This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

Dynamical, value-based decision making among NN options:
a constructive approach to unfolding
the symmetric pitchfork bifurcation

Paul Reverdy P. Reverdy is with the Department of Aerospace and Mechanical Engineering, University of Arizona, Tucson, AZ 85721. Email: [email protected]
Abstract

Decision making is a fundamental capability of autonomous systems. As decision making is a process which happens over time, it can be well modeled by dynamical systems. Often, decisions are made on the basis of perceived values of the underlying options and the desired outcome is to select the option with the highest value. This can be encoded as a bifurcation which produces a stable equilibrium corresponding to the high-value option. When some options have identical values, it is natural to design the decision-making model to be indifferent among the equally-valued options, leading to symmetries in the underlying dynamical system. For example, when all NN options have identical values, the dynamical system should have SNS_{N} symmetry. Unfortunately, constructing a dynamical system that unfolds the SNS_{N}-symmetric pitchfork bifurcation is non-trivial. In this paper, we develop a method to construct an unfolding of the pitchfork bifurcation with a symmetry group that is a significant subgroup of SNS_{N}. The construction begins by parsing the decision among NN options into a hierarchical set of N1N-1 binary decisions encoded in a binary tree. By associating the unfolding of a standard S2S_{2}-symmetric pitchfork bifurcation with each of these binary decisions, we develop an unfolding of the pitchfork bifurcation with symmetries corresponding to isomorphisms of the underlying binary tree.

keywords:
symmetries of dynamical systems, bifurcations, binary trees
{AMS}

37C80, 37G10, 37E25

A fundamental characteristic of systems that exhibit autonomy is the ability to decide among a set of possible actions. Such systems arise in the natural world, composed of individuals or groups of humans or animals, and also in the artificial world, composed of robots or other algorithmic agents. Studying decision making in both types of systems is of scientific and engineering interest. In natural systems, one tends to focus on developing models that explain observed behavior, while in artificial systems, one tends to focus on developing models that perform some desired behavior. In both cases, it is natural to develop models which make decisions using a dynamic mechanism that reacts to some perceived notion of action value. This paper focuses on the case of a single agent and an arbitrary number of actions and constructs such a mechanism based on the pitchfork bifurcation with symmetries corresponding to interchange of action labels.

As done in numerous recent papers [1, 2, 3, 4, 5, 6], it is natural to model the decision making process using a dynamical system. Decision making is itself a dynamic process, affected by the set of actions available to the individual and the context, including the individual’s internal state and external inputs. For example, consider an animal running in a rough environment. The actions available to the animal correspond to locations it can place its feet on each stride, the internal state to the animal’s level of fatigue and fear, and the external inputs to stimuli revealing the quality of available foot placements or the presence of a predator. Decisions must be made quickly, since one is required for each step, and the decision-making process must be flexible, so that decisions can be revised smoothly if a previously attractive location turns out to be undesirable or the appearance of a predator requires a startle response. As argued in [6], these requirements suggest viewing the decision-making process as a dynamical system operating in continuous time and on a continuous state space.

The example of the running animal is an example of the ecological theory of affordances. An affordance is an opportunity for action which the environment affords the agent [7], and the theory of affordances has gained significant traction as a model of ecological decision making [8, 9]. In particular, the so-called affordance competition hypothesis [10, 11] suggests that animals perform physical behaviors by continuously identifying affordances and selecting among them by a process which is biased by the desirability of their predicted outcomes. More recently, affordance theory has been investigated in robotics as an approach for developing novel control laws [12, 13]. In our own work we have begun to use dynamical systems which we call motivation dynamics to develop both theoretical [14] and practical [15] robot control systems. Therefore, a dynamical system for making decisions on the basis of perceived action values could be valuable for applications in both ecology and robotics.

This paper constructs a continuous dynamical system that allows a single agent to make decisions on the basis of perceived action values. Motivated by recent work in this area [1, 3, 6, 5], we base our dynamical system on the pitchfork bifurcation with certain symmetries. In the nominal case where all NN actions have the same value, we require that the dynamical system be equivariant under interchange of actions and impose symmetries which correspond to permutations of the actions. This allows us to use the tools of equivariant bifurcation theory [16, 17]. We introduce asymmetric action values as unfolding parameters of this pitchfork bifurcation, and construct the system so that decisions correspond to bifurcations.

The key difficulty in our approach lies in the fact that developing an unfolding for the pitchfork bifurcation with SNS_{N}-symmetry is non-trivial. In the case of a binary decision between N=2N=2 options, i.e., S2S_{2} symmetry, the unfolding properties are well understood [16], but generalizing to the case N>2N>2 is complex. For N=2N=2, the model of value-based decision making introduced by [1] and analyzed by [3] can be shown to embed an unfolding of the pitchfork bifurcation [18], but this structure does not naturally generalize even to the case N=3N=3 [4]. Reina et al. [4] show how to modify the Seeley et al. model to recover desirable bifurcation properties in the case of an arbitrary number of options NN but with only one best option, though they do not provide analysis for their general case. In the recent paper by Franci et al. [6] the authors use equivariant bifurcation theory to analyze a similar system in the general case of N>2N>2 options (and indeed, for a generic number M>1M>1 of interacting agents). However, Franci et al. present explicit equations only for the case of N=3N=3 options and M=3M=3 agents, and they consider only the symmetric case of equally-valued options.

In contrast to these preceding works, the present paper presents an explicit construction of an unfolding of a pitchfork bifurcation with a symmetry group that is a significant subgroup of SNS_{N} for N2N\geq 2. The core idea underlying our construction is that one can parse a decision among N2N\geq 2 options into a series of binary decisions; these decisions can be represented by a binary tree. This representation allows us to take advantage of the well-understood representation of binary decisions in terms of the unfolding of a standard pitchfork singularity. We then construct a dynamical system with the desired bifurcation for N2N\geq 2 by assembling a series of standard pitchfork systems in a way that reflects the parsing encoded by the tree structure. In the symmetric case where the options are equally-valued, the dynamical system is equivariant under transformations (i.e., relabeling of options) that correspond to isomorphisms of the binary tree. Breaking the symmetry of option values naturally results in an unfolding of the pitchfork bifurcation.

The contributions of this paper are three-fold. First, we develop a dynamical systems model of decision making where the structure of the decision, and of the dynamical system itself, is encoded in a tree graph structure. Such graph structures have often been used in studying multi-agent decision-making problems, [19, 5, 6], but to our knowledge this is the first time a graph structure has been used to organize the dynamical decision-making procedure of a single agent. Secondly, the dynamical system we develop has a novel feedforward structure due to its recursive definition, where the vector field is constructed by recursively parsing down the tree structure. The structure is such that dynamics flow from the root node towards the leaves, which represent options. Thirdly, in building our model we develop a constructive approach to unfolding the equivariant pitchfork bifurcation. Our model has symmetries which correspond to isomorphisms of the underlying binary tree structure, which results in our pitchfork having a symmetry group that is a significant subgroup of SNS_{N}. This significantly extends existing results, e.g., in [4] and [6].

1 Dynamical systems as models of NN-ary decision making

In this section we summarize our requirements for dynamical systems models of decision making. We suppose we have N2N\geq 2 actions. In the following, we refer to actions by the more generic term option to reflect that decision making need not be directly tied to action. The ithi^{th} option has value vi>0v_{i}>0. Inspired by previous work in the decision making literature [1, 3], we encode the decision state of the system in a variable mΔNm\in\Delta^{N}, where

ΔN={xN+1|xi0,ixi=1}\Delta^{N}=\left\{x\in\mathbb{R}^{N+1}\Big{|}x_{i}\geq 0,\sum_{i}x_{i}=1\right\}

denotes the NN-simplex. Let eiN+1,i{1,,N+1},e_{i}\in\mathbb{R}^{N+1},i\in\{1,\ldots,N+1\}, denote the ithi^{th} corner of ΔN\Delta^{N}, i.e., the vector with the ithi^{th} element equal to one and all other elements equal to zero. The ithi^{th} element of mm, denoted mim_{i}, represents the degree to which the system has committed to option ii. Note that mi[0,1]m_{i}\in[0,1]. The N+1N+1 element of mm represents the degree to which the system is uncommitted to any option, and the normalization condition ensures that the total commitment is finite.

We encode the decision-making process in a continuous vector field ff operating on the state space ΔN\Delta^{N}. A decision for option ii is represented by the state mm approaching the corresponding corner of the NN-simplex, i.e., the state eie_{i}. See Figure 1 for the case N=2N=2. We want the system to commit to an option ii when its corresponding value viv_{i} is sufficiently high. Thus, we construct the vector field ff to have an attracting equilibrium near eie_{i} when viv_{i} is high. When several different options have high values but there is no clear maximum, it is desirable to have several equilibria that encode partial commitment to the attractive options.

111111x1x_{1}x2x_{2}x3x_{3}
Figure 1: Plot of the 2-simplex Δ2\Delta^{2}. The simplex is the shaded section of the plane x1+x2+x3=1x_{1}+x_{2}+x_{3}=1.

To structure the equilibria, we require symmetry in the vector field ff that reflects the symmetry in the options. We model options to be distinguished only by their associated values. Thus symmetries in the options are permutations of the set {vi,i{1,,N}}\{v_{i},i\in\{1,\ldots,N\}\} that leave the set invariant. In the fully-symmetric case where vi=v>0i{1,,N}v_{i}=v>0\forall i\in\{1,\ldots,N\}, all options are equally valued and therefore identical, and the symmetry group is the symmetric group SNS_{N}. When the option values are not equal, we want the symmetry in ff to be broken. Regardless of any symmetries that may be present, we want the system to remain uncommitted when the values viv_{i} are low and to commit when the values are above a threshold. In the symmetric case, we want the system to remain uncommitted when the value vv is low, and to commit symmetrically to all options when vv increases beyond a threshold. Conversely, in the asymmetric case, we want the system to commit asymmetrically to the high-value options.

1.1 Equivariant bifurcation theory

The desire for a switching commitment process strongly suggests the use of a bifurcation in the dynamical system, and the symmetry requirement encourages the use of equivariant bifurcation theory [16, 20]. In the case of N=2N=2 options, our requirements are satisfied by selecting the vector field ff to embed a pitchfork bifurcation as shown in Figure 2. The mechanism can be viewed as follows. Let x=m1m2x=m_{1}-m_{2} represent the degree of commitment for option 1 over option 2. Then, in the symmetric case, selecting ff to be a pitchfork bifurcation, i.e.,

(1) x˙=f(x,μ)=x(μx2),μ\dot{x}=f(x,\mu)=x(\mu-x^{2}),\ \mu\in\mathbb{R}

will yield the desired behavior. Note that ff is an odd function of xx, so the system (1) obeys the S2S_{2} symmetry corresponding to switching the option labels. When μ\mu (which represents the value vv) is less than zero, the unique equilibrium of (1) is x=0x=0, corresponding to equal commitment to both options, and this equilibrium is stable. As μ\mu increases through zero, the equilibrium at x=0x=0 becomes unstable and two new stable equilibria appear at ±μ\pm\sqrt{\mu}. These facts are summarized in the bifurcation diagram shown in Figure 2(a). Equivariant bifurcation theory predicts that these two branches must appear symmetrically in the post-bifurcation regime; each branch represents a commitment to one option or the other. Due to symmetry, commitment to either option is possible; the option to which the system commits is determined by initial conditions.

The case of asymmetric option values is naturally modeled by breaking the symmetry of the vector field ff. The study of such symmetry breaking is a core part of equivariant bifurcation theory, and the fundamental concept is the so-called unfolding of the bifurcation. One begins with the concept of a perturbation of a bifurcation. Specifically, a perturbation FF of a bifurcation ff is a function F(x,μ,α)F(x,\mu,\alpha) such that F(x,μ,0)=f(x,μ)F(x,\mu,0)=f(x,\mu). A universal unfolding is a kk-parameter (i.e., αk\alpha\in\mathbb{R}^{k}) perturbation FF such that any small perturbation of ff can be expressed in terms of the kk parameters that define FF. The two-parameter family

(2) F(x,μ,α)=x(μx2)+α1+α2x2,α1,α2F(x,\mu,\alpha)=x(\mu-x^{2})+\alpha_{1}+\alpha_{2}x^{2},\ \alpha_{1},\alpha_{2}\in\mathbb{R}

is a universal unfolding of the pitchfork bifurcation (1) [16]. As the parameters α1,α2\alpha_{1},\alpha_{2} are varied, the bifurcation diagram for x˙=F(x,μ,α)\dot{x}=F(x,\mu,\alpha) varies as well. In particular, for appropriate parameter values, the system has a single equilibrium in the post-bifurcation regime, and this equilibrium is both nonzero and stable (see Figure 2(b)). Thus, by choosing appropriate values for the unfolding parameters, one can encode a globally-attracting preference for one option or the other. To complete the model for N=2N=2 options, it remains to relate the option values viv_{i} to the bifurcation parameter μ\mu and unfolding parameters αi\alpha_{i} in (2).

Refer to caption Refer to caption
(a) (b)
Figure 2: Bifurcation diagrams for (a) the symmetric pitchfork (1) and (b) the unfolded pitchfork (2). Stable equilibria are represented by solid lines and unstable equilibria by dashed lines. Equilibria above the μ\mu axis represent a preference for option 1, while those below represent a preference for option 2. In panel (a) the system has a single stable equilibrium representing no preference when μ\mu is small and two symmetric stable equilibria representing preferences for option 1 and option 2 respectively, when μ\mu is large. Note particularly in panel (b) that, for intermediate values of μ\mu, the system has a single stable equilibrium representing a preference for option 1.

