Simple Algorithm
Quantum Computing · Excited State · 24 min

Quantum Chemistry & Materials

The most natural application: simulating molecules and materials.

In 1929 Paul Dirac wrote that the laws needed for "the whole of chemistry" were completely known, and that the only difficulty was that their exact application leads to equations much too complicated to be soluble. Nearly a century later that difficulty is still the bottleneck, and it is exactly the kind of complexity a quantum computer stores natively: MM electron orbitals fit into MM qubits.

What you’ll learn

  • State the electronic structure problem in the Born–Oppenheimer approximation and explain the target of chemical accuracy.
  • Write the second-quantized molecular Hamiltonian and map it to a sum of Pauli strings with the Jordan–Wigner transformation.
  • Explain why strongly correlated systems such as FeMoco and the cuprates defeat standard classical methods.
  • Compare VQE and quantum phase estimation: what each needs, how its cost scales with precision, and what it guarantees.
  • Weigh the caveats: initial-state overlap, the QMA-hardness of general ground-state problems, and improving classical methods.

The electronic structure problem

Almost all of chemistry, from why water is bent to how an enzyme splits N2\mathrm{N_2}, is decided by how electrons arrange themselves around nuclei. A proton is about 1836 times heavier than an electron, so electrons adjust almost instantly to wherever the nuclei happen to be. The Born–Oppenheimer approximation freezes the nuclei at positions RA\mathbf R_A and asks only for the electronic ground state of the hamiltonian

Hel=−12∑i∇i2−∑i,AZA∣ri−RA∣+∑i<j1∣ri−rj∣H_{\text{el}}=-\frac12\sum_i\nabla_i^2-\sum_{i,A}\frac{Z_A}{\abs{\mathbf r_i-\mathbf R_A}}+\sum_{i<j}\frac{1}{\abs{\mathbf r_i-\mathbf r_j}}
Atomic units (ℏ=me=e=1\hbar=m_e=e=1, energies in hartree): electron kinetic energy, electron–nucleus attraction, electron–electron repulsion. The nuclear repulsion ∑A<BZAZB/∣RA−RB∣\sum_{A<B}Z_AZ_B/\abs{\mathbf R_A-\mathbf R_B} is a constant once the nuclei are fixed.

Its lowest eigenvalue E0(R)E_0(\mathbf R), scanned over nuclear positions, is the potential energy surface: stable geometries are its minima and reaction barriers its saddle points.

To compute, expand the wavefunction in MM spin-orbitals (one-electron functions). An NN-electron state is then a superposition of Slater determinants, antisymmetrized products that specify which NN of the MM spin-orbitals are occupied. Solving exactly in this basis is called full configuration interaction (FCI), and it requires (MN)\binom{M}{N} amplitudes.

Predict before you look

How many Slater determinants are there for N=50N=50 electrons in M=100M=100 spin-orbitals? Give the power of ten, that is, log⁡10\log_{10} of the count.

Show the answer

29 (power of 10)

(10050)≈1.0×1029\binom{100}{50}\approx1.0\times10^{29}. Fixing the spin (25 up and 25 down electrons in 50 orbitals each) still leaves (5025)2≈1.6×1028\binom{50}{25}^2\approx1.6\times10^{28}. A modest molecule in a modest basis outstrips any conceivable classical memory, while a quantum register would hold the same state in 100 qubits.

Second quantization: orbitals become qubits

Chemists keep track of determinants with second quantization. Each spin-orbital pp gets a creation operator ap†a_p^\dagger and an annihilation operator apa_p. A determinant is just a list of occupation numbers ∣n0n1⋯nM−1⟩\ket{n_0n_1\cdots n_{M-1}} with np∈{0,1}n_p\in\{0,1\}, and the antisymmetry of electrons is built into the algebra:

