188
Part Five · Convergence

Phase estimation for chemistry: the fault-tolerant route

With an error-corrected quantum computer, the method of choice for chemistry is not VQE but phase estimation (Chapter 113): it measures an energy eigenvalue directly, to precision limited only by run time, with no variational optimization. It has a catch that is easy to miss. Phase estimation returns the ground-state energy only with probability equal to the squared overlap between your starting state and the true ground state. If that overlap is tiny, you will mostly measure something else.

On the plate: how the starting overlap decides the cost.

Qubitization and factorized Hamiltonians

Modern chemistry algorithms do not Trotterize. They block-encode the Hamiltonian (Chapter 119) as a linear combination of unitaries, H=∑ℓcℓUℓH = \sum_\ell c_\ell U_\ell, and build a “qubitized” walk operator whose eigenphases are arccos⁡(Ek/λ)\arccos(E_k/\lambda), where λ=∑∣cℓ∣\lambda = \sum|c_\ell| is the Hamiltonian’s 1-norm. Phase estimation on the walk needs about πλ/(2ε)\pi\lambda/(2\varepsilon) steps for energy precision ε\varepsilon. So the cost is driven by λ\lambda, and much of the progress since 2017 came from rewriting the Coulomb tensor to shrink it: single factorization, double factorization (Mario Motta, Joonho Lee and colleagues; Vera von Burg and colleagues, 2021) and tensor hypercontraction (Joonho Lee, Dominic Berry and colleagues, 2021). Established

The overlap problem

Prepare a trial state ∣ϕ⟩=∑kγk∣Ek⟩|\phi\rangle = \sum_k\gamma_k|E_k\rangle. Each run of phase estimation projects onto ∣E0⟩|E_0\rangle with probability ∣γ0∣2|\gamma_0|^2, so you need about 1/∣γ0∣21/|\gamma_0|^2 runs to see the ground state once. For a molecule built from fragments, a product of good local approximations has overlap that shrinks exponentially with size: ∣γ0∣≈γ1 n|\gamma_0|\approx\gamma_1^{\,n}, the orthogonality catastrophe. Set the per-fragment overlap 0.950 and the number of fragments 20. For many molecules of interest, a well-chosen multi-configurational state from DMRG or selected CI gives a usable overlap; the question is whether that remains true precisely where classical methods fail (Chapter 183). Debated

Cost comparison

For FeMoco, the published logical cost fell as factorizations improved: Markus Reiher and colleagues’ 2017 Trotter-based estimate needed of order 101410^{14} T gates; Lee and colleagues’ 2021 tensor-hypercontraction estimate needed about 5.3×1095.3\times10^9 Toffoli gates and 2,142 logical qubits. Chapter 189 traces these numbers and turns them into physical hardware. The algorithmic side is mature; state preparation and hardware are the open pieces. Established for the algorithm; Debated for state preparation at scale.

What you now have

Block encodings and qubitized phase estimation, the role of the 1-norm λ, single, double and tensor-hypercontraction factorizations, the ground-state overlap requirement, the orthogonality catastrophe, and how logical costs for FeMoco have fallen.

· · ·

Three questions to carry with you.

What sets qubitized phase-estimation cost?tap
About πλ/2ε walk steps, where λ is the Hamiltonian’s 1-norm.
How many runs does an overlap γ₀ require?tap
About 1/|γ₀|².
What is the orthogonality catastrophe?tap
Overlaps of approximate product states with the true ground state shrink exponentially with system size.

And one to work out. With a ground-state overlap ∣γ0∣=0.3|\gamma_0| = 0.3, about how many phase-estimation runs are needed to see the ground state once on average?

runs ≈

Sources: G. H. Low and I. L. Chuang, Quantum 3, 163 (2019). D. W. Berry et al., Quantum 3, 208 (2019). V. von Burg et al., Phys. Rev. Research 3, 033055 (2021). J. Lee et al., PRX Quantum 2, 030305 (2021). M. Reiher et al., Proc. Natl Acad. Sci. USA 114, 7555 (2017). S. Lee et al., Nat. Commun. 14, 1952 (2023).

Next · Chapter 189
Resource estimation: how many qubits and gates real problems need

FeMoco's Toffoli count fell from about 10¹⁴ to about 5×10⁹ in four years ✓. How?

Turn the page