1.2 Seeley et al. model for N=2N=2

In their paper [1], Seeley et al. study the value-sensitive decision-making problem for N=2N=2 options and develop a dynamical systems model. Their model can be shown to embed an unfolded pitchfork [18], thus completing the model whose mechanism we laid out in the preceding section. Concretely, Seeley et al. let the state of their model be m=(m1,m2,mU)Δ2m=(m_{1},m_{2},m_{U})\in\Delta^{2} and set m˙=f(m,v)\dot{m}=f(m,v), with

(3) m˙i=vimUmi(1vivimU+σ(1mimU))\displaystyle\dot{m}_{i}=v_{i}m_{U}-m_{i}\left(\frac{1}{v_{i}}-v_{i}m_{U}+\sigma(1-m_{i}-m_{U})\right)

for each i=1,2i=1,2, and the dynamics of mUm_{U} are determined by the normalization constraint. As above, viv_{i} denotes the value of option ii, and σ>0\sigma>0 is a constant parameter.

In the symmetric case of v1=v2=vv_{1}=v_{2}=v the dynamics (3) obey an S2S_{2} symmetry. Seeley et al. [1] showed that, in the symmetric case, the dynamics (3) exhibits a pitchfork bifurcation as vv and σ\sigma increase above a threshold. In the pre-bifurcation regime, the system has a single stable equilibrium with the symmetry m1=m2m_{1}=m_{2}. Because of the symmetry, the system does not commit to either option, and the equilibrium is said to be a deadlock. In the post-bifurcation regime, the deadlock equilibrium is unstable and two additional equilibria emerge, each representing a decision to commit to one of the two options. In previous work [1], the parameter values at the bifurcation point were found to satisfy

(4) σ=4v3(v21)2.\sigma=\frac{4v^{3}}{(v^{2}-1)^{2}}.

Note that either σ\sigma or vv can be interpreted as the bifurcation parameter, with (4) defining the bifurcation value. In other words, fixing σ\sigma, the bifurcation occurs as vv increases through a threshold, while fixing vv, the bifurcation occurs as σ\sigma increases.

In the asymmetric case of v1v2v_{1}\neq v_{2}, the S2S_{2} symmetry of the dynamics is broken and the pitchfork bifurcation unfolds as studied by Pais et al. [3]. For a fixed value of σ\sigma, the number of equilibria of (3) depends on the parameters v1v_{1} and v2v_{2}. Certain parameter values result in a single stable equilibrium whose location is biased towards the high-value option, while others result in two stable equilibria representing each option and a saddle point in between. The complete phase diagram of the system is complex, but the two primary findings of [3] can be summarized as follows. First, the dynamics (3) remain deadlocked (i.e., have a single attractor with m1m2m_{1}\approx m_{2}) when the average option value v¯=(v1+v2)/2\bar{v}=(v_{1}+v_{2})/2 is small. Second, the dynamics decide for the high value option (i.e., for v1>v2v_{1}>v_{2}, have a single attractor with m1m2m_{1}\gg m_{2}) when the difference in option values Δv=v1v2\Delta v=v_{1}-v_{2} is sufficiently large relative to v¯=(v1+v2)/2\bar{v}=(v_{1}+v_{2})/2. In symbols, we have that the system makes a decision when

|Δv|v¯>κ(σ),\frac{|\Delta v|}{\bar{v}}>\kappa(\sigma),

where κ(σ)\kappa(\sigma) is a coefficient that depends on the parameter σ\sigma. Pais et al. [3] note that this behavior is analogous to Weber’s law of just-noticeable differences from psychology, which states that the minimum difference in stimulus intensity required to discriminate between two different stimuli varies linearly with their mean intensity.

The implication of these two findings is that the decision-making dynamics (3) has several desirable properties. First, when both options are poor (corresponding to a low value of v¯\bar{v}), the system remains deadlocked and avoids making a decision, e.g., to wait for more information. When at least one option is sufficiently satisfactory (corresponding to a high value of v¯\bar{v}), the system will quickly commit to an option, and preferentially select the one with a higher value. These are properties that we seek to generalize to the case of N>2N>2 options.

1.3 Reduction of the Seeley et al. model

As shown in the preceding sections, the Seeley et al. model (3) has desirable characteristics that we seek to generalize. However, the functional form of (3) obscures the unfolding of the pitchfork bifurcation which serves as the fundamental decision-making mechanism. In recent work [18], we studied the dynamics (3) using model reduction techniques to elucidate the unfolding.

The model reductions studied in [18] use singular perturbation theory. Specifically, the reduction approach maps viKviv_{i}\mapsto Kv_{i} for a constant gain K>0K>0 and takes the singular limit K+K\to+\infty. This approach is similar to an analysis performed in [3], where the authors studied the limit v¯+\bar{v}\to+\infty; however, the approach using the gain KK preserves the relative difference in values Δv/v¯\Delta v/\bar{v}. This ratio is key in defining the unfolding of the pitchfork bifurcation embedded in (3).

The bifurcation is more readily analyzed by expressing mΔ2m\in\Delta^{2} in terms of mean-difference coordinates defined by

Δm=m1m2,m¯=m1+m22\Delta m=m_{1}-m_{2},\bar{m}=\frac{m_{1}+m_{2}}{2}

which are analogous to the definitions of Δv\Delta v and v¯\bar{v} made above. Note that the definitions of these new coordinates and the definitions of Δ2\Delta^{2} and (v1,v2)+2(v_{1},v_{2})\in\mathbb{R}_{+}^{2} imply that m¯,v¯>0\bar{m},\bar{v}>0 and that 2m¯Δm2m¯-2\bar{m}\leq\Delta m\leq 2\bar{m} and 2v¯<Δv<2v¯-2\bar{v}<\Delta v<2\bar{v}.

In the mean-difference coordinates, the dynamics (3) of m=(Δm,m¯)m=({\Delta m},{\bar{m}}) take the form

(5) Δm˙\displaystyle\dot{\Delta m} =fΔm(Δm,m¯;v¯;Δv)\displaystyle=f_{\Delta m}(\Delta m,\bar{m};\bar{v};\Delta v)
=(2m¯+ΔmK(2v¯+Δv)2m¯ΔmK(2v¯Δv))+Kv¯Δm(12m¯)+KΔv(12m¯)(1+m¯),\displaystyle=-\left(\frac{2\bar{m}+\Delta m}{K(2\bar{v}+\Delta v)}-\frac{2\bar{m}-\Delta m}{K(2\bar{v}-\Delta v)}\right)+K\bar{v}\Delta m(1-2\bar{m})+K\Delta v(1-2\bar{m})(1+\bar{m}),
(6) m¯˙=\displaystyle\dot{\bar{m}}= fm¯(Δm,m¯;v¯,σ;Δv)\displaystyle f_{\bar{m}}(\Delta m,\bar{m};\bar{v},\sigma;\Delta v)
=\displaystyle= 12(2m¯+ΔmK(2v¯+Δv)2m¯ΔmK(2v¯Δv)+K(2v¯+Δv)2(12m¯)(1+2m¯+Δm2)\displaystyle\frac{1}{2}\biggl{(}-\frac{2\bar{m}+\Delta m}{K(2\bar{v}+\Delta v)}-\frac{2\bar{m}-\Delta m}{K(2\bar{v}-\Delta v)}+\frac{K(2\bar{v}+\Delta v)}{2}(1-2\bar{m})(1+\frac{2\bar{m}+\Delta m}{2})
+K(2v¯Δv)2(12m¯)(1+2m¯Δm2)σ2(2m¯+Δm)(2m¯Δm)).\displaystyle\ \ \ \ \ +\frac{K(2\bar{v}-\Delta v)}{2}(1-2\bar{m})(1+\frac{2\bar{m}-\Delta m}{2})-\frac{\sigma}{2}(2\bar{m}+\Delta m)(2\bar{m}-\Delta m)\biggr{)}.

A straightforward application of singular perturbation theory with small parameter ϵ=1/K\epsilon=1/K and coordinates x=Δmx=\Delta m, and y=(12m¯)/ϵy=(1-2\bar{m})/\epsilon yields the following result.

Theorem 1.1.

[18, Theorem 1] In the singular limit ϵ0\epsilon\to 0, the motivation dynamics (3) reduce to

(7) x˙=σ2v¯(1x2)2x+3α6+αx,\dot{x}=\frac{\sigma}{2\bar{v}}(1-x^{2})\frac{2x+3\alpha}{6+\alpha x},

where α=Δv/v¯\alpha=\Delta v/\bar{v}.

A standard nonlinear time scaling argument then allows one to eliminate the denominator 6+αx6+\alpha x from (7) and makes the connection to the unfolding of the pitchfork bifurcation (2) explicit.

Corollary 1.2.

[18, Corollary 2] The singularly-perturbed motivation dynamics (7) are equivalent to

x=x(1x2)+32α32αx2,x^{\prime}=x(1-x^{2})+\frac{3}{2}\alpha-\frac{3}{2}\alpha x^{2},

i.e., an unfolding of the pitchfork bifurcation (2) with bifurcation parameter μ1\mu\mapsto 1 and unfolding parameters α1=3α/2\alpha_{1}=3\alpha/2 and α2=3α/2\alpha_{2}=-3\alpha/2.

The equilibria of the singularly-perturbed system (7) are shown in Figure 3. Note that the system has three equilibria for all possible values of α[2,2]\alpha\in[-2,2]. When α=0\alpha=0 the options are equally valued, the pitchfork is unperturbed, and the equilibria correspond to those of the standard pitchfork (1) in the post-bifurcation regime. The equilibrium x=0x=0 is unstable, while those at x=±1x=\pm 1 are stable. For nonzero α\alpha the pitchfork unfolds. In the singularly-perturbed regime, the non-zero equilibria remain at x=±1x=\pm 1, while the intermediate equilibrium shifts to x=3α/2x=-3\alpha/2. The intermediate equilibrium is unstable for the values of α[2/3,2/3]\alpha\in[-2/3,2/3] where it exists. The equilibria at x=±1x=\pm 1 are stable when the value difference Δv\Delta v is not too biased against the corresponding option. For example, x=+1x=+1 is a stable equilibrium of (7) for α2/3\alpha\geq-2/3.

The structure of equilibria shown in Figure 3 determines the decision-making behavior of the model (3) in the singular limit. The state x=Δm=+1x=\Delta m=+1 corresponds to the system committing fully to option 1, i.e., to m=(1,0,0)m=(1,0,0). This state is stable for α=Δv/v¯2/3\alpha=\Delta v/\bar{v}\geq-2/3, and globally attracting for α>2/3\alpha>2/3. In other words, the system can commit to option 1 when when Δv2v¯/3\Delta v\geq-2\bar{v}/3, and will be globally attracted to committing to option 1 when Δv>2v¯/3\Delta v>2\bar{v}/3. Switching behavior can occur as α\alpha shifts. For example, suppose that the system (7) is initialized with α<0\alpha<0 and Δm=0.9\Delta m=-0.9, representing a commitment to option 2. If α\alpha is then raised above the value 2/32/3, the state Δm\Delta m will be attracted to the value +1+1, representing a decision to switch and commit to option 1. The rate at which Δm\Delta m is attracted to +1+1 is governed by the parameter σ\sigma, as can be seen from the form of the dynamics (7).

We note that the coefficient 3/23/2 arises from the last term KΔv(12m¯)(1+m¯)K\Delta v(1-2\bar{m})(1+\bar{m}) in (5) and can be adjusted by changing the coefficient 1+m¯1+\bar{m} to β+m¯\beta+\bar{m}, which changes the coefficient 3/2=1+1/23/2=1+1/2 to β+1/2\beta+1/2. This observation can be used to control the bifurcation properties of the system, e.g., by making it more or less sensitive to the relative value difference α\alpha. We do not pursue this line of investigation further in the present paper, but recognize that it is a natural point of departure for further work.

Refer to caption
Figure 3: Equilibria of the singularly-reduced dynamics (7) as a function of the unfolding parameter α\alpha. Solid lines represent stable equilibria; dashed lines unstable equilibria. For sufficiently large |α|>2/3|\alpha|>2/3, only one equilibrium corresponding to a preference for the high-value option is stable.

In this section, we introduced our requirements for a dynamical system model of value-sensitive decision making We showed how an unfolding of the pitchfork bifurcation can provide the fundamental mechanism for such a model in the case of N=2N=2 options, and reviewed a model due to Seeley et al. that embeds such a pitchfork mechanism along with some recent results reducing that model. In the following section we begin to construct a generalization of the Seeley et al. model to the case N>2N>2 by parsing the decision among NN options into a series of binary decisions represented by a tree structure.

2 Parsing NN-ary decisions into binary trees

Inspired by the ideas presented in the previous section, we seek to develop a dynamical systems model of value-sensitive decision-making among NN options using an unfolding of the pitchfork bifurcation. The difficulty of this approach is that constructing an unfolding of the pitchfork bifurcation in NN dimensions is non-trivial. In this paper, we construct such an unfolding, and thus the desired model, by composing a series of unfoldings of a standard pitchfork bifurcations. The composition is structured by parsing a decision among NN options into a series of binary decisions.

In this section we introduce the idea of parsing a decision among NN options into a series of binary decisions represented by a tree structure, and review a number of concepts, primarily from the computer science literature, on binary trees. As an example of an NN-ary decision, consider the case of a professor who has a variety of tasks and must decide which one to focus on at any given time. She may decide on a task by making a series of binary decisions as shown in Figure 4. For example, she might first decide between working on research or on teaching; given a decision to work on teaching, she may work on preparing a lecture or some assignments. The decision among five options is thus parsed into a series of binary decisions represented by a binary tree. A decision among an arbitrary number NN options can be parsed in this way.