{ap,aq†}=δpq,{ap,aq}={ap†,aq†}=0\acomm{a_p}{a_q^\dagger}=\delta_{pq},\qquad\acomm{a_p}{a_q}=\acomm{a_p^\dagger}{a_q^\dagger}=0
The canonical anticommutation relations. In particular (ap†)2=0(a_p^\dagger)^2=0: the Pauli exclusion principle.
H=∑pqhpq ap†aq+12∑pqrshpqrs ap†aq†arasH=\sum_{pq}h_{pq}\,a_p^\dagger a_q+\frac12\sum_{pqrs}h_{pqrs}\,a_p^\dagger a_q^\dagger a_ra_s
The molecular Hamiltonian in second quantization. The one-electron integrals hpqh_{pq} (kinetic energy and nuclear attraction) and two-electron integrals hpqrs=∫dx1dx2 ϕp∗(x1)ϕq∗(x2)ϕr(x2)ϕs(x1)/∣r1−r2∣h_{pqrs}=\int d\mathbf x_1d\mathbf x_2\,\phi_p^*(\mathbf x_1)\phi_q^*(\mathbf x_2)\phi_r(\mathbf x_2)\phi_s(\mathbf x_1)/\abs{\mathbf r_1-\mathbf r_2} are computed classically, once. There are O(M4)\order{M^4} terms.

The occupation-number basis of MM spin-orbitals has 2M2^M states, exactly as many as MM qubits. So let qubit pp be ∣0⟩\ket0 when orbital pp is empty and ∣1⟩\ket1 when it is occupied. The catch: apa_p must anticommute with operators on other orbitals, while operators on different qubits commute. The Jordan–Wigner transformation repairs this with a string of Pauli ZZ's that records the parity of all earlier orbitals.

DerivationJordan–Wigner in four lines
  1. ap†=Z0Z1⋯Zp−1⊗Xp−iYp2,X−iY2=∣1⟩ ⁣⟨0∣a_p^\dagger=Z_0Z_1\cdots Z_{p-1}\otimes\frac{X_p-iY_p}{2},\qquad\frac{X-iY}{2}=\ketbra{1}{0}

    Filling orbital pp flips qubit pp from ∣0⟩\ket0 to ∣1⟩\ket1; the ZZ string supplies the fermionic sign (−1)n0+⋯+np−1(-1)^{n_0+\cdots+n_{p-1}}.

  2. np=ap†ap=∣1⟩ ⁣⟨1∣p=I−Zp2n_p=a_p^\dagger a_p=\dyad{1}_p=\frac{\id-Z_p}{2}

    The strings cancel because Z2=IZ^2=\id: occupation numbers are single-qubit operators, and so are Coulomb terms npnqn_pn_q (two-qubit).

  3. ap†aq+aq†ap=12(XpXq+YpYq)Zp+1⋯Zq−1(p<q)a_p^\dagger a_q+a_q^\dagger a_p=\tfrac12\left(X_pX_q+Y_pY_q\right)Z_{p+1}\cdots Z_{q-1}\qquad(p<q)

    Hopping between two orbitals acts on every qubit in between, so a single term can touch up to MM qubits.

  4. H=∑jcj Pj,Pj∈{I,X,Y,Z}⊗M,cj∈RH=\sum_j c_j\,P_j,\qquad P_j\in\{\id,X,Y,Z\}^{\otimes M},\qquad c_j\in\R

    Expanding every term yields the qubit Hamiltonian: a real combination of O(M4)\order{M^4} Pauli strings. Every quantum chemistry algorithm starts here.

Why strong correlation is hard

If electrons moved independently, one determinant would suffice. Hartree–Fock (mean field) finds the best single determinant. What it misses is the correlation energy: a tiny fraction of the total energy, yet typically larger than the reaction energies chemists care about.

Correlation comes in two flavors. Dynamic correlation, electrons dodging one another, is captured well by perturbation theory and coupled cluster built on one dominant determinant. Static (strong) correlation appears when several determinants are nearly degenerate: stretched bonds, transition-metal clusters with many partially filled dd orbitals, doped Mott insulators. Then there is no good single starting point.

