---
identifier: iit:result/albantakis-2023-fig1-triad
title: The three-unit substrate of the IIT 4.0 paper condenses into the complexes C and aB
vetted_by: "Matteo Grasso"
formalism: iit-4.0-2026
summary: >-
  The substrate aBC that Albantakis et al. (2023) use to introduce the
  postulates (Fig 1) condenses into two complexes, C (φ_s = 0.21) and aB
  (φ_s = 0.04). The 2023 formulation gives the same complexes, with
  φ_s(C) = 0.21 and φ_s(aB) = 0.17 as in Fig 1E. Under the current
  formulation the intrinsic information requirement lowers φ_s(aB) below
  φ_s(aBC) = 0.13, so aB remains a complex only because C excludes aBC
  first.
substrate: >-
  Three binary units A, B, C with states in {−1, 1}. Each unit is a logistic
  of its weighted inputs (Eq 60) with slope k = 4. The weights, read from the
  Fig 1A causal model, are A→A −0.2, A→B 0.7, A→C 0.2, B→A 0.7, B→B −0.2,
  C→B −0.8, C→C 0.2; there is no B→C or C→A connection. PyPhi ships the
  substrate as `pyphi.examples.iit4_2023_fig1a_substrate`.
state: "(−1, 1, 1) — aBC, units B and C ON"
phi: 0.21221955291860525
mip: >-
  For the first complex C, the cut of its self-connection C→C. For the
  second complex aB, the directed bipartition cutting A→B. The same under
  both formulations.
pyphi_version: 2.0.0rc4
script: albantakis-2023-fig1-triad.py
see_also:
  - iit:measure/phi-s
sources:
  - ref: iit:ref/albantakis-2023b
    at: >-
      Fig 1C–E and the caption of Fig 1; "Exclusion: Determining maximal
      substrates (complexes)", Eqs 24–26
  - ref: iit:ref/mayner-2026
    at: '§2.3 "Intrinsic Information" and §2.4 "Integrated Information"'
---

Albantakis et al. (2023) introduce the postulates of existence,
intrinsicality, information, integration, and exclusion on a three-unit
substrate in state aBC (Fig 1). Fig 1E reports φ_s for four candidate
systems and marks aB and C as complexes. The script computes φ_s for all
seven candidate systems and runs the recursive exclusion search (Eqs 24–25)
under both formulations.

| candidate | φ_s, Fig 1E | iit-4.0-2023 | iit-4.0-2026 |
|---|---|---|---|
| C | 0.21 | 0.2122 | 0.2122 |
| aB | 0.17 | 0.1719 | 0.0405 |
| aBC | 0.13 | 0.1339 | 0.1339 |
| a | 0.04 | 0.0396 | 0.0389 |
| B | — | 0.0032 | 0.0016 |
| aC, BC | — | 0 | 0 |

Under both formulations C has the highest φ_s, so C is the first complex. C
excludes every candidate that overlaps it: aBC, aC, and BC. Among the
remaining candidates aB, a, and B, aB has the highest φ_s, so aB is the
second complex. The exclusion order is the same under both formulations.

The caption of Fig 1E justifies aB as a complex because its φ_s "is higher
than that of all other overlapping systems (for example, subset *a* with
φ_s = 0.04 and superset *aBC* with φ_s = 0.13)". Under the 2023 formulation
this comparison holds. Under the current formulation it does not hold as
written: aBC (0.13) exceeds aB (0.04). aB is a complex because C, which does
not overlap aB, excludes aBC first. aB then exceeds a by a small margin
(0.0405 against 0.0389). Fig 1E also marks C as a complex with
φ_s = 0.21, which matches the computed value.

The difference in φ_s(aB) comes from the intrinsic information requirement
of Mayner et al. (2026). The current formulation sets
φ_s = min{φ_c, φ_e, ii}, where ii is the minimum over cause and effect of
intrinsic differentiation and intrinsic specification (§2.3–2.4); the 2023
formulation sets φ_s = min{φ_c, φ_e}. The script prints each term for the
two complexes.

| term | aB, Fig 1C–D | aB, computed | C, computed |
|---|---|---|---|
| intrinsic specification, cause (Fig 1C: ii_c) | 0.84 | 0.8432 | 0.2122 |
| intrinsic specification, effect (Fig 1C: ii_e) | 1.1 | 1.0976 | 0.2911 |
| intrinsic differentiation, cause | — | 0.6640 | 0.6638 |
| intrinsic differentiation, effect | — | 0.0405 | 1.0000 |
| φ_c | 0.24 | 0.2446 | 0.2122 |
| φ_e | 0.17 | 0.1719 | 0.2911 |
| φ_s, 2023 | 0.17 | 0.1719 | 0.2122 |
| φ_s, 2026 | — | 0.0405 | 0.2122 |

In the 2023 formulation, intrinsic information is what the current
formulation calls intrinsic specification, so the Fig 1C values ii_c = 0.84
and ii_e = 1.1 correspond to the intrinsic specification rows. Intrinsic
differentiation is new in the current formulation and has no counterpart in
the figure. φ_c and φ_e are the same under both formulations and match
Fig 1D. For aB,
the specified effect state Ab has probability 0.9723, so the effect side's
intrinsic differentiation, −log₂ 0.9723 = 0.0405, is the minimum of all
terms. For C, the minimum is φ_c = 0.2122 under both formulations, so its
φ_s does not change.

The script uses the default configuration of each formulation, applied with
PyPhi's `IIT_4_0_2026` and `IIT_4_0_2023` presets. PyPhi codes the state −1
as 0. To reproduce, run the script with uv, which installs PyPhi 2.0.0rc4 for the run:

```
uv run --with "pyphi==2.0.0rc4" python canon/albantakis-2023-fig1-triad.py
```