We now define a number of terms associated with binary trees. A (rooted) tree is a connected acyclic undirected graph where one node is identified as the root. The parent of a node nn is the node connected to nn on the path to the root, and the children of a node nn are the nodes for which nn is the parent. Similarly, a descendant of a node nn is any node which is a child of nn or is a descendent of any of the children of nn. A sibling of a node nn is any other node which shares a parent with nn. A leaf is a node with no children, while an internal node is a node which is not a leaf. Finally, a binary tree is a tree where each node has at most two children, referred to as the left and right children. For such a tree, we refer to the descendants of a node nn associated with the right and left children as the right and left descendants, respectively. Note that when an arbitrary number NN of options is parsed into a binary tree TT, the tree can be selected such that each internal node has two children. Such a tree is referred to as a full or proper binary tree.

We now formally define a parsing of a decision set.

Definition 2.1 (Parsing).

Given a set of no2n_{o}\geq 2 options, a parsing of these options is a proper rooted binary tree TT where each leaf node represents an option.

Often, we will label the nodes with an index ii. Then, the function oo maps leaf nodes to their associated option, i.e., io(i)i\mapsto o(i). Figure 5 shows the node labels associated with the parsing shown in Figure 4. In this case, we have o(2)=o(2)= Experiment, o(3)=o(3)= Code, o(5)=o(5)= Lecture, o(7)=o(7)= Homework, and o(8)=o(8)= Exam.

ResearchExperimentCodeTeachingLectureAssignmentsHomeworkExam
Figure 4: Parsing a decision among NN options into a series of binary decisions.
012345678
Figure 5: Node labels associated with the parsing shown in Figure 4. The function oo maps leaf nodes to options, e.g., o(2)=o(2)= Experiment.

2.1 Tree traversal

Tree traversal is a fundamental process acting on a tree, wherein the process visits (and carries out an action on) each node in the tree exactly once. Trees may be traversed in either depth-first or breadth-first orders, as depicted in Figure 6. As their names imply, depth-first order operates by going as deep down the tree as possible before going to the next sibling, while breadth-first order operates by going through each sibling before going to a descendant. The nodes in Figure 6 are labeled according to the order in which they will be visited in depth-first or breadth-first traversal.

012345678 013784256
Ordered node list: (0,1,2,3,4,5,6,7,8)(0,1,2,3,4,5,6,7,8) Ordered node list: (0,1,3,7,8,4,2,5,6)
Ordered option list: (3,4,5,7,8) Ordered option list: (7,8,4,5,6)
(a) (b)
Figure 6: Depth-first (a) versus breadth-first (b) traversal of a binary tree. The nodes are labeled with numbers according to the order in which they will be visited during traversal.

2.2 Tree paths

A path in a finite tree is a finite sequence of edges which joins a sequence of nodes. For a rooted tree, there is always a unique shortest path from the root to any other node. We denote the sequence of nodes along the shortest path from the root to node ii as pip_{i} and denote its jthj^{th} element as pijp_{ij}. The sequence pip_{i} begins with the root node and ends with the node ii. The number of nodes in pip_{i} is denoted |pi||p_{i}|.

2.3 Tree isomorphisms

Trees may have important symmetries. For example, the nodes of a tree may be rearranged without changing the structure it represents. Two trees which share the same structure are said to be isomorphic. The concept of tree isomorphism is inherited from the concept of graph isomorphism [21], for which tree isomorphisms are a special case.

Definition 2.2 (Rooted tree isomorphism).

Let T1T_{1} and T2T_{2} be two rooted trees with node sets N1N_{1} and N2N_{2}, edge sets E1E_{1} and E2E_{2}, and roots r1N1r_{1}\in N_{1} and r2N2r_{2}\in N_{2}, respectively. An isomorphism of T1T_{1} and T2T_{2} is a bijection between the node sets φ:N1N2\varphi:N_{1}\to N_{2} such that

u,vN1(u,v)E1(φ(u),φ(v))E2\forall u,v\in N_{1}\ \ (u,v)\in E_{1}\Leftrightarrow(\varphi(u),\varphi(v))\in E_{2}

and φ(r1)=r2\varphi(r_{1})=r_{2}.

In words, a rooted tree isomorphism is a mapping between the node sets such that each edge is preserved, along with the root node. An example, Figure 7 shows two isomorphisms of the tree presented in Figure 6(a). Note that isomorphisms of binary trees are generated by flips at nodes, wherein the left and right descendants of a given node are exchanged.

Problems associated with tree isomorphisms arise in many applications. In particular, a standard problem in computer science is to determine whether two rooted trees T1T_{1} and T2T_{2} are isomorphic. A classic algorithm due to Aho, Hopcroft, and Ullman [22] solves the problem in O(n)O(n) time for trees with nn vertices.

067812345 015234678
Ordered node list: (0,6,7,8,1,2,3,4,5) Ordered node list: (0,1,5,2,3,4,6,7,8)
Ordered option list: (7,8,3,4,5) Ordered option list: (5,3,4,7,8)
(a) (b)
Figure 7: Tree isomorphisms are generated by flips at nodes. Here we show two isomorphisms of the tree presented in Figure 6(a), keeping the node numbers from the previous figure. Panel (a): isomorphism generated from flipping at node 0, i.e., applying γ0\gamma_{0}. Panel (b): isomorphism generated from flipping at node 1, i.e., applying γ1\gamma_{1}.

2.4 Symmetry group of a tree and of options

The set of isomorphisms of a given tree exhibit a group structure. The elements of the group are generated by the flips at internal nodes described above and the group operation is given by composition. It is clear that each flip is its own inverse, as exchanging left and right descendants of a node twice leaves the tree unchanged. Flips may be carried out in any order, so the operation is associative, and the identity is the element consisting of no flips.

When a tree TT is a parsing of a set of non_{o} options, isomorphisms of the tree generate isomorphisms of the option set {1,,no}\{1,\ldots,n_{o}\}. Recall that an isomorphism of TT is a bijection φ\varphi from the node set of TT to itself. Thus, a node ii is mapped to j=φ(i)j=\varphi(i) and the set {o(i):i{1,,no}}\{o(i):i\in\{1,\ldots,n_{o}\}\} is mapped to {o(φ(i)):i{1,,no}}\{o(\varphi(i)):i\in\{1,\ldots,n_{o}\}\}. The group of all possible isomorphisms of non_{o} objects is SnoS_{n_{o}}. Note, however, that not all such isomorphisms can be generated by the set of tree isomorphisms. For example, nodes that are siblings must remain siblings even under isomorphism operations.

Let the tree TT be a parsing of a set of non_{o} options. The set of isomorphisms the options that can be generated by isomorphisms of TT forms a group ΓTSno\Gamma_{T}\leq S_{n_{o}} whose structure is given by a wreath product of copies of S2S_{2}. This can be seen as follows. Let ii be an internal node in the tree TT and let ro(i)ro(i) and lo(i)lo(i) denote the set of options associated with its right and left descendants, respectively. For example, for the tree in Figure 7(a), ro(0)={o(3),o(4),o(5)}ro(0)=\{o(3),o(4),o(5)\} and lo(0)={o(7),o(8)}lo(0)=\{o(7),o(8)\}. Let γi\gamma_{i} represent the action of performing a flip at node ii. Then γi\gamma_{i} exchanges the sets ro(i)ro(i) and lo(i)lo(i). Explicitly, we have

(8) γi:(ro(i),lo(i))(lo(i),ro(i)).\gamma_{i}:(ro(i),lo(i))\mapsto(lo(i),ro(i)).

Note that applying γi\gamma_{i} twice results in the identity mapping, so γi\gamma_{i} generates the permutation group S2S_{2} acting on the set {ro(i),lo(i)}\{ro(i),lo(i)\}. Furthermore, any options which are not associated with the descendants of node ii are unaffected. Thus, one can think of γi\gamma_{i} as generating a representation of S2S_{2} acting on the set {1,,no}\{1,\ldots,n_{o}\}; this is trivially a subgroup of SnoS_{n_{o}}. Applying γj\gamma_{j} for another internal node jij\neq i generates another representation of S2S_{2}. The group generated by application of both actions γi\gamma_{i} and γj\gamma_{j} is the wreath product S2S2S_{2}\wr S_{2}. This process can be extended for each internal node ii, yielding a symmetry group which is the repeated wreath product of copies of S2S_{2}. Formally we have the following Proposition.

Proposition 2.3 (Symmetry group of a parsing).

Let the tree TT be a parsing of a set of non_{o} options. Denote the number of internal nodes of TT by nin_{i}. The symmetry group ΓT\Gamma_{T} associated with the isomorphisms of TT is given by the nin_{i}-fold wreath product of S2S_{2}

(9) ΓT=S2S2S2ni timesSno.\Gamma_{T}=\underbrace{S_{2}\wr S_{2}\wr\cdots\wr S_{2}}_{n_{i}\text{ times}}\leq S_{n_{o}}.

3 A recursively-defined vector field

In this section we show how to construct a decision-making vector field for a given parsing of a finite set of options. We suppose we have a finite set of non_{o} options and a tree TT which is a parsing of the options. Furthermore, each option ii is associated with a scalar vi>0v_{i}>0 that encodes its importance or value. We seek a vector field ff operating on the state space Δno\Delta^{n_{o}} with attracting fixed points associated with the high-value options.

We label the nnn_{n} nodes of TT with an index ii, with the root node having the index i=0i=0. We define the following notation to describe the tree structure. For a node ii, we denote its parent by p(i)\operatorname{p}(i) and its left and right child nodes by lc(i)\operatorname{lc}(i) and rc(i)\operatorname{rc}(i), respectively. The descendants of a node ii consist recursively of the node’s children lc(i)\operatorname{lc}(i) and rc(i)\operatorname{rc}(i) along with the descendants of the children. We denote the descendants of a node ii by d(i)\operatorname{d}(i). The descendants can be partitioned into left descendants and right descendants, which consist of the left child and its descendants and the right child and its descendants, respectively. For a node ii, we denote the left descendants by ld(i)\operatorname{ld}(i) and the right descendants by rd(i)\operatorname{rd}(i), respectively. Recall that leaf nodes represent options. As above, let o(i)o(i) be the option associated with a leaf node ii.

To each node ii we associate the state ziz_{i}\in\mathbb{R} and the value ui>0u_{i}>0. Furthermore, to each internal (non-leaf) node ii we associate states 𝐳i2,𝐦iΔ2\mathbf{z}_{i}\in\mathbb{R}^{2},\mathbf{m}_{i}\in\Delta^{2}, and 𝐯i+2\mathbf{v}_{i}\in\mathbb{R}_{+}^{2}. These additional states represent quantities associated with the node’s children. We denote the jthj^{th} component of 𝐳i,𝐦i\mathbf{z}_{i},\mathbf{m}_{i}, and 𝐯i\mathbf{v}_{i} by zij,mij,z_{ij},m_{ij}, and vijv_{ij}, respectively.

The states ziz_{i} are defined recursively as follows. Let z0=1z_{0}=1. Then, for i0,zlc(i)=zimi1i\geq 0,z_{lc(i)}=z_{i}m_{i1} and zrc(i)=zimi2z_{rc(i)}=z_{i}m_{i2}. Alternatively, for i1,zi=zp(i)mp(i)ji\geq 1,z_{i}=z_{\operatorname{p}(i)}m_{\operatorname{p}(i)j}, where

j={1,if i=lc(p(i)),2,if i=rc(p(i)).j=\begin{cases}1,&\text{if }i=\operatorname{lc}(\operatorname{p}(i)),\\ 2,&\text{if }i=\operatorname{rc}(\operatorname{p}(i)).\end{cases}

The state vector 𝐳i\mathbf{z}_{i} associated with internal node ii has components 𝐳i=(zlc(i),zrc(i))T=zi𝐦i\mathbf{z}_{i}=(z_{\operatorname{lc}(i)},z_{\operatorname{rc}(i)})^{T}=z_{i}\mathbf{m}_{i}. Note that the definition of ziz_{i} is invertible for zp(i)0z_{\operatorname{p}(i)}\neq 0: in this case, we have 𝐦i=𝐳i/zi\mathbf{m}_{i}=\mathbf{z}_{i}/z_{i}. The components are mp(i)j=zi/zp(i)m_{\operatorname{p}(i)j}=z_{i}/z_{\operatorname{p}(i)}, where j=1j=1 if i=lc(p(i))i=\operatorname{lc}(\operatorname{p}(i)) and j=2j=2 if i=rc(p(i))i=\operatorname{rc}(\operatorname{p}(i)). Similarly, for a node ii, uiu_{i} is equal to the mean of the associated children’s values

(10) ui={vo(i),if i is a leaf,(vlc(i)+vrc(i))/2,if i is an internal node.u_{i}=\begin{cases}v_{o(i)},&\text{if $i$ is a leaf},\\ (v_{\operatorname{lc}(i)}+v_{\operatorname{rc}(i)})/2,&\text{if $i$ is an internal node.}\end{cases}

Then, for an internal node ii, vi1=ulc(i)v_{i1}=u_{\operatorname{lc}(i)} and vi2=urc(i)v_{i2}=u_{\operatorname{rc}(i)}.

