What is the cost of probabilistic error cancellation?

It can require the device to be recalibrated after every shot, because the quasi-probabilities drift as the qubits warm up

It can require large sampling overhead because noisy operations are represented using quasi-probability decompositions

It can require many extra physical qubits because each noisy operation is encoded into a larger error-correcting block

It can require the circuit to be rerun on a noiseless simulator, since the correction is computed from the exact amplitudes