Quantum computing · Experiment

Quantum phase estimation on a Rydberg quantum simulator

Hendrik Timme, Benjamin F. Schiffer, Ohad Lib, Shijia Sun, Maximilian Ammenwerth, Dominik S. Wild, J. Ignacio Cirac, Immanuel Bloch, Johannes Zeiher

arXiv:2611.18122v1 ↗PDF ↗Submitted 13 November 2026

Question
Phase estimation finds energies from the complex overlap \(g(t)=\langle\psi_0\rvert e^{-iHt}\lvert\psi_0\rangle\), whose Fourier transform is the spectrum weighted by the probe state. Measuring the return probability gives only \(\lvert g\rvert\); the phase normally needs a controlled time evolution with an auxiliary qubit, which an analog simulator cannot do. Can an analog Rydberg array recover the phase and so measure absolute many-body energies?
Result
Using strontium atoms in tweezers, the experiment recovers the phase of \(g(t)\) from two extra runs that each start with a small single-spin rotation imitating a short imaginary-time step. The Cauchy–Riemann equations turn the difference in their return amplitudes into the phase velocity. For chains of 5 and 19 atoms the reconstructed spectra give ground-state energies and the gap to the first excited band in agreement with exact numerics. Extending the method to every output bitstring yields energy-filtered observables, including the sign change of the drive energy from negative near the ground state to positive in excited bands, and density order resolved by energy in a \(5\times3\) array.

The paper in detail

Quantum phase estimation is the standard route to energies on a quantum computer, but in its textbook form it needs the Hamiltonian evolution to be controlled by an auxiliary qubit. Statistical variants use only the complex Loschmidt echo \(g(t)\). Its Fourier transform, smoothed by a Gaussian window of width \(\delta\), is the density of states seen by the probe, \(\rho_{\psi_0}(E)\propto\langle\psi_0\rvert e^{-(H-E)^2/(2\delta^2)}\lvert\psi_0\rangle\). The difficulty on an analog machine is that experiments measure probabilities, so only \(\lvert g(t)\rvert\) is directly accessible.

The array realizes \(H=\tfrac\Omega2\sum_iX_i-\Delta\sum_in_i+\sum_{i<j}V_{ij}n_in_j\) with nearest-neighbour interactions of \(2\pi\times35\) MHz, far above \(\Omega\approx2\pi\times3\) MHz, so neighbouring atoms cannot both be excited. The probe is the Néel state \(\lvert Z_2\rangle\), whose energy is known exactly because it is a product state, and the return probability is the fraction of shots that come back to the same bitstring after a quench.

The phase comes from analytic continuation. If the evolution could run for complex time \(t+i\tau\), the logarithm of \(g\) would be analytic, and its real and imaginary parts, \(\log\lvert g\rvert\) and the phase \(\varphi\), would obey the Cauchy–Riemann equations. A short imaginary-time step \(e^{\mp H\tau}\) applied to a product state is, to first order and under the blockade, a small \(Y\) rotation of each excited atom. Two extra branches with rotations \(\pm\tau\), followed by the same analog evolution, therefore give \(\lvert g(t\mp i\tau)\rvert\), and their log-ratio is the phase velocity. Integrating from \(\varphi(0)=0\) gives the full echo.

The spectra show a sharp ground-state peak and a first excited band. Their positions match exact numerics at \(\Delta/\Omega=1.75\), \(2.25\) and \(2.75\) for 5 and 19 atoms, and the ground state lies below \(E_{Z_2}\) by the amount quantum fluctuations predict. Changing the probe to the complementary Néel state shifts weight to higher bands. Reconstructing the amplitude of every bitstring gives an energy-filtered state from which observables can be computed: the drive term \((\Omega/2)\langle O_X\rangle_E\) lowers the energy near the ground state and raises it in excited bands, a coherence that populations cannot show. The same holds for staggered magnetization and, in a \(5\times3\) lattice, for checkerboard and stripe density modes.

What limits the method is statistics and the approximations behind the rotation trick. Longer evolution and faster repetition would sharpen the spectra toward the critical region, and suitable probes would scan other windows. Open targets named by the authors include frustrated kagome arrays, where quantum fluctuations decide between competing orders, and pseudogap physics in Fermi–Hubbard simulators.

Transferable methods

  1. Phase from magnitudes by analytic continuation

    The missing phase of the echo is recovered without any controlled evolution by measuring magnitudes at slightly complex times. Because \(\log g\) is analytic in the complex time plane, how its magnitude changes in imaginary time fixes how its phase changes in real time.

