Infinite systems of interacting chains with memory of variable length - a stochastic model for biological neural nets

• Each component’s spiking probability depends on accumulated system activity since its last spike, producing a non-Markovian process with memory of variable length.Source: §1 · pp. 1–2
• The proposed approach extends neural point-process models to countably infinite systems without assuming linearity or attractiveness.Source: §1 · p. 2
• A Kalikow-type decomposition, in random environment or space-time, supports proofs of existence, uniqueness, and loss of memory for the stationary process.Source: §1 · p. 2
• In a critical directed Erdős–Rényi-type graph with many but finitely many components, the paper derives an explicit upper bound on successive inter-spike-interval correlations.The bound involves the number of components and the typical length of an inter-spike interval.Source: §1 · p. 3
• The model accommodates small empirical correlations in large systems and potentially substantial negative correlations between neighboring intervals in small systems.The introduction relates these patterns to findings discussed by Gerstner and Kistler (2002) and Nawrot et al. (2007).Source: §1 · pp. 2–3

2 Systems of interacting chains with memory of variable length: Existence, uniqueness and loss of memory

• Neurons take binary spike values and update independently given the past, with transition probabilities specified using synaptic weights and functions of the history since the last spike.The synaptic weight W_{j→i} represents neuron j's influence on neuron i.Source: §2 · pp. 3–4
• For systems with spontaneous spikes, Theorem 1 gives a unique probability law when its assumptions hold and δ exceeds the critical parameter δ*.The model allows spontaneous spikes to represent external stimulus or autonomous brain activity.Source: §2 · pp. 4–5
• Theorem 2 removes the minimum-spiking-rate assumption but requires sufficiently fast decay of aging factors as well as fast decay of synaptic weights; it yields a unique law.Under this additional decay assumption, the specification is continuous.Source: §2 · pp. 5–6
• Under either theorem’s assumptions, Corollary 1 gives a nonincreasing ϕ; when β>β*, Theorem 2’s (2.10) holds with ϕ(s)=Cρ^s, where ρ∈(0,1) depends only on β.Source: §2 · p. 6
• Theorem 1's proof uses a conditional Kalikow-type decomposition and a branching process in random environment, whereas Theorem 2 uses a space-time decomposition.Both proofs imply a perfect simulation algorithm that samples finite space-time windows from the stationary law.Source: §2 · p. 6

3 Correlations between inter-spike intervals in the critical directed Erd¨os-R´enyi random graph

• The finite network uses random Bernoulli synaptic weights, with p_N = λ/N and λ = 1 + ϑ/N, yielding a critical directed Erdős–Rényi graph.The directed link i → j is present exactly when W_i→j = 1.Source: §3 · p. 7
• In this setup, each neuron's spiking function depends on accumulated weighted spikes from interacting neurons, not directly on elapsed time since its last spike.The analysis fixes the connectivity when studying covariance between successive inter-spike intervals.Source: §3 · p. 7
• For large N on the good graph set A, neighboring inter-spike intervals are asymptotically uncorrelated, while small N or graphs outside A may have sizable correlations.The latter is compatible with reported negative correlations between neighboring intervals; the model accommodates both behaviors across scales.Source: §3 · pp. 7–8

Conditional decomposition

• The decomposition uses the spontaneous-spike environment to handle transition probabilities that may be discontinuous.The authors identify dependence on the realization of spontaneous spikes as the new ingredient in their approach.Source: §4 · p. 8
• Conditioned on the spontaneous-spike environment, each transition is a mixture over spatial interaction ranges, with finite-range conditional probabilities.The range weights depend on the environment, while each conditional probability uses only a finite neighborhood.Source: §4 · p. 10

Proof of Theorem 1

• The forward coloring procedure recursively traces the ancestors of a neuron-time pair until its spike decision can be made from spontaneous spikes or range −1 choices.The construction assigns interaction neighborhoods and then works backward through their ancestors before assigning values recursively.Source: §4 · pp. 14–15
• The forward coloring procedure stops almost surely, and the resulting measure is the unique invariant measure of the process.The proof also shows that the probability of stopping before time t tends to 1 almost surely.Source: §4 · p. 15
• For δ ≥ δ*, e(δ) < 1, so e(δ)^n → 0 and the proof’s target result follows.The argument bounds the expected size of an upper-coupled ancestor process by controlling its reproduction mean conditional on the environment.Source: §4 · pp. 16, 18
• The value of δ* depends on the aging functions; when gj(s) = 1 for all j,s, G(n) = n and δ* follows from the third moment of a geometric distribution.Source: §4 · p. 18
• The proof extends the Clan of ancestors method of Fernández et al. (2001, 2002), whose finite ancestor sets yield perfect simulation algorithms.Source: §4 · p. 18

Proof of Corollary 1, part 1

• The total number of elements in the ancestor process gives a rough upper bound on how far into the past one must look to choose Xi(t).The construction makes Xt(i) measurable with respect to the interaction-region choices and Bernoulli and uniform variables used by that process.Source: §4 · p. 18

5 A space-time Kalikow-type decomposition and proof of Theorem 2

• Condition (2.9) permits a space-time Kalikow-type decomposition without assuming spontaneous spikes, using increasing space-time neighborhoods.The neighborhoods are V_i(k) × [−k−1, −1], with V_i(−1) = ∅.Source: §5 · p. 20
• The decomposition's conditional probabilities depend either on the previous state of neuron i alone or on a finite space-time neighborhood.For k ≥ 1, the relevant variables are from neurons in V_i(k) over times t−k−1 through t−1.Source: §5 · p. 21
• Condition (2.9) and the stated estimates give a branching-process reproduction bound m := sup_i m_i < 1.The explored sets are compared with a classical multi-type branching process.Source: §5 · pp. 21–22

6 Proof of Theorem 3

• The proof tracks the first time information emitted by neuron i can return to neuron i itself.This return time is used in the analysis of successive inter-spike intervals.Source: §6 · p. 23
• The proof uses the critical scaling λ = 1 + ϑ/N, with ϑ > 0.It also chooses k(N) so that k(N) → ∞ and k(N) ≤ N.Source: §6 · pp. 23, 25
• The explored influence sets are coupled to a Galton–Watson process with offspring mean μ = (N − 2)λ/N.The N − 2 factor excludes the originating neuron and the target neuron from possible arrow choices.Source: §6 · p. 24