Most of the complex systems can not be understood using traditional physicists approach. Complex systems exhibit behavior essence of which lies in the interaction of the different parts inside the system. As complex systems include not only physical, but also social and economical systems, it is convenient to introduce agents - very general concept describing varying parts of the systems (ex. atoms, humans, animals). One of the signature features of human society as a complex system is that it creates orderly structures - political ideologies, religions, fan clubs, brand-name phenomena and etc. While speaking very generally these observations can be blamed on the human herding instinct. Thus a successful agent based model should pay attention of how effective is its description of agents' herding.
Evidently there are two distinct simple ways (actually there is a third more complex - network modeling - way) to treat herding behavior - agents might be allowed to interact with single colleague per time step (ex. a friend recommends to watch the movie) [1] or might be in some way coupled to the source of information expressing opinion of each agent in the system (ex. mass media, social networks) [2]. Unidirectional version of Kirman's model and the GLM model were shown to have the same asymptotic behavior as renowned Bass diffusion model, which describes diffusion of successful products in the market [3, 4]. From our point of view it is rather interesting to see how well these distinct definitions of herding behavior work in case of bidirectional transitions. Note that while Kirman's model was originally defined as bidirectional birth-death process [1], the GLM model was defined only as unidirectional process [2]. Thus introducing the bidirectionality into the GLM model is the main novelty presented in this work.
Following the mathematical formalism of the GLM model and applying ideas of bidirectional transitions by Kirman, let each agent occupying state $s_1$, here X is their number, to move to state $s_2$ with probability:
$$p(s_1 \to s_2) = 1 - (1 - \Sigma_1)(1 - H)^{N - X},\tag{1}$$consequently N-X agents occupying state $s_2$ can move to $s_1$ with probability:
$$p(s_2 \to s_1) = 1 - (1 - \Sigma_2)(1 - H)^X. \tag{2}$$In the above probabilities H is present in the herding terms (thus it has a meaning of probability to be recruited by the herd during the time step), while $\Sigma_i$ represents individual transitions (also during the time step). Both H and $\Sigma_i$ can be assumed to be linearly dependent on the time step ($\Sigma_i = \sigma_i \Delta t$, $H = h \Delta t$) [2].
By expanding the bidirectional transitions probabilities of the GLM model in the short time step limit, $\Delta t \rightarrow 0$, one can obtain per-agent transition probabilities identical to the Kirman's model:
$$p(s_1 \to s_2) \approx [\sigma_1 + h(N - X)]\Delta t = \frac{p(X \to X - 1)}{X}, \quad p(s_2 \to s_1) \approx [\sigma_2 + hX]\Delta t = \frac{p(X \to X + 1)}{N - X}, \tag{3}$$where $p(X \to X \pm 1)$ is a Kirman model one-step transition probabilities. This result serves as analytical proof that in the aforementioned limit of short time steps we can use the results obtained for Kirman model and its financial market treatment [5] towards the GLM model.
In this contribution we have shown that the GLM model and Kirman model describe asymptotically identical behavior. These models agree well in the limit of small time steps, though note that the selection of the time step should inversely depend on the square of the size of the system, $\Delta t \sim N^{-2}$.
To reach the main conclusion one needs to recall that the GLM model describes agent-to-community interaction during the time step, while Kirman's model is concerned with agent-to-agent interaction during the time step. The asymptotic agreement of the models might be interpreted via Monte Carlo method formalism - completely random polling is statistically the same as polling whole population.