    Write \(g(t+i\tau)=r(t,\tau)e^{i\varphi(t,\tau)}\). Away from zeros of \(g\), \(\log g\) is analytic, and the Cauchy–Riemann equations give \(\partial_\tau\log r=-\partial_t\varphi\). A central difference in \(\tau\) yields \[\dot\varphi(t)=\frac{\log\lvert g(t-i\tau)\rvert-\log\lvert g(t+i\tau)\rvert}{2\tau}+O(\tau^2).\] Editorial check on a five-site blockaded chain at \(\Delta=\Omega\) with \(\tau=10^{-3}\): the estimate matches the derivative of the exact phase to within \(5\times10^{-4}\) over \(0<\Omega t<6\).

    The imaginary step must be implemented. Split \(H=H_c+H_q\) into the diagonal part and the drive; for a product probe \(H_c\lvert\psi_0\rangle=\zeta_c\lvert\psi_0\rangle\), so \[e^{-H\tau}\lvert\psi_0\rangle=e^{-\zeta_c\tau}e^{-H_q\tau}\lvert\psi_0\rangle+O(\tau^2).\] Under the blockade only excited atoms can flip, and \(e^{-H_q\tau}\) on them is a small \(Y\) rotation, which is unitary up to normalization. The known normalizations restore the magnitudes.

    How to reuse it

    Any platform that can measure return probabilities and apply a cheap approximation to a short imaginary-time step can recover phases this way, provided the echo stays away from zero (integration fails across zeros) and the step's error is controlled. It is a general alternative to Hadamard tests when controlled evolution is unavailable.

  2. Energy filtering as post-processing

    With the complex amplitudes \(g_b(t)=\langle b\rvert e^{-iHt}\lvert\psi_0\rangle\) for all bitstrings, a Gaussian filter in time becomes a projection in energy, applied classically after the experiment. Expectation values in an energy window then follow without preparing eigenstates.

    Multiplying by a Gaussian in time and Fourier transforming gives \[c_b(E)\propto\int dt\,e^{iEt}e^{-\delta^2t^2/2}g_b(t)=\langle b\rvert e^{-(H-E)^2/(2\delta^2)}\lvert\psi_0\rangle\cdot\text{const},\] so \(\lvert\psi_{E,\delta}\rangle=\sum_bc_b(E)\lvert b\rangle\) is the probe filtered to a window of width \(\delta\) around \(E\). Diagonal observables need only \(\lvert c_b\rvert^2\); the drive \(O_X\) needs relative phases.

    Why the drive energy changes sign (editorial gloss on the paper's argument): by Hellmann–Feynman, \(\langle O_X\rangle_n=2\,\partial E_n/\partial\Omega\). The ground state is pushed down by the drive, so its value is negative, and since \(O_X\) is traceless the values over all eigenstates sum to zero, so some excited states must be positive. Editorial check for five atoms at \(\Delta=\Omega\): \(\langle O_X\rangle=-2.19\) in the ground state and a sum of zero to \(10^{-15}\).

    For generalized echoes the integration constant \(\varphi_b(0^+)\) is fixed from perturbation theory in the PXP model, and the rotation trick leaves an \(O(\tau)\) bias for amplitudes that start at zero. For one atom with classical energy offset \(e\), the integrated bias tends to \(\pi e\tau/4\), which the authors remove as a constant offset.

    How to reuse it

    Spectral filtering of measured amplitudes gives microcanonical-like expectation values and, with Boltzmann reweighting, finite-temperature ones. The cost is shots: every bitstring whose amplitude matters must be sampled well enough to fix its phase, so the approach suits small or strongly constrained Hilbert spaces.

Source: Main text, Eqs. (1)–(2) and Figs. 1–3; Supplementary Information D (Eqs. (D1)–(D4)) and E (Eqs. (E10)–(E17)). Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. It is the first statistical phase estimation on an analog simulator, done with native evolution and no auxiliary qubits or controlled dynamics. Because the evolution is analog, long times and fine energy resolution cost nothing in circuit depth, which matters where tensor-network methods lose accuracy, as in two dimensions.

Limitations. The local rotations reproduce imaginary-time evolution only to first order, and only because the blockade freezes the neighbours of excited atoms in the Néel probe; in two dimensions this is a further approximation. Generalized echoes need a phase anchor fixed analytically within the PXP model and an \(O(\tau)\) bias correction, and reliable relative phases require many shots per bitstring, which does not scale. Resolution was limited by the shot budget, so individual eigenstates within a band were not resolved, and the method works only for probe states whose classical energy is known.

Description based on the full arXiv HTML of v1, read 2026-11-16 (main text, Figs. 1–5), and Supplementary Sections D–E from the PDF (the Cauchy–Riemann derivation, the local-rotation approximation and the bias correction); Sections A–C were skimmed. Editorial checks on a 5-site blockaded chain: 13 allowed states, the Cauchy–Riemann estimate matching the exact phase velocity to \(5\times10^{-4}\), the perturbative shift \(-(N+1)\Omega^2/(8\Delta)\) within 2% of exact diagonalization at \(\Delta/\Omega=8\), and the drive term \(\langle O_X\rangle\) negative in the ground state and summing to zero over eigenstates.