MethodCostWorks well forStruggles with
Hartree–FockO(M4)\order{M^4}orbitals, rough geometriesall correlation
CCSD(T), the "gold standard"O(M7)\order{M^7}weakly correlated moleculesbond breaking, multireference systems
DMRG (tensor networks)polynomial at fixed bond dimension1D-like systems; active spaces up to roughly a hundred orbitalshighly entangled 2D/3D systems
Quantum Monte Carlopolynomial per samplebosons, sign-problem-free modelsthe fermion sign problem (needs approximate constraints)
Full CI (exact)exponentialactive spaces of about 20 orbitalsanything much larger
Classical electronic-structure methods in brief. Each fails somewhere; strongly correlated, high-entanglement systems defeat all of them.
HHubbard=−t∑⟨ij⟩,σ(ciσ†cjσ+cjσ†ciσ)+U∑ini↑ni↓H_{\text{Hubbard}}=-t\sum_{\langle ij\rangle,\sigma}\left(c_{i\sigma}^\dagger c_{j\sigma}+c_{j\sigma}^\dagger c_{i\sigma}\right)+U\sum_i n_{i\uparrow}n_{i\downarrow}
Hopping tt between neighboring lattice sites and on-site repulsion UU. One line to write down, yet the most-studied open problem in correlated materials.

Route 1: the variational eigensolver

The variational quantum eigensolver (VQE) uses the quantum processor only to prepare a parameterized trial state ∣ψ(θ)⟩\ket{\psi(\boldsymbol\theta)} and to measure its energy; a classical optimizer updates θ\boldsymbol\theta. It rests on the Rayleigh–Ritz variational principle.

DerivationThe variational principle
  1. ∣ψ⟩=∑kck∣Ek⟩,H∣Ek⟩=Ek∣Ek⟩,E0≤E1≤⋯\ket\psi=\sum_kc_k\ket{E_k},\qquad H\ket{E_k}=E_k\ket{E_k},\qquad E_0\le E_1\le\cdots

    Expand any normalized trial state in the energy eigenbasis.

  2. ⟨ψ|H|ψ⟩=∑k∣ck∣2Ek ≥ E0∑k∣ck∣2=E0\mel{\psi}{H}{\psi}=\sum_k\abs{c_k}^2E_k\ \ge\ E_0\sum_k\abs{c_k}^2=E_0

    The expectation value is a weighted average of eigenvalues, so it can never fall below the smallest one.

  3. min⁡θ⟨ψ(θ)|H|ψ(θ)⟩ ≥ E0,with equality iff the ansatz reaches a ground state\min_{\boldsymbol\theta}\mel{\psi(\boldsymbol\theta)}{H}{\psi(\boldsymbol\theta)}\ \ge\ E_0,\quad\text{with equality iff the ansatz reaches a ground state}

    Minimizing over an ansatz gives an upper bound on E0E_0, tight only if the circuit family is expressive enough.

Try it on a one-qubit toy with the shape of tapered H₂: H=0.8 Z+0.6 XH=0.8\,Z+0.6\,X. Read ∣1⟩\ket1 as the Hartree–Fock determinant (diagonal energy −0.8-0.8) and ∣0⟩\ket0 as a doubly excited configuration; the XX term couples them. The exact ground energy is −1-1.

Predict before you look

The exact ground state of H=0.8Z+0.6XH=0.8Z+0.6X mixes ∣0⟩\ket0 and ∣1⟩\ket1. What percentage of its weight sits on the Hartree–Fock-like configuration ∣1⟩\ket1?

Show the answer

90 %

For E=−1E=-1, the first row of (H−E)v=0(H-E)v=0 reads 1.8a+0.6b=01.8a+0.6b=0, so b=−3ab=-3a and the normalized ground state is (∣0⟩−3∣1⟩)/10(\ket0-3\ket1)/\sqrt{10}, with ∣b∣2=0.9\abs{b}^2=0.9. The single determinant ∣1⟩\ket1 has energy −0.8-0.8; mixing in 10% of the other configuration lowers it to −1-1. That −0.2-0.2 is this toy's correlation energy.