We endow the motivation states 𝐦i\mathbf{m}_{i} associated with an internal node ii with dynamics ˙𝐦i=f(𝐦i,𝐯i)\dot{}\mathbf{m}_{i}=f(\mathbf{m}_{i},\mathbf{v}_{i}), where ff is the Seeley et al. dynamics (3). The dynamics of the overall tree TT consists of copies of the dynamics ff defined as follows. Recall that nin_{i} is the number of internal nodes of TT and let 𝐦2ni\mathbf{m}\in\mathbb{R}^{2n_{i}} be the vector consisting of the stacked node states 𝐦i\mathbf{m}_{i}. Note that the definition of 𝐦\mathbf{m}, i.e., the order in which the 𝐦i\mathbf{m}_{i} are stacked, is arbitrary: different orders correspond to permutations of the coordinates. The structure of the dynamics is encoded in the tree structure, i.e., the parent-child relationships given by the functions p,lc\operatorname{p},\operatorname{lc}, and rc\operatorname{rc}. For practical purposes of performing computations, one chooses a scheme for numbering the coordinates that corresponds to a scheme for traversing the nodes of the tree. We choose to traverse the tree TT in a depth-first manner and define 𝐦\mathbf{m} by

𝐦=(𝐦0T,𝐦lc(0)T,𝐦lc(lc(0))T,left descendants of node 0,𝐦rc(0)T,𝐦lc(rc(0))T,right descendants of node 0)T.\mathbf{m}=(\mathbf{m}_{0}^{T},\underbrace{\mathbf{m}_{\operatorname{lc}(0)}^{T},\mathbf{m}_{\operatorname{lc}(\operatorname{lc}(0))}^{T},\ldots}_{\text{left descendants of node }0},\underbrace{\mathbf{m}_{\operatorname{rc}(0)}^{T},\mathbf{m}_{\operatorname{lc}(\operatorname{rc}(0))}^{T},\ldots}_{\text{right descendants of node }0})^{T}.

The dynamics of 𝐦\mathbf{m} are defined by stacking the dynamics of the component states 𝐦i\mathbf{m}_{i}:

(11) ˙𝐦=fm(𝐦,𝐯)\displaystyle\dot{}\mathbf{m}=f_{m}(\mathbf{m},\mathbf{v}) =(˙𝐦0T,˙𝐦lc(0)T,˙𝐦lc(lc(0))T,,˙𝐦rc(0)T,˙𝐦lc(rc(0))T,)T\displaystyle=(\dot{}\mathbf{m}_{0}^{T},\dot{}\mathbf{m}_{\operatorname{lc}(0)}^{T},\dot{}\mathbf{m}_{\operatorname{lc}(\operatorname{lc}(0))}^{T},\ldots,\dot{}\mathbf{m}_{\operatorname{rc}(0)}^{T},\dot{}\mathbf{m}_{\operatorname{lc}(\operatorname{rc}(0))}^{T},\ldots)^{T}
=(f(𝐦0,𝐯0)T,f(𝐦lc(0),𝐯lc(0))T,f(𝐦lc(lc(0)),𝐯lc(lc(0)))T,,\displaystyle=(f(\mathbf{m}_{0},\mathbf{v}_{0})^{T},f(\mathbf{m}_{\operatorname{lc}(0)},\mathbf{v}_{\operatorname{lc}(0)})^{T},f(\mathbf{m}_{\operatorname{lc}(\operatorname{lc}(0))},\mathbf{v}_{\operatorname{lc}(\operatorname{lc}(0))})^{T},\ldots,
f(𝐦rc(0),𝐯rc(0))T,f(𝐦lc(rc(0)),𝐯lc(rc(0)))T,)T,\displaystyle\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ f(\mathbf{m}_{\operatorname{rc}(0)},\mathbf{v}_{\operatorname{rc}(0)})^{T},f(\mathbf{m}_{\operatorname{lc}(\operatorname{rc}(0))},\mathbf{v}_{\operatorname{lc}(\operatorname{rc}(0))})^{T},\ldots)^{T},

where 𝐯2ni\mathbf{v}\in\mathbb{R}^{2n_{i}} is the vector of the node value states 𝐯i\mathbf{v}_{i} stacked in depth-first order.

The dynamics of 𝐦i\mathbf{m}_{i} then defines dynamics of the states 𝐳i\mathbf{z}_{i} by a simple change of coordinates. Recall that 𝐳i=(zlc(i),zrc(i))T=zi𝐦i\mathbf{z}_{i}=(z_{\operatorname{lc}(i)},z_{\operatorname{rc}(i)})^{T}=z_{i}\mathbf{m}_{i} with z0=1z_{0}=1. Then ˙𝐳i=z˙i𝐦i+zi˙𝐦i=z˙i𝐦i+zif(𝐦i,𝐯i)\dot{}\mathbf{z}_{i}=\dot{z}_{i}\mathbf{m}_{i}+z_{i}\dot{}\mathbf{m}_{i}=\dot{z}_{i}\mathbf{m}_{i}+z_{i}f(\mathbf{m}_{i},\mathbf{v}_{i}). As above, we construct the vector 𝐳2ni\mathbf{z}\in\mathbb{R}^{2n_{i}} by stacking the individual 𝐳i\mathbf{z}_{i} in depth-first order. When zi0z_{i}\neq 0, we have 𝐦i=𝐳i/zi\mathbf{m}_{i}=\mathbf{z}_{i}/z_{i}, so ˙𝐳i\dot{}\mathbf{z}_{i} can be written in terms of 𝐳\mathbf{z}. We denote the resulting dynamics of 𝐳\mathbf{z} by

(12) ˙𝐳=fz(𝐳,𝐯).\dot{}\mathbf{z}=f_{z}(\mathbf{z},\mathbf{v}).

In the following section we show that the functions fmf_{m} and fzf_{z} are equivariant under changes of coordinates that correspond to isomorphisms of the tree TT.

Let 𝐦oΔno\mathbf{m}_{o}\in\Delta^{n_{o}} be the state that represents the system’s decision among the non_{o} options. We relate the state 𝐳\mathbf{z} to 𝐦o\mathbf{m}_{o} by projection onto no+1\mathbb{R}^{n_{o}+1}. Let o1:{1,,no}{1,,nn}o^{-1}:\{1,\ldots,n_{o}\}\to\{1,\ldots,n_{n}\} be the function that maps from an option ii to its associated leaf node.

Definition 3.1 (Projected dynamics).

Let TT be a parsing of non_{o} options, let gg be the dynamics defined by (12), and let h:nnno+1h:\mathbb{R}^{n_{n}}\to\mathbb{R}^{n_{o}+1} be the projection that reads off the elements of 𝐳\mathbf{z} that correspond to the leaves of the tree TT. Explicitly, h:𝐳𝐦oh:\mathbf{z}\mapsto\mathbf{m}_{o} by

(13) (𝐦o)i=moi={zo1(i),i{1,2,,no}1i=1nomoi,i=no+1.(\mathbf{m}_{o})_{i}=m_{oi}=\begin{cases}z_{o^{-1}(i)},&i\in\{1,2,\ldots,n_{o}\}\\ 1-\sum_{i=1}^{n_{o}}m_{oi},&i=n_{o}+1.\end{cases}

The projection hh then defines the dynamics of the projected state 𝐦o\mathbf{m}_{o} by

(14) ˙𝐦o=fo(𝐳,𝐯)=h(fz(𝐳,𝐯)).\dot{}\mathbf{m}_{o}=f_{o}(\mathbf{z},\mathbf{v})=h(f_{z}(\mathbf{z},\mathbf{v})).

The projected dynamics (14) leave the simplex Δno\Delta^{n_{o}} invariant.

Theorem 3.2.

Let TT be a parsing of non_{o} options, let fzf_{z} be the dynamics defined by (12), and let h:nnno+1h:\mathbb{R}^{n_{n}}\to\mathbb{R}^{n_{o}+1} be the projection (13) that reads off the elements of 𝐳\mathbf{z} that correspond to the leaves of the tree TT. Let fof_{o} be the dynamics defined by (14). Then fof_{o} leaves the simplex Δno\Delta^{n_{o}} invariant.

Proof 3.3.

Let 𝐦\mathbf{m} be the projected state. Proving the claim reduces to showing that mi0m_{i}\geq 0 and that i=1nomi1\sum_{i=1}^{n_{o}}m_{i}\leq 1. We proceed by induction from the bottom of the tree. Let ii be a generic leaf node of TT. By definition, TT is a proper binary tree, so ii has a sibling. Furthermore, since ii is a leaf node, it has a parent. Let jj denote the sibling of ii and p(i)\operatorname{p}(i) the common parent of ii and jj. By the definition of the projection hh, we have mo(i)=zi,mo(j)=zjm_{o(i)}=z_{i},m_{o(j)}=z_{j}. By the definition of 𝐳i\mathbf{z}_{i}, we have

𝐳p(i)=(zi,zj)=zp(i)𝐦p(i)=zp(i)(mp(i)1,mp(i)2).\mathbf{z}_{\operatorname{p}(i)}=(z_{i},z_{j})=z_{\operatorname{p}(i)}\mathbf{m}_{\operatorname{p}(i)}=z_{\operatorname{p}(i)}(m_{\operatorname{p}(i)1},m_{\operatorname{p}(i)2}).

Note that the dynamics (3) of 𝐦p(i)\mathbf{m}_{\operatorname{p}(i)} leave the simplex Δ2\Delta^{2} invariant. This implies that the components mp(i)j0m_{\operatorname{p}(i)j}\geq 0 for i=1,2i=1,2, and

(15) mp(i)1+mp(i)21.m_{\operatorname{p}(i)1}+m_{\operatorname{p}(i)2}\leq 1.

Furthermore, we have that zi=mo(i)z_{i}=m_{o(i)} and zj=mo(j)z_{j}=m_{o(j)} are both non-negative. Multiplying the expression (15) by zp(i)z_{\operatorname{p}(i)} yields the bound

mo(i)+mo(j)=zi+zjzp(i).m_{o(i)}+m_{o(j)}=z_{i}+z_{j}\leq z_{\operatorname{p}(i)}.

Thus, the sum of the mim_{i} associated with descendants of node k=p(i)k=\operatorname{p}(i) is upper bounded by zkz_{k}. The analogous argument holds for the parent of node kk, and thus we can inductively work our way up the tree. At each node ll, the sum of the mim_{i} associated with descendants of node ll is upper bounded by zlz_{l}.

The base case of the inductive argument is the root node i=0i=0. All non_{o} options are descendant from the root node, so we have i=1nomiz0=1\sum_{i=1}^{n_{o}}m_{i}\leq z_{0}=1, as desired.

4 The vector field is equivariant under tree isomorphisms

Recall from Definition 2.2 above that two rooted trees T1T_{1} and T2T_{2} are said to be isomorphic if there exists a bijection mapping between the nodes of T1T_{1} and T2T_{2} that preserves the root node. The vector field fzf_{z} defined by (12) and its projection fof_{o} defined by (14) are equivariant under changes of coordinates which correspond to tree isomorphisms.

Let T1T_{1} and T2T_{2} be two isomorphic trees. By definition they must have the same number nnn_{n} of nodes. The isomorphism between the trees is a bijection between the node sets of T1T_{1} and T2T_{2}. In other words, it is a bijective map σ:{1,,nn}{1,,nn}\sigma:\{1,\ldots,n_{n}\}\to\{1,\ldots,n_{n}\}. This is precisely the definition of a permutation.

Definition 4.1.

Let T1T_{1} and T2T_{2} be two isomorphic trees each with nnn_{n} nodes. The map σ:{1,,nn}{1,,nn}\sigma:\{1,\ldots,n_{n}\}\to\{1,\ldots,n_{n}\} associated with the isomorphism between the trees is called the node permutation corresponding to the isomorphism.

The dynamics (12) obey symmetries that correspond to isomorphisms of the underlying tree TT. Formally, the dynamics (12) are said to be equivariant.

Definition 4.2.

Let X=nX=\mathbb{R}^{n} and suppose that Γ\Gamma is a compact Lie group acting on XX. Then a mapping F:X×XF:X\times\mathbb{R}\to X is Γ\Gamma-equivariant if and only if

F(γx,λ)=γF(x,λ)F(\gamma x,\lambda)=\gamma F(x,\lambda)

for all γΓ\gamma\in\Gamma, where λ\lambda\in\mathbb{R} is a parameter.

The Seeley et al. dynamics (3) obey a symmetry that correspond to swapping the labels of the two options. When the option values are identical, the dynamics are S2S_{2}-equivariant.

Lemma 4.3.

Let π2S2\pi_{2}\in S_{2} represent the permutation of two elements. The Seeley et al. dynamics ff defined by (3) are preserved under the action of π2\pi_{2}. Specifically, we have

f(π2𝐦,π2𝐯)=π2f(𝐦,𝐯).f(\pi_{2}\mathbf{m},\pi_{2}\mathbf{v})=\pi_{2}f(\mathbf{m},\mathbf{v}).

When 𝐯=(v,v)T\mathbf{v}=(v,v)^{T}, f(π2𝐦,𝐯)=π2f(𝐦,𝐯)f(\pi_{2}\mathbf{m},\mathbf{v})=\pi_{2}f(\mathbf{m},\mathbf{v}), and the dynamics are S2S_{2}-equivariant.

Proof 4.4.

The first statement is proven by straightforward substitution. From (3), we have

f(π2𝐦,π2𝐯)=[v2(1m1m2)m2(1v2v2(1m1m2)+σ(m1m2))v1(1m1m2)m1(1v1v1(1m1m2)+σ(m1m2))]=π2f(𝐦,𝐯).f(\pi_{2}\mathbf{m},\pi_{2}\mathbf{v})=\begin{bmatrix}v_{2}(1-m_{1}-m_{2})-m_{2}\left(\frac{1}{v_{2}}-v_{2}(1-m_{1}-m_{2})+\sigma(m_{1}m_{2})\right)\\ v_{1}(1-m_{1}-m_{2})-m_{1}\left(\frac{1}{v_{1}}-v_{1}(1-m_{1}-m_{2})+\sigma(m_{1}m_{2})\right)\end{bmatrix}=\pi_{2}f(\mathbf{m},\mathbf{v}).

