VQE on NISQ hardware collapses under its own measurement budget long before any quantum advantage appears. Teams chasing ground states for molecules larger than a few atoms watch their variance explode after a few hundred parameters, and the classical optimizer never recovers.
Ansatz choice is the hidden tax
Most published circuits reuse hardware-efficient layers that ignore molecular symmetries. The result is a search space filled with barren plateaus where gradients vanish below shot noise. One practical fix is to seed the circuit with a classically pre-optimized Hartree-Fock state and add only symmetry-preserving entangling blocks. That single step cuts the number of required parameters by roughly half on small organics and keeps the optimizer from wandering into flat regions.
Shot overhead kills iteration counts
Each energy evaluation needs enough shots to resolve the difference between two nearby parameter sets. On current devices that difference sits near 10^-3 hartree, so you burn 105 to 106 shots per evaluation. Multiply by the hundreds of evaluations needed for a single optimization run and the wall-time cost exceeds what most cloud queues will sustain. The concrete workaround is to switch to a Bayesian optimizer that reuses previous shot data through a surrogate model; it reduces total shots by a factor of three to five on the same hardware without changing the circuit.
Noise turns the variational principle into a liability
VQE minimizes the expectation value of the Hamiltonian, yet decoherence adds an uncontrolled positive bias. The algorithm therefore reports a lower energy than the true noisy state actually possesses. When the target is chemical accuracy, this bias hides whether you have reached the ground state or simply converged to a noise-corrupted local minimum. Mitigation requires interleaving zero-noise extrapolation with the variational loop rather than applying it only at the end; the extra calibration circuits add 20 percent overhead but remove the systematic error that otherwise invalidates downstream geometry optimizations.
When to stop and hand the problem back to classical methods
If your molecule contains more than twenty qubits worth of active space and the required circuit depth exceeds the device’s two-qubit gate fidelity limit, VQE will not outperform density-matrix renormalization group or selected configuration interaction on the same timeline. Track the quantity “shots per millihartree of improvement.” Once that metric rises above 107, pivot to embedding techniques that fold the quantum subroutine into a larger classical active-space calculation. The hybrid route keeps the quantum device inside its viable regime while still capturing the strong-correlation region that classical methods struggle with.
Teams that treat VQE as a measurement-budget problem rather than a circuit-depth problem finish useful calculations first.