Route 2: phase estimation

The fault-tolerant route reads the energy off directly. An eigenstate of HH with energy EE is an eigenstate of U=e−iHτU=e^{-iH\tau} with eigenvalue e−iEτe^{-iE\tau}, and phase estimation writes the phase φ=−Eτ/2π(mod1)\varphi=-E\tau/2\pi\pmod 1 into a register of counting qubits. Started from a trial state ∣ψ⟩=∑kck∣Ek⟩\ket\psi=\sum_kc_k\ket{E_k}, it returns EkE_k with probability ∣ck∣2\abs{c_k}^2 and leaves the system in that eigenstate.

First you must build UU from gates. For H=∑jHjH=\sum_jH_j with non-commuting terms, the simplest tool is Trotterization: e−iHτ≈(∏je−iHjτ/r)re^{-iH\tau}\approx\bigl(\prod_je^{-iH_j\tau/r}\bigr)^r.

Predict before you look

Choose τ\tau so that a window of 1 hartree fills the whole phase range [0,1)[0,1). How many counting qubits tt make the resolution 2−t2^{-t} finer than chemical accuracy, 0.00160.0016 hartree?

Show the answer

10

2−9≈0.00202^{-9}\approx0.0020 is too coarse; 2−10≈0.000982^{-10}\approx0.00098 suffices, so t=10t=10. The controlled powers U2kU^{2^k} then need a total evolution time of about 210τ2^{10}\tau. Phase estimation's cost grows like 1/ε1/\varepsilon (Heisenberg scaling), versus 1/ε21/\varepsilon^2 samples for sampling-based energy estimates.

AspectVQEPhase estimation
Hardwarenear-term: shallow circuits, many repetitionsfault-tolerant: long coherent circuits
Cost vs precision ε\varepsilonO(1/ε2)\order{1/\varepsilon^2} samples per energyO(1/ε)\order{1/\varepsilon} total evolution time
What you geta variational upper bound, limited by the ansatzan eigenvalue to precision ε\varepsilon, by projective measurement
What it needsan expressive yet trainable ansatz; error mitigationa trial state with large overlap ∣c0∣2\abs{c_0}^2
Main riskmeasurement cost, noise, barren plateausmillions of physical qubits and long runtimes

The honest picture

Chemistry is the most natural application, not a guaranteed one.

  • Complexity. Estimating ground-state energies of general local Hamiltonians is QMA-complete (Kitaev), so no one expects quantum computers to solve every ground-state problem efficiently. Phase estimation needs a trial state with non-negligible overlap, and for large systems simple trial states can have overlap that shrinks exponentially with system size.
  • Competition. Classical methods keep improving. Lee et al. (2023) reviewed the evidence and found no clear case yet for an exponential quantum advantage in ground-state chemistry, while polynomial speedups for hard, strongly correlated cases remain plausible.
  • Resources. Careful fault-tolerant estimates for FeMoco-class problems call for millions of physical qubits running for days.
  • Beyond ground states. Dynamics, spectra and finite-temperature properties of correlated materials may be cleaner targets, since there is no state-preparation bottleneck of the same kind and classical methods struggle with long-time dynamics.
Think first If ground-state energy estimation is QMA-complete, why is chemistry still considered the most promising application?

QMA-completeness is a worst-case statement over all local Hamiltonians, including contrived ones whose ground states are hard to reach even for nature. Molecules are physical systems that do relax to their ground states, chemists can often supply good trial states, and the quantum algorithm then needs only polynomial resources. The open question is whether hard-for-classical, easy-for-quantum instances are common in practice.

Key takeaways

  • Born–Oppenheimer reduces chemistry to the electronic ground-state problem; the target is chemical accuracy, about 1.6 mHa in energy differences.
  • Second quantization maps MM spin-orbitals to MM qubits; Jordan–Wigner turns the Hamiltonian into O(M4)\order{M^4} Pauli strings, with ZZ strings enforcing fermionic signs.
  • Strong (static) correlation, as in FeMoco or doped Hubbard models, defeats single-reference classical methods, and exact diagonalization is exponential.
  • VQE gives a variational upper bound on near-term hardware but pays O(1/ε2)\order{1/\varepsilon^2} in samples; phase estimation gives eigenvalues at O(1/ε)\order{1/\varepsilon} cost but needs fault tolerance and a trial state with good overlap.
  • Advantage is plausible but not guaranteed: general ground-state problems are QMA-hard, and classical methods keep improving.