The second statement follows by noting that 𝐯=(v,v)T\mathbf{v}=(v,v)^{T} implies π2𝐯=𝐯\pi_{2}\mathbf{v}=\mathbf{v}. Then, we have f(π2𝐦,𝐯)=f(π2𝐦,π2𝐯)=π2f(𝐦,𝐯)f(\pi_{2}\mathbf{m},\mathbf{v})=f(\pi_{2}\mathbf{m},\pi_{2}\mathbf{v})=\pi_{2}f(\mathbf{m},\mathbf{v}).

Let TT be a parsing of non_{o} options. Recall from Proposition 2.3 that the set of isomorphisms associated with TT form a group denoted ΓT\Gamma_{T}. These isomorphisms are represented by permutations σ\sigma. The dynamics (12) obey symmetries corresponding to the permutations associated with ΓT\Gamma_{T}. When the option values are all identical, the dynamics are ΓT\Gamma_{T}-equivariant.

Lemma 4.5.

The dynamics (11) defined by a tree TT are preserved under isomorphisms of TT. Explicitly, let TT be a parsing of non_{o} options, fmf_{m} be the dynamics (11), and let γΓT\gamma\in\Gamma_{T}. Then, fm(γ𝐦,γ𝐯)=γfm(𝐦,𝐯)f_{m}(\gamma\mathbf{m},\gamma\mathbf{v})=\gamma f_{m}(\mathbf{m},\mathbf{v}).

Proof 4.6.

Let 𝐦\mathbf{m} represent the coordinates of (11) that result from the default depth-first parsing of the tree TT. Let γiΓT\gamma_{i}\in\Gamma_{T} represent the operation of flipping the tree TT at internal node ii. Note that any γΓT\gamma\in\Gamma_{T} can be represented as a composition of several flips γi\gamma_{i}, so it suffices to show that fm(γi𝐦,γi𝐯)=γifm(𝐦,𝐯)f_{m}(\gamma_{i}\mathbf{m},\gamma_{i}\mathbf{v})=\gamma_{i}f_{m}(\mathbf{m},\mathbf{v}) for any flip γi\gamma_{i}.

Let TiT_{i}^{\prime} be the tree that results from flipping TT at the internal node ii, and let 𝐦\mathbf{m}^{\prime} represent the coordinates of (11) that result from the depth-first traversal of the tree TT. The dynamics (11) take the form ˙𝐦=fm(𝐦,𝐯)\dot{}\mathbf{m}=f_{m}(\mathbf{m},\mathbf{v}) in the coordinates associated with tree TT and ˙𝐦=fm(𝐦,𝐯)\dot{}\mathbf{m}^{\prime}=f_{m}(\mathbf{m}^{\prime},\mathbf{v}^{\prime}) in the coordinates associated with TT^{\prime}. Note that 𝐯\mathbf{v}^{\prime} represents 𝐯\mathbf{v} in the coordinates associated with TT^{\prime}.

The action of γi\gamma_{i} permutes the descendants of node ii, and in particular swaps the right and left children of ii: γi:(mi1,mi2)(mi2,mi1)\gamma_{i}:(m_{i1},m_{i2})\mapsto(m_{i2},m_{i1}). Compactly, this can be written as π2(mi1,mi2)=(mi2,mi1),\pi_{2}(m_{i1},m_{i2})=(m_{i2},m_{i1}), where π2S2\pi_{2}\in S_{2} represents the permutation of two elements. The relation between 𝐦\mathbf{m} and 𝐦\mathbf{m}^{\prime} is as follows

𝐦\displaystyle\mathbf{m} =(𝐦0T,,𝐦iT,𝐦lc(i)T,𝐦lc(lc(i))T,left descendants of node i,𝐦rc(i)T,𝐦lc(rc(i))T,right descendants of node i,)T\displaystyle=(\mathbf{m}_{0}^{T},\ldots,\mathbf{m}_{i}^{T},\underbrace{\mathbf{m}_{\operatorname{lc}(i)}^{T},\mathbf{m}_{\operatorname{lc}(\operatorname{lc}(i))}^{T},\ldots}_{\text{left descendants of node }i},\underbrace{\mathbf{m}_{\operatorname{rc}(i)}^{T},\mathbf{m}_{\operatorname{lc}(\operatorname{rc}(i))}^{T},\ldots}_{\text{right descendants of node }i},\ldots)^{T}
𝐦\displaystyle\mathbf{m}^{\prime} =(𝐦0T,,π2𝐦iT,𝐦rc(i)T,𝐦lc(rc(i))T,right descendants of node i,𝐦lc(i)T,𝐦lc(lc(i))T,left descendants of node i,)T,\displaystyle=(\mathbf{m}_{0}^{T},\ldots,\pi_{2}\mathbf{m}_{i}^{T},\underbrace{\mathbf{m}_{\operatorname{rc}(i)}^{T},\mathbf{m}_{\operatorname{lc}(\operatorname{rc}(i))}^{T},\ldots}_{\text{right descendants of node }i},\underbrace{\mathbf{m}_{\operatorname{lc}(i)}^{T},\mathbf{m}_{\operatorname{lc}(\operatorname{lc}(i))}^{T},\ldots}_{\text{left descendants of node }i},\ldots)^{T},

where lc\operatorname{lc} and rc\operatorname{rc} are the child relationships associated with tree TT.

Recall that 𝐦i=(mi1,mi2)\mathbf{m}_{i}=(m_{i1},m_{i2}) obeys the dynamics ˙𝐦i=f(𝐦i,𝐯i)\dot{}\mathbf{m}_{i}=f(\mathbf{m}_{i},\mathbf{v}_{i}) given by (3). By Lemma 4.3, we have f(π2𝐦i,π2𝐯i)=π2f(𝐦i,𝐯i)f(\pi_{2}\mathbf{m}_{i},\pi_{2}\mathbf{v}_{i})=\pi_{2}f(\mathbf{m}_{i},\mathbf{v}_{i}), so the action of γi\gamma_{i} leaves the dynamics of 𝐦i\mathbf{m}_{i} equivariant. It remains to study the action of γi\gamma_{i} on the descendants of node ii. As seen above, the action of γi\gamma_{i} on these descendants is a block permutation, mapping (𝐦lc(i),𝐦rc(i))(𝐦rc(i),𝐦lc(i))(\mathbf{m}_{\operatorname{lc}(i)},\mathbf{m}_{\operatorname{rc}(i)})\mapsto(\mathbf{m}_{\operatorname{rc}(i)},\mathbf{m}_{\operatorname{lc}(i)}), etc. The dynamics of each block jj is given by ˙𝐦j=f(𝐦j,𝐯j)\dot{}\mathbf{m}_{j}=f(\mathbf{m}_{j},\mathbf{v}_{j}) and the overall dynamics fmf_{m} is a simple stacking of copies of ff. Since the action of γi\gamma_{i} permutes both the blocks of 𝐦\mathbf{m} and 𝐯\mathbf{v} in the same way, the dynamics fm(𝐦,𝐯)f_{m}(\mathbf{m}^{\prime},\mathbf{v}^{\prime}) consists of a permutation of the blocks of fm(𝐦,𝐯)f_{m}(\mathbf{m},\mathbf{v}). Thus, we have

fm(𝐦,𝐯)=fm(γi𝐦,γi𝐯)=γifm(𝐦,𝐯)f_{m}(\mathbf{m}^{\prime},\mathbf{v}^{\prime})=f_{m}(\gamma_{i}\mathbf{m},\gamma_{i}\mathbf{v})=\gamma_{i}f_{m}(\mathbf{m},\mathbf{v})

for any flip γi\gamma_{i}. The result follows by recalling that a generic γΓT\gamma\in\Gamma_{T} can be represented by the composition of several flips γi\gamma_{i}.

Theorem 4.7.

Let the conditions for Lemma 4.5 be satisfied and suppose that 𝐯=v𝟏no\mathbf{v}=v\mathbf{1}_{n_{o}}, where 𝟏nono\mathbf{1}_{n_{o}}\in\mathbb{R}^{n_{o}} is the vector with all entries equal to 1, i.e., when vi=vi{1,,no}v_{i}=v\forall i\in\{1,\ldots,n_{o}\}. Then, the dynamics fzf_{z} defined by (12) are equivariant under permutations of 𝐳\mathbf{z} corresponding to isomorphisms of TT.

Proof 4.8.

Let ˙𝐳=fz(𝐳,𝐯)\dot{}\mathbf{z}=f_{z}(\mathbf{z},\mathbf{v}) be the dynamics (12). Let ˙𝐦=fm(𝐦,𝐯)\dot{}\mathbf{m}=f_{m}(\mathbf{m},\mathbf{v}) be the dynamics (11). Note that the vector fields fzf_{z} and fmf_{m} are related by a change of coordinates 𝐳=g(𝐦)\mathbf{z}=g(\mathbf{m}) that is invertible away from the origin 𝐦=0\mathbf{m}=0. It is clear that g(γ𝐦)=γg(𝐦)γΓTg(\gamma\mathbf{m})=\gamma g(\mathbf{m})\forall\gamma\in\Gamma_{T}. Then, elementary calculus yields

(16) fz(𝐳,𝐯)=˙𝐳=ddtg(𝐦)=g𝐦˙𝐦=g𝐦fm(𝐦,𝐯)=g𝐦(g1(𝐳))fm(g1(𝐳),𝐯).f_{z}(\mathbf{z},\mathbf{v})=\dot{}\mathbf{z}=\frac{d}{dt}g(\mathbf{m})=\frac{\partial g}{\partial\mathbf{m}}\dot{}\mathbf{m}=\frac{\partial g}{\partial\mathbf{m}}f_{m}(\mathbf{m},\mathbf{v})=\frac{\partial g}{\partial\mathbf{m}}(g^{-1}(\mathbf{z}))f_{m}(g^{-1}(\mathbf{z}),\mathbf{v}).

Analogously, we have fz(γ𝐳,γ𝐯)=g𝐦(g1(γ𝐳))fm(g1(γ𝐳),γ𝐯)f_{z}(\gamma\mathbf{z},\gamma\mathbf{v})=\frac{\partial g}{\partial\mathbf{m}}(g^{-1}(\gamma\mathbf{z}))f_{m}(g^{-1}(\gamma\mathbf{z}),\gamma\mathbf{v}). The fact that g(γ𝐦)=γg(𝐦)=γ𝐳g(\gamma\mathbf{m})=\gamma g(\mathbf{m})=\gamma\mathbf{z} implies that γg1(𝐳)=γ𝐦=g1(γ𝐳)\gamma g^{-1}(\mathbf{z})=\gamma\mathbf{m}=g^{-1}(\gamma\mathbf{z}). The chain rule yields g(γ𝐦)𝐦=g(γ𝐦)𝐦γ\frac{\partial g(\gamma\mathbf{m})}{\partial\mathbf{m}}=\frac{\partial g(\gamma\mathbf{m})}{\partial\mathbf{m}}\gamma. Similarly, g(γ𝐦)=γg(𝐦)g(\gamma\mathbf{m})=\gamma g(\mathbf{m}) implies that g(γ𝐦)𝐦=γg(𝐦)𝐦\frac{\partial g(\gamma\mathbf{m})}{\partial\mathbf{m}}=\gamma\frac{\partial g(\mathbf{m})}{\partial\mathbf{m}}. Finally, note that γγ\gamma\cdot\gamma is equal to the identity for any γΓT\gamma\in\Gamma_{T}. Putting these facts together yields

fz(γ𝐳,γ𝐯)\displaystyle f_{z}(\gamma\mathbf{z},\gamma\mathbf{v}) =g𝐦(g1(γ𝐳))fm(g1(γ𝐳),γ𝐯)\displaystyle=\frac{\partial g}{\partial\mathbf{m}}(g^{-1}(\gamma\mathbf{z}))f_{m}(g^{-1}(\gamma\mathbf{z}),\gamma\mathbf{v})
=γg1(𝐳)𝐦γfm(γg1(𝐳),γ𝐯)\displaystyle=\gamma\frac{\partial g^{-1}(\mathbf{z})}{\partial\mathbf{m}}\gamma f_{m}(\gamma g^{-1}(\mathbf{z}),\gamma\mathbf{v})
=γg1(𝐳)𝐦γγfm(g1(𝐳),𝐯)=γfz(𝐳,𝐯).\displaystyle=\gamma\frac{\partial g^{-1}(\mathbf{z})}{\partial\mathbf{m}}\gamma\cdot\gamma f_{m}(g^{-1}(\mathbf{z}),\mathbf{v})=\gamma f_{z}(\mathbf{z},\mathbf{v}).

Thus, the dynamics fzf_{z} obey the same tree isomorphism symmetry as the dynamics fmf_{m}. When 𝐯=v𝟏\mathbf{v}=v\mathbf{1}, γ𝐯=𝐯γΓT\gamma\mathbf{v}=\mathbf{v}\forall\gamma\in\Gamma_{T}. Then fz(γ𝐳,𝐯)=g(γ𝐳,γ𝐯)=γg(𝐳,𝐯)f_{z}(\gamma\mathbf{z},\mathbf{v})=g(\gamma\mathbf{z},\gamma\mathbf{v})=\gamma g(\mathbf{z},\mathbf{v}), the desired result.

The implication of Lemma 4.5 and Theorem 4.7 is that the fundamental structure of the dynamics (11) and (12) is encoded in structure of the parsing TT. Furthermore, when all the options have equal values, they are treated the same in the sense that by the dynamics of the corresponding states are unchanged by permutation of the coordinates. When the option values differ, however, these symmetries can be broken. The symmetry breaking can be understood by studying the bifurcation properties of the vector field.

