HeH⁺ PES

BasQ Qiskit Fall Fest 2026 · Challenge 1 · Beginner track

The simplest molecule with a lopsided bond, from exact maths to a noisy quantum chip.

HeH⁺ is a helium atom and a proton sharing two electrons. We trace its energy as the two nuclei are pulled apart, using the variational quantum eigensolver (VQE), and check every step: exact references, a noiseless quantum simulation, a realistic noisy one, and real hardware. This page is a short visual tour; the notebook has the full story.

4 → 2qubits after removing symmetries, with no loss of accuracy
3×10⁻⁷ Haworst noiseless VQE error, 5000× below chemical accuracy
45 → 2two-qubit gates on an IBM Heron chip
2.1 mHanoisy error at equilibrium after zero-noise extrapolation

Visuals on the left, the exact notebook code on the right. 1 Ha (Hartree) ≈ 27.2 eV; chemical accuracy = 1.6 mHa.

Section 1

Two qubits are enough

Two electrons in two orbitals, each with a spin-up and a spin-down slot, is 4 qubits. But the number of spin-up and spin-down electrons never changes, so two qubits only ever store values we already know. We delete them.

↑ orbital 1↑ orbital 2↓ orbital 1↓ orbital 2 tapering qubit A qubit B the deleted qubits stored "1 spin-up electron" and "2 electrons in total": always true
  • 27 → 9 Pauli terms to measure, same exact energy.
  • H₂ goes to 1 qubit, HeH⁺ stops at 2. H₂ is symmetric (two identical atoms), which is one more symmetry to remove. HeH⁺ is not.

Section 2

Trust, but verify

Before believing any quantum number, we compared Qiskit's exact answer with an independent chemistry code (PySCF) at every distance. They agree to 10⁻¹⁵ Ha. We also caught a trap from the tutorial.

The trap: an unfiltered solver searches states with any number of electrons. For HeH⁺ it returns a 3-electron state, 0.164 Ha too low. We filter on electron number and spin.
Agreement: PySCF FCI vs Qiskit (2-qubit) at all 14 distances: ≤ 3×10⁻¹⁵ Ha.

Section 3

The energy curve

Squeeze the nuclei and they repel; pull them apart and the bond breaks. The lowest point is the bond length. VQE (dots) lands exactly on the exact curve everywhere; Hartree-Fock, which ignores how electrons dodge each other, sits slightly above.

HeH+ potential energy curve in STO-3G -2.80 -2.70 123 He–H distance (Å) Energy (Ha) minimum 0.914 Å
Hartree-FockExactVQE (2 qubits)
  • Breaks into He + H⁺: helium holds its electrons far more tightly than hydrogen does.
  • Bond energy 1.49 eV in this small basis (real value 2.04 eV).

Section 4

The real error is the basis set

STO-3G describes each atom with the bare minimum of functions. Bigger basis sets move the bond length from 0.91 Å to the true 0.774 Å. That error is about 100,000× larger than VQE's, so the quantum algorithm is not the bottleneck here.

Bond length vs basis set size 0.80 0.85 0.90 204060 spatial orbitals in the basis Bond length (Å) literature 0.774 Å
FCI bond length per basis

STO-3G 0.914 Å · 6-31G 0.769 Å · 6-31G** 0.776 Å · cc-pVDZ 0.790 Å · cc-pVTZ 0.776 Å · cc-pVQZ 0.775 Å. cc-pVQZ would need 118 qubits; classically it takes seconds.

Optimisers under shot noise

On hardware every energy is a noisy estimate. Worst error from 6 noisy runs:

SLSQP 1302 mHa gradients drown in noise
COBYLA 5.5 mHa
SPSA 9.0×10^-1 mHa 600 evaluations
chemical accuracy (1.6 mHa) · log scale

Section 5

Fitting it onto a real chip

On IBM's heavy-hex chips, qubits only talk to neighbours, so circuits get rewritten. The tapered circuit compiles to a handful of gates; the 4-qubit version balloons. Then we pick the physical qubits with the best calibration.

4 qubits (JW) 45.0 CZ
2 qubits (tapered) 2.0 CZ
noise-aware (best) 7.7 mHa [22, 23]
worst qubits 17.6 mHa [113, 119]

Section 6

Removing noise by adding more

Zero-noise extrapolation: run the circuit with its noise deliberately tripled and quintupled, see how the error grows, and draw the line back to zero noise.

Zero-noise extrapolation at 0.774 Å 0 10 20 30 012345 noise scale factor error vs exact (mHa) chemical accuracy
measuredstraight-line fit → zero noise
Noisy error along the curve 0 5 10 11.522.5 He–H distance (Å) error (mHa) chemical accuracy
raw noisyafter ZNE

Section 7

Ready for real hardware

The whole IBM job is built: 3 distances × 3 noise levels in one submission, 32,768 shots each. We did not get real-QPU time, so we ran the identical job on the fake device with real measurements. That adds two things the exact-average simulation hides: readout error and shot noise.

dry run error with shot-noise error bars 0 5 10 15 11.522.5 He–H distance (Å) error vs exact (mHa) chemical accuracy exact
dry run rawdry run + ZNEgate noise only + ZNE (Section 6)
  • Readout error adds ~3 mHa, and ZNE can't remove it: folding amplifies gate noise, not measurement errors. On a real device TREX readout mitigation is designed to handle that part.
  • Shot noise sets the error bars (±2–6 mHa). Confirming chemical accuracy at equilibrium would need ~12× more shots.
  • One command away: RUN_HARDWARE = True submits the identical job to ibm_fez.

Section 8

Where the error comes from

Every error source at the equilibrium distance, side by side on a log scale. The algorithm is essentially perfect; noise and shot noise cost a few mHa each; the small basis costs more than everything else together.

Basis set (STO-3G → 6-31G) 80.9 mHa
Gate + readout noise (dry run) 11.4 mHa
Gate noise (simulation) 7.7 mHa
After ZNE (dry run) 6.5 mHa
Shot noise (32,768 shots) 4.5 mHa
After ZNE (simulation) 2.1 mHa
VQE algorithm 5.2×10^-5 mHa
chemical accuracy (1.6 mHa) · log scale

Takeaways

Small molecule, honest numbers.

  1. Symmetry first. Removing two known qubits turned 45 two-qubit gates into 2, the difference between 121 mHa and 7.7 mHa of error.
  2. VQE is exact when noiseless, so the basis set, not the algorithm, limits accuracy.
  3. Mitigation works, partly. A good qubit choice and ZNE bring gate noise to 1.4–2.6 mHa, near chemical accuracy. Readout error (~3 mHa) needs its own fix.
  4. Shot noise is the next wall. At 32,768 shots one energy wobbles by ~4.5 mHa; optimising on hardware would also need noise-robust optimisers such as SPSA.

Read the full notebook Source on GitHub