Check yourself

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

Question 1

A molecule is described with MM spin-orbitals and NN electrons. How many system qubits does the Jordan–Wigner encoding use?

  1. NN
  2. ⌈log⁡2(MN)⌉\lceil\log_2\binom{M}{N}\rceil
  3. MM
  4. M4M^4
Show the answer

C. MM

Jordan–Wigner assigns one qubit per spin-orbital, storing its occupation number, so it uses MM qubits for all particle numbers at once.

Question 2

Why is FeMoco a hard case for classical electronic-structure methods?

  1. It has too many atoms for Hartree–Fock
  2. The Born–Oppenheimer approximation fails for iron
  3. Relativistic effects dominate its chemistry
  4. Many nearly degenerate open-shell iron dd orbitals make many determinants important at once
Show the answer

D. Many nearly degenerate open-shell iron dd orbitals make many determinants important at once

FeMoco is strongly (statically) correlated: no single determinant dominates, so single-reference methods like CCSD(T) are unreliable, and its active space (54 electrons in 54 orbitals in Reiher et al.) is far beyond exact diagonalization.

Question 3

A VQE experiment reports an energy below the exact full-CI ground energy of the same Hamiltonian. What is the most likely explanation?

  1. The ansatz found a state below the ground state
  2. Statistical (shot-noise) or error-mitigation error in the estimated energy
  3. The variational principle fails for strongly correlated molecules
  4. The Pauli terms do not commute, so the energy is unbounded below
Show the answer

B. Statistical (shot-noise) or error-mitigation error in the estimated energy

For any physical state ⟨ψ|H|ψ⟩≥E0\mel{\psi}{H}{\psi}\ge E_0. An estimate below E0E_0 must come from how the number was estimated: finite sampling, or a mitigation technique such as extrapolation overshooting.

Question 4

You need ten times better energy precision. Roughly how do the costs change for sampling-based VQE energy estimation and for phase estimation?

  1. VQE: about 100× more samples; QPE: about 10× more evolution time
  2. Both about 10×
  3. Both about 100×
  4. VQE: about 10× more samples; QPE: about 100× more evolution time
Show the answer

A. VQE: about 100× more samples; QPE: about 10× more evolution time

Sampling error falls like 1/S1/\sqrt{S}, so S∝1/ε2S\propto1/\varepsilon^2. Phase estimation is Heisenberg-limited: precision ε\varepsilon needs total evolution time ∝1/ε\propto1/\varepsilon.

Question 5

You run phase estimation on e−iHτe^{-iH\tau} starting from ∣ψ⟩=∑kck∣Ek⟩\ket\psi=\sum_kc_k\ket{E_k}. With what probability does it return (an estimate of) the ground energy E0E_0?

  1. ∣c0∣\abs{c_0}
  2. Always 1, since phase estimation finds the lowest eigenvalue
  3. ∣c0∣2\abs{c_0}^2
  4. 1/2t1/2^t for tt counting qubits
Show the answer

C. ∣c0∣2\abs{c_0}^2

Phase estimation measures in the energy eigenbasis, so by the Born rule it lands on E0E_0 with probability ∣c0∣2\abs{c_0}^2 (times a resolution factor close to 1). A poor trial state means many repetitions, and an exponentially small overlap means exponentially many.

Go deeper

Further reading

  • S. McArdle, S. Endo, A. Aspuru-Guzik, S. C. Benjamin & X. Yuan, "Quantum computational chemistry", Rev. Mod. Phys. 92, 015003 (2020).
  • A. Aspuru-Guzik, A. D. Dutoi, P. J. Love & M. Head-Gordon, "Simulated quantum computation of molecular energies", Science 309, 1704 (2005).
  • A. Peruzzo et al., "A variational eigenvalue solver on a photonic quantum processor", Nat. Commun. 5, 4213 (2014).
  • S. Lee et al., "Evaluating the evidence for exponential quantum advantage in ground-state quantum chemistry", Nat. Commun. 14, 1952 (2023).