5 Bifurcation properties of the equivariant vector field

The Seeley et al. dynamics (3) decide between two options using a pitchfork bifurcation that unfolds as the values of the two options differ. The dynamics (11) and (12) introduced in Section 3 embed multiple copies of the pitchfork bifurcation inherited from (3). In this section we make this statement precise. We begin by recalling the definition of a kk-parameter unfolding of a bifurcation.

Definition 5.1 ([16]).

Let f(x,λ)=0f(x,\lambda)=0 be an equation which undergoes a bifurcation as λ\lambda\in\mathbb{R} is varied. An unfolding of ff is a parametrized family of functions F(x,λ,α),αkF(x,\lambda,\alpha),\alpha\in\mathbb{R}^{k}, such that F(x,λ,0)=f(x,λ)F(x,\lambda,0)=f(x,\lambda). One refers to FF as a kk-parameter unfolding of ff.

We now recall the formal bifurcation result concerning the Seeley et al. dynamics (3).

Theorem 5.2.

[1, 3] Let ˙𝐦=f(𝐦,𝐯)\dot{}\mathbf{m}=f(\mathbf{m},\mathbf{v}) be the dynamics (3) and let 𝐯=v𝟏+2\mathbf{v}=v\mathbf{1}\in\mathbb{R}_{+}^{2} be the vector with both entries equal to v>1v>1. The dynamics undergo a pitchfork bifurcation as the parameter σ\sigma increases through a critical value given by

(17) σ=4v3(v21)2.\sigma=\frac{4v^{3}}{(v^{2}-1)^{2}}.

Equivalently, for fixed σ\sigma, the dynamics undergo a pitchfork bifurcation as the parameter vv increases through the critical value v=vv=v^{*} solving (17).

Proof 5.3.

When 𝐯=v𝟏\mathbf{v}=v\mathbf{1}, straightforward computation shows that the dynamics ff defined by (3) have an equilibrium 𝐦=m¯𝟏\mathbf{m}=\bar{m}\mathbf{1}, where m¯\bar{m} satisfies

(18) m¯=(1+v2)+1+2v2+4σv3+9v42v(2v+σ).\bar{m}=\frac{-(1+v^{2})+\sqrt{1+2v^{2}+4\sigma v^{3}+9v^{4}}}{2v(2v+\sigma)}.

Evaluating the Jacobian of ff yields

(19) J\displaystyle J =[1v1v1(1+m1)+v1mUσm2v1(1+m1)σm1v2(1+m2)1v2v2(1+m2)+v2mUσm1]|(𝐦,𝐯)=(m¯𝟏,v𝟏)\displaystyle=\left.\begin{bmatrix}-\frac{1}{v_{1}}-v_{1}(1+m_{1})+v_{1}m_{U}-\sigma m_{2}&-v_{1}(1+m_{1})-\sigma m_{1}\\ -v_{2}(1+m_{2})&-\frac{1}{v_{2}}-v_{2}(1+m_{2})+v_{2}m_{U}-\sigma m_{1}\end{bmatrix}\right|_{(\mathbf{m},\mathbf{v})=(\bar{m}\mathbf{1},v\mathbf{1})}
=[1v3vm¯σm¯v(1+m¯)σm¯v(1+m¯)σm¯1v3vm¯σm¯].\displaystyle=\begin{bmatrix}-\frac{1}{v}-3v\bar{m}-\sigma\bar{m}&-v(1+\bar{m})-\sigma\bar{m}\\ -v(1+\bar{m})-\sigma\bar{m}&-\frac{1}{v}-3v\bar{m}-\sigma\bar{m}\end{bmatrix}.

The eigenvalues of JJ are λ1=2m¯v2+v21v,λ2=4m¯v22m¯σvv21v\lambda_{1}=\frac{-2\bar{m}v^{2}+v^{2}-1}{v},\lambda_{2}=\frac{-4\bar{m}v^{2}-2\bar{m}\sigma v-v^{2}-1}{v}. Consider λ1\lambda_{1} and λ2\lambda_{2} as functions of σ\sigma. Simple substitution shows that λ2<0σ>0\lambda_{2}<0\forall\sigma>0 and that λ1\lambda_{1} smoothly increases through zero as σ\sigma increases through the value σ\sigma^{*} defined by (17).

As shown in Corollary 1.2, the pitchfork bifurcation embedded in the S2S_{2}-equivariant dynamics (3) unfolds as a function of a single parameter α=2(v1v2)/(v1+v2)\alpha=2(v_{1}-v_{2})/(v_{1}+v_{2}). Note that when v1=v2,α=0v_{1}=v_{2},\alpha=0. The bifurcation and unfolding properties of (3) carry over naturally to the dynamics (11).

Note that the dynamics fmf_{m} defined by (11) consist of stacked copies of (3), so the Jacobian of fmf_{m} is a block diagonal matrix whose diagonal blocks are copies of JJ defined in (19). The singularity of fmf_{m} then unfolds as a function of ni=no1n_{i}=n_{o}-1 parameters αi=2(vi1vi2)/(vi1+vi2)\alpha_{i}=2(v_{i1}-v_{i2})/(v_{i1}+v_{i2}). Formally, we have the following theorem.

Theorem 5.4.

Let TT be a parsing of non_{o} options consisting of nin_{i} internal nodes. Let ΓT\Gamma_{T} be the isomorphism group of TT. Let ˙𝐦=fm(𝐦,𝐯)\dot{}\mathbf{m}=f_{m}(\mathbf{m},\mathbf{v}) be the dynamics (11) defined by TT and let vi=v>1v_{i}=v>1 for each option i{1,,no}i\in\{1,\ldots,n_{o}\} Then,

  1. i).

    The vector field fmf_{m} has an equilibrium 𝐦=m¯𝟏2ni\mathbf{m}=\bar{m}\mathbf{1}_{2n_{i}}, where m¯\bar{m} is defined by (18).

  2. ii).

    The vector field has a singularity at (𝐦,v,σ)=(m¯𝟏2ni,v,σ)(\mathbf{m},v,\sigma)=(\bar{m}\mathbf{1}_{2n_{i}},v,\sigma), where σ\sigma and vv are related by (17). The singularity is a ΓT\Gamma_{T}-equivariant bifurcation that consists of nin_{i} copies of the standard S2S_{2} pitchfork bifurcation.

  3. iii).

    When 𝐯\mathbf{v} is perturbed away from 𝐯=v𝟏\mathbf{v}=v\mathbf{1} the system fm(𝐦,𝐯)f_{m}(\mathbf{m},\mathbf{v}) is a no1n_{o}-1-parameter unfolding of the ΓT\Gamma_{T}-equivariant bifurcation.

Proof 5.5.

Recall from (11) that fm(𝐦,𝐯)f_{m}(\mathbf{m},\mathbf{v}) consists of stacked copies of the dynamics f(𝐦i,𝐯i)f(\mathbf{m}_{i},\mathbf{v}_{i}) defined by (3). Since vi=vi{1,,no}v_{i}=v\forall i\in\{1,\ldots,n_{o}\}, 𝐯i=v𝟏2\mathbf{v}_{i}=v\mathbf{1}_{2} and 𝐯=v𝟏2ni\mathbf{v}=v\mathbf{1}_{2n_{i}}. Thus, the ithi^{th} block of fm(𝐦,𝐯)f_{m}(\mathbf{m},\mathbf{v}) is equal to f(𝐦i,𝐯i)f(\mathbf{m}_{i},\mathbf{v}_{i}), which has an equilibrium 𝐦i=m¯𝟏2\mathbf{m}_{i}=\bar{m}\mathbf{1}_{2} as seen in the proof of Theorem 5.2. The equilibrium of ff follows by stacking the blocks 𝐦i\mathbf{m}_{i}, which proves statement i).

For statement ii), note that since the elements of 𝐦\mathbf{m} are stacked in the same order as those of fm(𝐦,𝐯)f_{m}(\mathbf{m},\mathbf{v}), the Jacobian of fm(𝐦,𝐯)f_{m}(\mathbf{m},\mathbf{v}) is a block diagonal matrix with the ithi^{th} diagonal block being the Jacobian of f(𝐦i,𝐯i)f(\mathbf{m}_{i},\mathbf{v}_{i}). Thus, evaluating the Jacobian of fm(𝐦,𝐯)f_{m}(\mathbf{m},\mathbf{v}) at the equilibrium in statement i) yields a block diagonal matrix

Jm=fm(𝐦,𝐯)𝐦|(𝐦,𝐯)=(m¯𝟏2ni,v𝟏2ni)=[J000J000J]2ni×2ni,J_{m}=\left.\frac{\partial f_{m}(\mathbf{m},\mathbf{v})}{\partial\mathbf{m}}\right|_{(\mathbf{m},\mathbf{v})=(\bar{m}\mathbf{1}_{2n_{i}},v\mathbf{1}_{2n_{i}})}=\begin{bmatrix}J&0&\ldots&0\\ 0&J&\ldots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&0&\ldots&J\end{bmatrix}\in\mathbb{R}^{2n_{i}\times 2n_{i}},

where JJ is the matrix defined in (19). The eigenvalues of JmJ_{m} are λ1=2m¯v2+v21v\lambda_{1}=\frac{-2\bar{m}v^{2}+v^{2}-1}{v} and λ2=4m¯v22m¯σvv21v\lambda_{2}=\frac{-4\bar{m}v^{2}-2\bar{m}\sigma v-v^{2}-1}{v}, each with multiplicity nin_{i}. As shown in the proof of Theorem 5.2, λ1=0\lambda_{1}=0 when σ\sigma and vv are related by (17), so there is a singularity at (𝐦,v,σ)=(m¯𝟏2ni,v,σ)(\mathbf{m},v,\sigma)=(\bar{m}\mathbf{1}_{2n_{i}},v,\sigma). This singularity consists of nin_{i} copies of the S2S_{2} pitchfork bifurcation embedded in JJ.

For statement iii), consider how the option values vi,i{1,,no}v_{i},i\in\{1,\ldots,n_{o}\} are related to the value vector 𝐯2ni\mathbf{v}\in\mathbb{R}^{2n_{i}}. The vector 𝐯\mathbf{v} consists of nin_{i} blocks 𝐯i\mathbf{v}_{i} whose components are defined by (10), one for each internal node. Note that, since TT is a full binary tree, it is a well-established fact [23] that ni=n01n_{i}=n_{0}-1. Each 𝐯i,i{1,,ni}\mathbf{v}_{i},i\in\{1,\ldots,n_{i}\} corresponds to an unfolding parameter αi=2(vi1vi2)/(vi1+vi2)\alpha_{i}=2(v_{i1}-v_{i2})/(v_{i1}+v_{i2}). Each αi\alpha_{i} is an unfolding parameter, since 𝐯=v𝟏\mathbf{v}=v\mathbf{1} implies that αi\alpha_{i}. The result follows by noting that ni=no1n_{i}=n_{o}-1.

The implication of this result is that the dynamics (11) embeds a bifurcation which consists of multiple copies of the standard pitchfork bifurcation (1). The structure of the equilibria of (11) in the post-bifurcation regime reflects the symmetry properties of the vector field and can be studied in detail by appeal to the equivariant branching lemma [20, Theorem 3.3]. The detailed analysis is beyond the scope of this paper, but we show numerical results in Section 7.1 below.

6 Model reduction via singular perturbation

The dynamics fmf_{m} defined in (11) inherit a complicated rational form from the Seeley et al. dynamics (3). As shown in Theorem 1.1, the dynamics (3) can be reduced by singular perturbation. In this section, we carry out an analogous model reduction for (11) and show that the equilibria of the resulting reduced model can be readily understood.

6.1 Change of coordinates

As in [18], we apply singular perturbation theory to the dynamics (11) by mapping 𝐯K𝐯\mathbf{v}\mapsto K\mathbf{v} for a constant gain K>0K>0 and take the singular limit K+K\to+\infty, or equivalently ϵ=1/K0\epsilon=1/K\to 0. The singular perturbation allows us to eliminate half of the state variables and thus to express equilibria in a straightforward way as a function of the option values viv_{i}. The singular perturbation is more readily analyzed by expressing 𝐦i=(mi1,mi2)Δ2\mathbf{m}_{i}=(m_{i1},m_{i2})\in\Delta^{2} and 𝐯i=(vi1,vi2)+2\mathbf{v}_{i}=(v_{i1},v_{i2})\in\mathbb{R}^{2}_{+} in terms of mean-difference coordinates defined by

Δmi=mi1mi2,m¯i=mi1+mi22,and Δvi=vi1vi2,v¯i=vi1+vi22,\Delta m_{i}=m_{i1}-m_{i2},\ \bar{m}_{i}=\frac{m_{i1}+m_{i2}}{2},\text{and }\Delta v_{i}=v_{i1}-v_{i2},\ \bar{v}_{i}=\frac{v_{i1}+v_{i2}}{2},

respectively. Expressing 𝐦i\mathbf{m}_{i} in mean-difference coordinates results in expressing 𝐳i2\mathbf{z}_{i}\in\mathbb{R}^{2} in corresponding mean-difference coordinates

(zi1,zi2)=𝐳i=zi𝐦i=(zimi1,zimi2)T=(zi2m¯i+Δmi2,zi2m¯iΔmi2)T.(z_{i1},z_{i2})=\mathbf{z}_{i}=z_{i}\mathbf{m}_{i}=(z_{i}m_{i1},z_{i}m_{i2})^{T}=\left(z_{i}\frac{2\bar{m}_{i}+\Delta m_{i}}{2},z_{i}\frac{2\bar{m}_{i}-\Delta m_{i}}{2}\right)^{T}.

