quantum Fisher information of a GHZ state is only NN
  4. It determines NφN\varphi only modulo 2π2\pi, so φ\varphi is ambiguous up to multiples of 2π/N2\pi/N; a ladder of NN values or adaptive scheme is needed
Show the answer

D. It determines NφN\varphi only modulo 2π2\pi, so φ\varphi is ambiguous up to multiples of 2π/N2\pi/N; a ladder of NN values or adaptive scheme is needed

The fringe repeats NN times over [0,2π)[0, 2\pi). Resolving which period you are in requires coarser measurements as well, exactly as phase estimation combines U,U2,U4,…U, U^2, U^4, \dots

Question 4

When can quantum error correction restore Heisenberg scaling in the presence of Markovian noise?

  1. When the signal Hamiltonian is not in the Lindblad span of the noise (HNLS)
  2. Always, given enough ancilla qubits
  3. Never: noise always forces 1/N1/\sqrt N
  4. Only for photonic sensors
Show the answer

A. When the signal Hamiltonian is not in the Lindblad span of the noise (HNLS)

If the signal generator lies outside the span of the noise operators (and their products), a code can correct the noise while leaving the signal intact, giving 1/N1/N scaling. Otherwise any code that removes the noise also removes the signal.

Further reading

  • S. L. Braunstein & C. M. Caves, "Statistical distance and the geometry of quantum states", Phys. Rev. Lett. 72, 3439 (1994).
  • V. Giovannetti, S. Lloyd & L. Maccone, "Quantum metrology", Phys. Rev. Lett. 96, 010401 (2006).
  • S. F. Huelga, C. Macchiavello, T. Pellizzari, A. K. Ekert, M. B. Plenio & J. I. Cirac, "Improvement of frequency standards with quantum entanglement", Phys. Rev. Lett. 79, 3865 (1997).
  • B. L. Higgins, D. W. Berry, S. D. Bartlett, H. M. Wiseman & G. J. Pryde, "Entanglement-free Heisenberg-limited phase estimation", Nature 450, 393 (2007).
  • L. Pezzè, A. Smerzi, M. K. Oberthaler, R. Schmied & P. Treutlein, "Quantum metrology with nonclassical states of atomic ensembles", Rev. Mod. Phys. 90, 035005 (2018).