Conditional Independence

This example explores the assumption of conditional independence, and the behaviour of the program when it is not satisfied.

Every state-by-node TPM corresponds to a unique state-by-state TPM which satisfies the conditional independence assumption. If a state-by-node TPM is given as input for a network, the program assumes that it is from a system with the corresponding conditionally independent state-by-state TPM.

When a state-by-state TPM is given as input for a network, the state-by-state TPM is first converted to a state-by-node TPM. The program then assumes that the system corresponds to the unique conditionally independent representation of the state-by-node TPM. If a non-conditionally independent TPM is given, the analyzed system will not correspond to the original TPM. Note that every deterministic state-by-state TPM will automatically satisfy the conditional independence assumption.

Consider a system of two binary nodes (\(A\) and \(B\)) which do not change if they have the same value, but flip with probability 50% if they have different values.

We’ll load the state-by-state TPM for such a system from the examples module:

>>> import pyphi
>>> tpm = pyphi.examples.cond_depend_tpm()
>>> print(tpm)
[[ 1.   0.   0.   0. ]
 [ 0.   0.5  0.5  0. ]
 [ 0.   0.5  0.5  0. ]
 [ 0.   0.   0.   1. ]]

This system does not satisfy the conditional independence assumption; given a past state of (1, 0), the current state of node \(A\) depends on whether or not \(B\) has flipped.

When creating a network, the program will convert this state-by-state TPM to a state-by-node form, and issue a warning if it does not satisfy the assumption:

>>> sbn_tpm = pyphi.convert.state_by_state2state_by_node(tpm)

“The TPM is not conditionally independent. See the conditional independence example in the documentation for more information on how this is handled.”

>>> print(sbn_tpm)
[[[ 0.   0. ]
  [ 0.5  0.5]]

 [[ 0.5  0.5]
  [ 1.   1. ]]]

The program will continue with the state-by-node TPM, but since it assumes conditional independence, the network will not correspond to the original system.

To see the corresponding conditionally independent TPM, convert the state-by-node TPM back to state-by-state form:

>>> sbs_tpm = pyphi.convert.state_by_node2state_by_state(sbn_tpm)
>>> print(sbs_tpm)
[[ 1.    0.    0.    0.  ]
 [ 0.25  0.25  0.25  0.25]
 [ 0.25  0.25  0.25  0.25]
 [ 0.    0.    0.    1.  ]]

A system which does not satisfy the conditional independence assumption exhibits “instantaneous causality.” In such situations, there must be additional exogenous variable(s) which explain the dependence.

Consider the above example, but with the addition of a third node (\(C\)) which is equally likely to be ON or OFF, and such that when nodes \(A\) and \(B\) are in different states, they will flip when \(C\) is ON, but stay the same when \(C\) is OFF.

>>> tpm2 = pyphi.examples.cond_independ_tpm()
>>> print(tpm2)
[[ 0.5  0.   0.   0.   0.5  0.   0.   0. ]
 [ 0.   0.5  0.   0.   0.   0.5  0.   0. ]
 [ 0.   0.   0.5  0.   0.   0.   0.5  0. ]
 [ 0.   0.   0.   0.5  0.   0.   0.   0.5]
 [ 0.5  0.   0.   0.   0.5  0.   0.   0. ]
 [ 0.   0.   0.5  0.   0.   0.   0.5  0. ]
 [ 0.   0.5  0.   0.   0.   0.5  0.   0. ]
 [ 0.   0.   0.   0.5  0.   0.   0.   0.5]]

The resulting state-by-state TPM now satisfies the conditional independence assumption.

>>> sbn_tpm2 = pyphi.convert.state_by_state2state_by_node(tpm2)
>>> print(sbn_tpm2)
[[[[ 0.   0.   0.5]
   [ 0.   0.   0.5]]

  [[ 0.   1.   0.5]
   [ 1.   0.   0.5]]]


 [[[ 1.   0.   0.5]
   [ 0.   1.   0.5]]

  [[ 1.   1.   0.5]
   [ 1.   1.   0.5]]]]

The node indices are 0 and 1 for \(A\) and \(B\), and 2 for \(C\):

>>> AB = [0, 1]
>>> C = [2]

From here, if we marginalize out the node \(C\);

>>> tpm2_marginalizeC = pyphi.tpm.marginalize_out(C, sbn_tpm2)

And then restrict the purview to only nodes \(A\) and \(B\);

>>> import numpy as np
>>> tpm2_purviewAB = np.squeeze(tpm2_marginalizeC[:,:,:,AB])

We get back the original state-by-node TPM from the system with just \(A\) and \(B\).

>>> np.all(tpm2_purviewAB == sbn_tpm)
True