Note that the recursive definition of the ziz_{i} coordinates as 𝐳i=zi𝐦i\mathbf{z}_{i}=z_{i}\mathbf{m}_{i} is such that the value of ziz_{i} can be expressed as the product of 𝐦i\mathbf{m}_{i} along the path pip_{i} from the root to node ii. Recall that we define pip_{i} as the sequence of nodes traversed along the unique shortest path from the root to node ii. The sequence pip_{i} begins with the root node and ends with node ii. We denote the number of nodes in the sequence by |pi||p_{i}|, and the jthj^{th} node of pip_{i} by pijp_{ij}. Explicitly, we have

(20) zi=j=1|pi|12m¯pij+ajΔmpij2, where aj={+1,pi(j+1)=lc(pij)1,pi(j+1)=rc(pij).z_{i}=\prod_{j=1}^{|p_{i}|-1}\frac{2\bar{m}_{p_{ij}}+a_{j}\Delta m_{p_{ij}}}{2},\text{ where }a_{j}=\begin{cases}+1,&p_{i(j+1)}=\operatorname{lc}(p_{ij})\\ -1,&p_{i(j+1)}=\operatorname{rc}(p_{ij}).\end{cases}

The dynamics of ziz_{i} follow from the dynamics (11) and can be reduced by applying singular perturbation theory. As in [14, 18], we map 𝐯\mathbf{v} to K𝐯K\mathbf{v}, where K>0K>0 is a constant gain. To apply singular perturbation theory, we set ϵ=1/K\epsilon=1/K to be a small parameter and define coordinates x,yx,y with components

xi=Δmi,yi=12m¯iϵ.x_{i}=\Delta m_{i},\ y_{i}=\frac{1-2\bar{m}_{i}}{\epsilon}.

Since the dynamics fmf_{m} defined in (11) is composed of stacked copies of the dynamics ff defined in (3), singular perturbation of (11) can be carried out by singularly perturbing its components which consist of copies of ff. We can express the dynamics in the coordinates (Δmi,m¯i)(\Delta m_{i},\bar{m}_{i}) using the dynamics (5), (6)

(21) ˙Δmi\displaystyle\dot{}\Delta m_{i} =fΔm(Δmi,m¯i;Kv¯i;KΔvi)\displaystyle=f_{\Delta m}(\Delta m_{i},\bar{m}_{i};K\bar{v}_{i};K\Delta v_{i})
(22) ˙m¯i\displaystyle\dot{}\bar{m}_{i} =fm¯(Δmi,m¯i,σ;KΔvi).\displaystyle=f_{\bar{m}}(\Delta m_{i},\bar{m}_{i},\sigma;K\Delta v_{i}).

In the singular perturbation coordinates (xi,yi)(x_{i},y_{i}), these dynamics become

(23) x˙i\displaystyle\dot{x}_{i} =fx(xi,yi;Δvi,v¯i,ϵ)\displaystyle=f_{x}(x_{i},y_{i};\Delta v_{i},\bar{v}_{i},\epsilon)
=ϵ(1ϵyi+xi2v¯i+Δvi1ϵyixi2v¯iΔvi)+v¯ixiyi+Δviyi3ϵyi2\displaystyle=-\epsilon\left(\frac{1-\epsilon y_{i}+x_{i}}{2\bar{v}_{i}+\Delta v_{i}}-\frac{1-\epsilon y_{i}-x_{i}}{2\bar{v}_{i}-\Delta v_{i}}\right)+\bar{v}_{i}x_{i}y_{i}+\Delta v_{i}y_{i}\frac{3-\epsilon y_{i}}{2}
(24) ϵy˙i\displaystyle\epsilon\dot{y}_{i} =gy(xi,yi;Δvi,v¯i,ϵ)\displaystyle=g_{y}(x_{i},y_{i};\Delta v_{i},\bar{v}_{i},\epsilon)
=ϵ(1ϵyi+xi2v¯i+Δvi1ϵyixi2v¯iΔvi)+σ2((1ϵyi)2xi2)\displaystyle=\epsilon\left(\frac{1-\epsilon y_{i}+x_{i}}{2\bar{v}_{i}+\Delta v_{i}}-\frac{1-\epsilon y_{i}-x_{i}}{2\bar{v}_{i}-\Delta v_{i}}\right)+\frac{\sigma}{2}((1-\epsilon y_{i})^{2}-x_{i}^{2})
yi2((2v¯i+Δvi)(1+1ϵyi+xi2))yi2((2v¯iΔvi)(1+1ϵyixi2)).\displaystyle\ \ \ \ -\frac{y_{i}}{2}\left((2\bar{v}_{i}+\Delta v_{i})\left(1+\frac{1-\epsilon y_{i}+x_{i}}{2}\right)\right)-\frac{y_{i}}{2}\left((2\bar{v}_{i}-\Delta v_{i})\left(1+\frac{1-\epsilon y_{i}-x_{i}}{2}\right)\right).

6.2 Reduced node dynamics

Taking the singular limit of the dynamics (23), (24) associated with node ii yields a reduced system whose dynamics are given by a rational polynomial. This is formalized in the following theorem, which is a straightforward application of [18, Theorem 1] stated above as Theorem 1.1 and whose proof is reproduced here.

Theorem 6.1.

In the singular limit ϵ0\epsilon\to 0, the dynamics (23), (24) reduce to

(25) x˙i=σ2v¯i(1xi2)2xi+3αi6+αixi,\dot{x}_{i}=\frac{\sigma}{2\bar{v}_{i}}(1-x_{i}^{2})\frac{2x_{i}+3\alpha_{i}}{6+\alpha_{i}x_{i}},

where αi=Δvi/v¯i\alpha_{i}=\Delta v_{i}/\bar{v}_{i}.

Proof 6.2.

The proof follows the standard procedure for analyzing singularly-perturbed systems. First note that xix_{i} is the slow and yiy_{i} the fast variable. Taking the singular limit ϵ0\epsilon\to 0 of (23) and (24) yields

(26) x˙i\displaystyle\dot{x}_{i} =fx(xi,yi;Δvi,v¯i,0)=v¯ixiyi+3Δviyi2\displaystyle=f_{x}(x_{i},y_{i};\Delta v_{i},\bar{v}_{i},0)=\bar{v}_{i}x_{i}y_{i}+\frac{3\Delta v_{i}y_{i}}{2}
(27) 0\displaystyle 0 =gy(xi,yi;Δvi,v¯i,0)=yi2(6v¯i+Δvixi)+σ2(1xi2).\displaystyle=g_{y}(x_{i},y_{i};\Delta v_{i},\bar{v}_{i},0)=-\frac{y_{i}}{2}\left(6\bar{v}_{i}+\Delta v_{i}x_{i}\right)+\frac{\sigma}{2}\left(1-x_{i}^{2}\right).

Solving Equation (27) for the fast variable yiy_{i} yields

yi=h(xi):=σ(1xi2)6v¯i+Δvixi=σv¯i1xi26+(Δvi/v¯i)xi,y_{i}=h(x_{i}):=\frac{\sigma(1-x_{i}^{2})}{6\bar{v}_{i}+\Delta v_{i}x_{i}}=\frac{\sigma}{\bar{v}_{i}}\frac{1-x_{i}^{2}}{6+(\Delta v_{i}/\bar{v}_{i})x_{i}},

which defines the slow manifold {(xi,yi)=(xi,h(xi))}\{(x_{i},y_{i})=(x_{i},h(x_{i}))\}. The system quickly converges to the slow manifold and then xix_{i} slowly evolves on the slow manifold. Using the expression yi=h(xi)y_{i}=h(x_{i}) for the fast variable yiy_{i} in terms of the slow variable xix_{i} yields the reduced slow dynamics

x˙i=fx(xi,h(xi);Δvi,v¯i,0)=σ2v¯i(1xi2)2xi+3Δvi/v¯i6+(Δvi/v¯i)xi.\dot{x}_{i}=f_{x}(x_{i},h(x_{i});\Delta v_{i},\bar{v}_{i},0)=\frac{\sigma}{2\bar{v}_{i}}(1-x_{i}^{2})\frac{2x_{i}+3\Delta v_{i}/\bar{v}_{i}}{6+(\Delta v_{i}/\bar{v}_{i})x_{i}}.

Defining αi=Δvi/v¯i\alpha_{i}=\Delta v_{i}/\bar{v}_{i} yields the desired result (25).

The implication of this result is that, in the singular limit ϵ0\epsilon\to 0, m¯i1/2\bar{m}_{i}\to 1/2 and Δmi=xi\Delta m_{i}=x_{i} follows the dynamics (25). The coordinates of 𝐦i\mathbf{m}_{i} associated with a node ii then reduce to

𝐦i=(2m¯i+Δmi2,2m¯iΔmi2)T=(1+Δmi2,1Δmi2)T,\mathbf{m}_{i}=\left(\frac{2\bar{m}_{i}+\Delta m_{i}}{2},\frac{2\bar{m}_{i}-\Delta m_{i}}{2}\right)^{T}=\left(\frac{1+\Delta m_{i}}{2},\frac{1-\Delta m_{i}}{2}\right)^{T},

where Δmi\Delta m_{i} evolves according to (25). As shown in in Figure 3, the dynamics (25) have equilibria xi=±1,3αi/2x_{i}=\pm 1,-3\alpha_{i}/2 whose existence and stability properties depend on the unfolding parameter αi\alpha_{i}.

6.3 Reduced tree dynamics

The reduced dynamics of the full tree then follow from the reduction at each internal node ii introduced in Theorem 6.1. As noted above, the dynamics (11) of the tree state 𝐦\mathbf{m} consists of stacked copies of the node dynamics ˙𝐦i=fm(𝐦i,𝐯i)\dot{}\mathbf{m}_{i}=f_{m}(\mathbf{m}_{i},\mathbf{v}_{i}). The reduced dynamics consist of stacked copies of the reduced dynamics (25). The equilibria of each state follow from the component dynamics. Formally, we have the following.

Corollary 6.3.

In the singular limit ϵ0\epsilon\to 0, the dynamics (11) reduce to

(28) ˙𝐱=f𝐱(𝐱,𝐯),\dot{}\mathbf{x}=f_{\mathbf{x}}(\mathbf{x},\mathbf{v}),

where the ithi^{th} component of 𝐱\mathbf{x} is equal to xix_{i}, the states xix_{i} associated with each node ii are stacked in depth-first order, and f𝐱f_{\mathbf{x}} consists of stacked copies of the singularly-reduced dynamics (25).

The dynamics (28) have equilibrium states 𝐱0\mathbf{x}_{0} whose ithi^{th} component xi,0x_{i,0} is equal to ±1\pm 1 or 3αi/2-3\alpha_{i}/2. The existence and stability properties of these equilibrium values depends on the unfolding parameter αi\alpha_{i}, as follows:

xi,0={+1,αi[2,2], stable if αi>2/31,αi[2,2], stable if αi<2/32αi/3,αi[2/3,2/3], unstable.x_{i,0}=\begin{cases}+1,&\forall\alpha_{i}\in[-2,2],\text{ stable if }\alpha_{i}>-2/3\\ -1,&\forall\alpha_{i}\in[-2,2],\text{ stable if }\alpha_{i}<2/3\\ -2\alpha_{i}/3,&\forall\alpha_{i}\in[-2/3,2/3],\text{ unstable}.\end{cases}

The equilibria of the reduced dynamics (28) imply a set of equilibria of the projected dynamics fof_{o} defined in (14) whose values can be cleanly expressed in terms of the recursive definition (20). Formally, we have the following.

Corollary 6.4.

Take the singular limit ϵ0\epsilon\to 0 of the dynamics (11) and consider the projection of these dynamics to the leaf states ˙𝐦o=fo(𝐳,𝐯)\dot{}\mathbf{m}_{o}=f_{o}(\mathbf{z},\mathbf{v}) given by (14). Denote the ithi^{th} component of 𝐦o\mathbf{m}_{o} as moi=zo1(i)m_{oi}=z_{o^{-1}(i)}. The projected dynamics have equilibria with components

(29) moi=zo1(i)=j=1|pi|11+ajΔmj2,m_{oi}=z_{o^{-1}(i)}=\prod_{j=1}^{|p_{i}|-1}\frac{1+a_{j}\Delta m_{j}^{*}}{2},

where aja_{j} and Δmj\Delta m_{j}^{*} are given by

aj\displaystyle a_{j} ={+1,po1(i)(j+1)=lc(po1(i)j)1,po1(i)(j+1)=rc(po1(i)j) and\displaystyle=\begin{cases}+1,&p_{o^{-1}(i)(j+1)}=\operatorname{lc}(p_{o^{-1}(i)j})\\ -1,&p_{o^{-1}(i)(j+1)}=\operatorname{rc}(p_{o^{-1}(i)j})\end{cases}\text{ and }
Δmj\displaystyle\Delta m_{j}^{*} =Δmpo1(i)j={+1,αpo1(i)j[2,2],1,αpo1(i)j[2,2],2αpo1(i)j/3,αpo1(i)j[2/3,2/3].\displaystyle=\Delta m_{p_{o^{-1}(i)j}}=\begin{cases}+1,&\forall\alpha_{p_{o^{-1}(i)j}}\in[-2,2],\\ -1,&\forall\alpha_{p_{o^{-1}(i)j}}\in[-2,2],\\ -2\alpha_{p_{o^{-1}(i)j}}/3,&\forall\alpha_{p_{o^{-1}(i)j}}\in[-2/3,2/3].\\ \end{cases}

Equilibria are stable if each Δmj\Delta m_{j}^{*} is a stable equilibrium, where the stability of each Δmj\Delta m_{j}^{*} is given in Corollary 6.3.

