pyphi-troubleshooting

PyPhi common pitfalls

1. Numerical precision issues

Problem: Floating-point comparisons fail due to precision.

# Bad
if phi == 0.0:  # May fail due to floating point error

# Good
from pyphi import numerics
if numerics.is_zero(phi):  # Respects config.numerics.precision

Key functions (in pyphi.numerics):

• numerics.is_zero(x)
• numerics.is_positive(x)
• numerics.eq(x, y)

Comparisons are tolerant up to config.numerics.precision (default 13 decimal places). Tie resolution clusters candidates equal up to this tolerance, so equivalent selections co-select regardless of iteration order.

2. Configuration not in working directory

Problem: pyphi_config.yml must be in the directory where Python starts.

cd /somewhere/else
python -c "import pyphi"  # Uses defaults!

cd /my/project
python -c "import pyphi"  # Loads ./pyphi_config.yml

The file uses the nested layout with top-level keys formalism, infrastructure, and numerics. A legacy flat YAML (pre-2.0 UPPER_CASE keys at the top level) is rejected at load time with a rename map — port it to the nested keys. At runtime, a top-level write such as pyphi.config.precision = 6 routes to the correct layer; the explicit path pyphi.config.numerics.precision also works. Use with pyphi.config.override(...) for temporary scopes.

3. Missing optional dependencies

Problem: Import errors for optional features. The real extras are:

uv pip install pyphi[visualize]  # matplotlib, plotly, seaborn, networkx
uv pip install pyphi[caching]    # redis
uv pip install pyphi[emd]        # POT (ot.emd2), for EMD distance measures
uv pip install pyphi[xarray]     # xarray-backed TPM export
uv pip install pyphi[mcp]        # the pyphi-mcp server

Parallelism needs no extra — it uses the standard-library process pool. There is no parallel or graphs extra (networkx ships under visualize).

4. TPM format confusion

Problem: A TPM can be in different formats.

• State-by-node: rows are states, columns are nodes (P(node = 1) for binary; per-unit distributions for k-ary).
• State-by-state: rows are current states, columns are next states.
• Multidimensional: high-dimensional array indexed by node states.

Solution: Use the conversion utilities in pyphi.convert. The core TPM types live in pyphi.core.tpm (there is no top-level pyphi.tpm); a Substrate holds a FactoredTPM, and Substrate.joint_tpm() gives the read-only joint view.

5. System state validation

Problem: Building a System with an inconsistent state.

config.infrastructure.validate_system_states (default on) checks the state against the substrate. Note the 2.0 rename: the core value types are Substrate (formerly Network) and System (formerly Subsystem).

Solution: Use a reachable state, or scope the check off with pyphi.config.override(validate_system_states=False) for special cases.

6. Cache memory growth

Problem: Large substrates can fill memory with cached repertoires.

Solution: The content cache evicts by refcount as substrates are garbage collected, so releasing references frees entries. To bound or disable it:

pyphi.config.cache_repertoires = False
# or bound the footprint:
pyphi.config.maximum_cache_memory_percentage = 25

config.infrastructure.disk_cache_results (default off) additionally persists System.sia()/ces() results to __pyphi_cache__/.

7. Parallel computation not engaging

Problem: Set parallel = True but computation still runs sequentially.

Checklist:

1. Is config.infrastructure.parallel on, and the specific operation's flag enabled? Each site is a dict: e.g. config.infrastructure.parallel_concept_evaluation["parallel"] = True (siblings: parallel_purview_evaluation, parallel_partition_evaluation, parallel_relation_evaluation, parallel_complex_evaluation, parallel_mechanism_partition_evaluation, parallel_macro_system_evaluation).
2. Is the problem size above that site's sequential_threshold? Below it, the dispatch stays sequential by design.
3. Backend is stdlib ProcessPoolExecutor by default (config.infrastructure.parallel_backend); no Ray, no cluster setup.

8. IIT version mismatch

Problem: Mixing IIT 3.0 and IIT 4.0 code paths or config.

Solution: The version is config.formalism.iit.version (e.g. IIT_4_0_2023, IIT_4_0_2026, IIT_3_0); presets live in pyphi.conf.presets. The entry points are formalism-agnostic — there is no pyphi.new_big_phi or pyphi.compute in 2.0:

• System(substrate, state).sia() — system irreducibility (Φ).
• System(substrate, state).ces() — the cause-effect structure.
• pyphi.analyze(substrate, state, formalism=…) — one high-level entry point.

The 2026 formalism selects the MIP exactly as 2023 does, then applies the Eq. 23 ii-cap, so IIT_4_0_2026 Φ is never above IIT_4_0_2023 Φ.