Package to preserve decoding test studies at Fermilab

Dependencies

The needed packages are as follows:

• Stim: You can get through pip by running pip install stim (with ~=1.12 version number if you need to specify it explicitly) and importing via import stim.
• PyMatching comes in a separate repository. You can install it through pip install pymatching and include it via import pymatching.
• Routines inside the notebooks use a few utilities from scikit-learn and scipy, and produce the plots using matplotlib. If you do not have any one of these, they can also be installed through pip: pip install scikit-learn, pip install scipy, and pip install matplotlib, respectively.
• You would also need Tensorflow (pip install tensorflow) and QKeras (pip install qkeras) for NNs.

What is inside this repository?

• The Stim playground notebook introduces a few different wrapper functions I wrote to make the collection of statistics and plotting easier. It also reproduces some of the exercises in the Stim Getting Started notebook.

  • A note on the Stim Getting Started notebook: Not everything is described there; some information may need to be searched within other documents in Stim the repository. For example, the meaning of different gates and other shorthands in the circuit printout are described here.

• The notebook for distance-4 surface codes outlines my progress during the week of Feb. 26 on initial studies for a simple surface code over which one could run different decoding algorithms.

  • In this notebook, I have been able to run PyMatching successfully.
  • It looks like running a Stim-based BP+OSD implementation takes a very long time for a very small number of shots on my personal computer, so I stopped that for now.
  • I was able to run an MLP with a few layers, testing NN decoder performance against PyMatching. When packed bit information about detection events is used, there does not appear to be a good edge over PyMatching.

• The d=4, r=1 notebook includes studies during the week of March 4. Here, I followed the study by Google to include in NN training features both detection events and individual measurements that result in these detections. I use two dense layers + softmax output with labels being the d=4-bit observation information (in the end, I measure the fraction of mispredicted XORs of the unpacked predicted category bits, not the fraction of mispredicted category labels themselves). This seems to outperform PyMatching. There is a hint of overtraining at large number of nodes n for the same number of epochs, which is probably why n=120 in the middle performs better (if there is overtraining, I suppose it could be such that the performance over the XOR worsens).

In these studies, I assume the probabilities of all noise channels are the same, and their built-in implementation in Stim assumes they are constant over time.

Relevant papers

• October '23 paper from Google: https://arxiv.org/abs/2310.05900
• Delft group's paper on NN decoders and hardware costs: https://arxiv.org/abs/2202.05741