Case Study: The Measurement Budget That Sinks VQE

Executive Summary

VQE is usually presented as depth-limited: the circuits are shallow, so the problem is fitting the ansatz within coherence. That framing is incomplete and hides the constraint that actually binds. VQE's dominant cost is measurement, and the measurement bill grows in a way that no hardware improvement addresses.

This case study computes the bill for a specific molecule, applies every standard optimization to reduce it, and finds that even after a 40× improvement the calculation takes weeks — for a molecule that a laptop solves in under a second.

Skills applied

• Decomposing a molecular Hamiltonian into Pauli terms (§19.6).
• Computing shot requirements from a target precision (§19.9).
• Applying commuting-group measurement reduction (§19.10).
• Assessing an algorithm's cost against its classical baseline.

Phase 1: The molecule and its Hamiltonian

BeH₂ in a minimal basis, 6 spatial orbitals → 12 spin orbitals → 12 qubits under Jordan–Wigner.

The Hamiltonian expands to 666 Pauli terms:

$$H = \sum_{k=1}^{666} c_k P_k$$

with coefficients ranging over several orders of magnitude and $\sum_k|c_k| \approx 22$ Hartree.

Target precision: chemical accuracy, 1.6 mHa. Below that threshold the calculation cannot distinguish reaction pathways, which is the entire point of doing it.

Phase 2: The naive shot count

Each Pauli term's expectation is estimated independently. For a term with $|c_k|$ and variance at most 1, the shots needed to reach total error $\varepsilon$ when allocating optimally across terms is approximately

$$N_{\text{shots}} \approx \frac{\left(\sum_k |c_k|\right)^2}{\varepsilon^2}$$

$$N_{\text{shots}} \approx \frac{22^2}{(0.0016)^2} = \frac{484}{2.56\times10^{-6}} \approx 1.9\times10^{8}$$

190 million shots for a single energy evaluation.

And VQE evaluates the energy repeatedly. With ~30 parameters and SPSA needing ~2 evaluations per iteration over ~300 iterations:

$$1.9\times10^{8} \times 600 \approx 1.1\times10^{11} \text{ shots}$$

At 10,000 shots per second — generous, ignoring queue and reset overhead — that is ~360 years.

Phase 3: Every optimization, applied

1. Commuting-group measurement (≈8× reduction). Terms that commute qubit-wise share a measurement basis. The 666 terms collapse to roughly 80 measurement settings. This does not reduce shots proportionally — variance still governs — but it removes redundant state preparations and improves the constant substantially. Call it 8×.

2. Optimal shot allocation (≈2×). Allocate shots proportional to $|c_k|$ rather than uniformly. Terms with tiny coefficients need almost no shots. Roughly halves the total.

3. Active-space reduction (≈4×). Freeze core orbitals and restrict to a chemically relevant active space. 12 qubits → 8 qubits, and term count falls from 666 to ~180 with $\sum|c_k| \approx 11$. Since cost goes as $(\sum|c_k|)^2$, this is a 4× saving — and it is the largest single lever.

4. Better optimizer (≈2.5×). SPSA with a good learning schedule, warm-started from a classical Hartree–Fock or CCSD guess, converges in ~120 iterations rather than 300.

Cumulative: ~160×.

$$\frac{1.1\times10^{11}}{160} \approx 6.9\times10^{8} \text{ shots} \approx \textbf{19 hours}$$

Nineteen hours of continuous device time — and this assumes no queue, no recalibration, and no repeated runs for error bars. Realistically several days to a week.

Phase 4: The classical baseline

BeH₂ in a minimal basis:

The quantum calculation is roughly $10^{5}$ times slower and less accurate.

Phase 5: How the cost scales

The forgiving reading is that this is a small-molecule artifact and the crossover comes later. Check that.

The term count grows as $O(N^4)$ and $\sum|c_k|$ grows with it, so shot cost grows roughly as $O(N^8)$ in the worst case. The scaling is polynomial but with an exponent high enough to be prohibitive.

Meanwhile the classical crossover — where full CI becomes infeasible — sits around 20–24 orbitals, and DMRG or selected-CI push it considerably further. VQE reaches multi-week runtimes well before the classical methods run out.

Phase 6: What this means

VQE's problem is not depth. Better qubits and deeper circuits do not fix a measurement bill that scales as $O(N^8)$ with a $1/\varepsilon^2$ prefactor. Shot count is set by statistics, not by hardware quality — the one term in the budget that hardware improvement does not touch.

The research directions that address it: - Classical shadows and randomized measurement schemes, reducing the number of distinct settings. - Better Hamiltonian factorizations (low-rank, tensor hypercontraction) that shrink $\sum|c_k|$ — attacking the quantity that is squared. - Abandoning VQE for QPE at fault-tolerant scale, where precision costs $O(1/\varepsilon)$ rather than $O(1/\varepsilon^2)$ — a quadratically better scaling in the parameter that hurts most.

That last point is the strategic one. VQE was designed as the NISQ-compatible substitute for QPE. Its measurement scaling means that for problems large enough to be classically hard, QPE's requirements may arrive before VQE's become tolerable.

Discussion Questions

1. Shot cost scales as $(\sum_k|c_k|)^2/\varepsilon^2$. Explain each factor's origin and which is most attackable.
2. Active-space reduction gave the largest single saving. What does it cost in accuracy, and how would you know?
3. VQE was designed for NISQ hardware but is limited by measurement, not depth. Was the original premise wrong, or has the bottleneck moved?
4. QPE costs $O(1/\varepsilon)$ versus VQE's $O(1/\varepsilon^2)$. At what precision does that difference dominate all other considerations?

Your Turn: Extensions

• Build the H₂ Hamiltonian with Qiskit Nature, count terms, compute $\sum|c_k|$, and estimate shots for chemical accuracy.
• Implement qubit-wise commuting grouping and measure the reduction in distinct settings.
• Compare uniform against $|c_k|$-proportional shot allocation on the same Hamiltonian.
• Run VQE on H₂ with 1,000 versus 100,000 shots per evaluation and observe the effect on convergence.

Key Takeaways

• VQE's binding constraint is measurement, not circuit depth — and shot count is a statistical requirement that better hardware does not reduce.
• Shots scale as $(\sum_k|c_k|)^2/\varepsilon^2$, with Pauli-term count growing as $O(N^4)$, giving a steep polynomial overall.
• Every standard optimization together bought ~160× on BeH₂, taking 360 years to ~19 hours — still $10^5$× slower than an exact classical calculation.
• The classical crossover arrives after VQE's runtime has already become impractical, which is the core objection to the approach.
• Progress requires attacking $\sum|c_k|$ through better Hamiltonian factorizations, or moving to QPE's $O(1/\varepsilon)$ scaling at fault-tolerant scale.