Proof 6.5.

Recall that the value of moi=zo1(i)m_{oi}=z_{o^{-1}(i)} associated with an option ii can be expressed using the recursive definition (20) in terms of m¯j\bar{m}_{j} and Δmj\Delta m_{j} associated with nodes on the shortest path from the root node to the leaf node that represents the option ii. In the singular limit ϵ0\epsilon\to 0, we have m¯j1/2j\bar{m}_{j}\to 1/2\forall j and ˙Δmj\dot{}\Delta m_{j} following the dynamics (28) with equilibria given in Corollary 6.3. Substituting the values of m¯j\bar{m}_{j} and Δmj\Delta m_{j} yields the expression (29).

The stability property follows by contradiction. If any Δmj\Delta m_{j}^{*} in the product (29) corresponds to an unstable equilibrium of the singularly-perturbed dynamics (28), then the overall equilibrium will be unstable.

The implication of this result is clear from noting that the intermediate equilibrium Δmj=2αpo1(i)j/3\Delta m_{j}^{*}=-2\alpha_{p_{o^{-1}(i)j}}/3 is always unstable when it exists. Furthermore, the negative signs from aja_{j} and Δmj\Delta m_{j}^{*} cancel out when the path pip_{i} passes from a parent to a right child. Thus, when the option value viv_{i} is sufficiently high so that αj>2/3\alpha_{j}>-2/3 whenever pip_{i} passes from a parent to a left child and that αj<2/3\alpha_{j}<2/3 whenever pip_{i} passes from a parent to a right child. In this case, each element in the product (29) is equal to one. Thus, when viv_{i} is sufficiently large relative to the other option values, the unique stable equilibrium of the projected dynamics (14) is the state 𝐦o=ei\mathbf{m}_{o}=e_{i}, where eie_{i} is the indicator vector with entry ii equal to 1 and all other entries equal to zero. Therefore, in the singular limit and when one option value is sufficiently large relative to the others, the dynamics (14) carries out an argmax\arg\max operation on the value vector 𝐯\mathbf{v}. When several option values are relatively large, the dynamics (14) effectively performs a sort of dynamical argmax\arg\max operation whose output depends on initial conditions. See Section 7.2 for a numerical example.

7 Numerical examples

In this section we show the results of numerical simulations of the dynamics. All the computations have been carried out with code that is publicly available from the author’s website [24]. The code is completely general in the sense that it implements the dynamics (11) for a generic binary tree TT. For clarity of presentation, all the simulations in this section are based on a binary tree parsing four options, as shown in Figure 8. In all simulations, the parameter σ\sigma in (3) is set equal to 4 wherever it appears (once in (11) for each internal node).

0123456

Ordered node list: (0,1,2,3,4,5,6)(0,1,2,3,4,5,6)
Ordered option list: (2,3,5,6)

Figure 8: Tree TT used in the simulations. TT is a parsing of four options, corresponding to the nodes (2,3,5,6)(2,3,5,6). The nodes are labeled with numbers according to the order in which they will be visited during a depth-first traversal.

7.1 Bifurcation characteristics

Here we present the results of several simulations illustrating the bifurcation characteristics of the system as studied in Section 5. Figures 9 and 10 study the case of equal option values (i.e., vi=vv_{i}=v for each option i{1,,4}i\in\{1,\ldots,4\}) and show how the system (11) bifurcates from having a single stable equilibrium to having five equilibria as the option value vv is increased past the critical value v1.9058v^{*}\approx 1.9058 defined by (17).

For the simulation shown in Figure 9, v=1.25<1.9058vv=1.25<1.9058\approx v^{*}, so the system is in the pre-bifurcation regime. As predicted by Theorem 5.4, the dynamics (11) have an equilibrium 𝐦=m¯𝟏\mathbf{m}=\bar{m}\mathbf{1}, where m¯\bar{m} is defined by (18). This equilibrium value of 𝐦\mathbf{m} gets projected to an equilibrium value 𝐦o\mathbf{m}_{o} of (14) whose ithi^{th} component, corresponding to the ithi^{th} option, is equal to m¯d\bar{m}^{d}, with d=|po1(i)|d=|p_{o^{-1}(i)}| being the distance from the root of the tree TT to the ithi^{th} option.

The simulation shown in Figure 10 is identical to that shown in Figure 9, except that now we set vi=v=5v_{i}=v=5 so that the system is in the post-bifurcation regime. The four panels show trajectories resulting from four different initial conditions along with the (now unstable) deadlock equilibrium located at 𝐦o=m¯𝟏\mathbf{m}_{o}=\bar{m}\mathbf{1}. As suggested by Corollary 11, in the post-bifurcation regime there is an additional set of four symmetric stable equilibria corresponding to a clear preference for each of the four equally-valued options. The equilibrium to which a trajectory is attracted depends on initial conditions. In panel (a), initial conditions were 𝐦=(𝐦0T,𝐦1T,𝐦4T)T=(0.2,0.1,0.3,0.2,0.4,0.2)T\mathbf{m}=(\mathbf{m}_{0}^{T},\mathbf{m}_{1}^{T},\mathbf{m}_{4}^{T})^{T}=(0.2,0.1,0.3,0.2,0.4,0.2)^{T}, corresponding to a weak initial preference for option 1. This initial preference determines the attracting equilibrium. The other three panels (b)–(d) use initial conditions that are permutations of those from panel (a). These permutations correspond to tree isomorphisms that exchange option 1 with options 2–4, respectively. As expected, the attracting equilibrium changes from option 1 to option 2–4 accordingly.

Refer to caption
Figure 9: Setting vi=v=1.25v_{i}=v=1.25 for each option ii puts the system (11) in the pre-bifurcation regime. This results in a stable deadlock equilibrium where no option wins. The traces show the trajectories of the components of the projected state 𝐦o\mathbf{m}_{o}, each of which converges to the deadlock value represented by the dashed line.
Refer to caption Refer to caption
(a) (b)
Refer to caption Refer to caption
(c) (d)
Figure 10: Setting vi=v=5v_{i}=v=5 for each option ii puts the system (11) in the post-bifurcation regime. The traces show the trajectories of the components of the projected state 𝐦o\mathbf{m}_{o}. As suggested by Corollary 6.4, there is a set of four symmetric stable equilibria corresponding to a clear preference for each of the four equally-valued options, along with an unstable deadlock equilibrium. The simulations producing the four plots are identical except for a permutation of the initial conditions that cause the system to converge to different stable states. Setting a symmetric initial condition 𝐳0=z𝟏\mathbf{z}_{0}=z\mathbf{1} where all initial state values are equal results in deadlock.

7.2 Singularly-perturbed system

In Figure 11 we present the results of a simulation illustrating the results of Section 6, particularly Corollary 6.4. The values of the four options are set equal to (100,100,300,100)T=100(1,1,3,1)T(100,100,300,100)^{T}=100(1,1,3,1)^{T}, which puts the system close to the singularly-perturbed regime with ϵ=1/1001\epsilon=1/100\ll 1. These option values are such that the unfolding parameters of the internal nodes are α0=2/3,α1=0,\alpha_{0}=-2/3,\alpha_{1}=0, and α4=1\alpha_{4}=1. In this case, Corollary 6.4 predicts that there should be a unique stable equilibrium at 𝐦o=(0,0,1,0)T\mathbf{m}_{o}=(0,0,1,0)^{T} corresponding to an absolute preference for the high-value option 3.

The results shown in Figure 11 confirm this prediction of a unique stable equilibrium, as the trajectories of the projected dynamics (14) all converge to 𝐦0=(0,0,1,0)T\mathbf{m}_{0}=(0,0,1,0)^{T} for four different simulations with initial conditions corresponding to the simulations shown in Figure 10(a)–(d).

Refer to caption
Figure 11: Setting the option values 𝐯=(100,100,300,100)T=100(1,1,3,1)T\mathbf{v}=(100,100,300,100)^{T}=100(1,1,3,1)^{T} puts the system (11) close to the singularly-perturbed regime. Note that these option values are such that α0=2/3\alpha_{0}=-2/3 and α4=1\alpha_{4}=1, so the only stable equilibrium should be the one corresponding to an absolute preference for the high-value option 3. As predicted by Corollary 6.4, there is a single stable equilibrium corresponding to a preference for option 3. The traces show the trajectories of the components of the projected state 𝐦o\mathbf{m}_{o} for the four different initial conditions considered in Figure 10(a)–(d). In all cases, 𝐦o\mathbf{m}_{o} converges to the state (0,0,1,0)T(0,0,1,0)^{T}.

8 Conclusion

In this paper, we have developed a dynamical systems model of value-based decision making. The structure of the decision, and of the dynamical system itself, is encoded in a binary tree structure. The binary tree structure allows us to decompose a decision among NN options into a set of N1N-1 binary decisions arranged in a hierarchical structure. We then represent this decomposed decision as a dynamical system (11) whose vector field is defined recursively by parsing down the binary tree. At each internal node of the tree, the system makes decisions based on the values associated with the node’s two children, putting higher weight on the higher-value child. The NN leaf nodes of the tree represent the NN options.

The vector field (11) has symmetries that correspond to isomorphisms of the underlying tree and associated option values. When the NN options all have the same value, all isomorphisms of the tree leave the vector field equivariant; when only some options have the same value, a smaller set of isomorphisms leave the vector field equivariant. The equilibria of the vector field have significant structure, organized around an N1N-1-parameter unfolding of a pitchfork singularity as shown in Theorem 5.4. The unfolding parameters of the pitchfork consist of the relative difference in values between the children of each internal node of the underlying tree. As shown in Corollary 6.4, the system equilibria correspond to point attractors at states that correspond to a preference for the high-value option. In a singular limit, this preference becomes absolute in the sense that no weight is accorded to any other option.

Further work remains to be done to understand the structure of the symmetry group of the vector fields (11) and (14), particularly in the case that only a subset of the options have identical values. In this case, the symmetry group will be a subgroup of the original group ΓT\Gamma_{T}, and such subgroups likely have interesting structure. Similarly, further work remains to be done to understand the structure of equilibria of the vector fields in the post-bifurcation regime. The main tool here is the equivariant branching lemma [20, Theorem 3.3], which again leverages the subgroup structure of the symmetry group ΓT\Gamma_{T}.

There are a number of interesting questions raised by the binary tree structure of our model. For example, consider a generic case of deciding among N>2N>2 options. The binary tree structure appears to be a strong constraint on the structure of the decision-making process. A more general value-based decision-making model, such as the one whose analysis was begun in [6], could have similar unfolding characteristics with fewer structural constraints. An open question is to understand the effect of the constraints imposed by the binary tree structure. Are there decisions that can be made by a model encoded as a flat graph (i.e., without a hierarchy structure) that cannot be made by our binary-tree-based model?

As discussed in the introduction, we anticipate the model developed in this paper to be valuable for a variety of problems requiring models of value-based decision-making behavior. We are actively pursuing applications in the area of control systems and robotics where options correspond to control vector fields and the present model affords a method to compose multiple such vector fields. In particular, we are developing methods to derive dynamics of the option values viv_{i} such that the overall system achieves a complex behavior specified, e.g., in terms of temporal logic. This work has the potential to unite dynamical systems with so-called formal methods tools [25] for control.

Acknowledgement

We thank Daniel Koditschek for discussions that led to the concept of a parsing. This work was supported in part by Air Force Research Laboratory grant FA8650-15-D-1845 subcontract 669737-6 and grant FA8650-19-C-1712 subcontract 670956-1.

References

  • [1] T. D. Seeley, P. K. Visscher, T. Schlegel, P. M. Hogan, N. R. Franks, and J. A. Marshall, “Stop signals provide cross inhibition in collective decision-making by honeybee swarms,” Science, vol. 335, no. 6064, pp. 108–111, 2012.
  • [2] J. A. Marshall, R. Bogacz, A. Dornhaus, R. Planqué, T. Kovacs, and N. R. Franks, “On optimal decision-making in brains and social insect colonies,” Journal of the Royal Society Interface, vol. 6, no. 40, pp. 1065–1074, 2009.
  • [3] D. Pais, P. M. Hogan, T. Schlegel, N. R. Franks, N. E. Leonard, and J. A. Marshall, “A mechanism for value-sensitive decision-making,” PloS one, vol. 8, no. 9, p. e73216, 2013.
  • [4] A. Reina, J. A. R. Marshall, V. Trianni, and T. Bose, “Model of the best-of-nn nest-site selection process in honeybees,” Physical Review E, vol. 95, no. 5, 2017.
  • [5] R. Gray, A. Franci, V. Srivastava, and N. E. Leonard, “Multi-agent decision-making dynamics inspired by honeybees,” IEEE Transactions on Control of Network Systems, vol. 5, no. 2, pp. 793–806, 2018.
  • [6] A. Franci, M. Golubitsky, and N. E. Leonard, “The dynamics of multi-agent multi-option decision making,” 2019.
  • [7] J. J. Gibson, The theory of affordances. The ecological approach to visual perception.   Houghton Mifflin Boston, MA, 1979.
  • [8] K. Friston, P. Schwartenbeck, T. FitzGerald, M. Moutoussis, T. Behrens, and R. J. Dolan, “The anatomy of choice: active inference and agency,” Frontiers in human neuroscience, vol. 7, p. 598, 2013.
  • [9] N. F. Lepora and G. Pezzulo, “Embodied choice: how action influences perceptual decision making,” PLoS computational biology, vol. 11, no. 4, p. e1004110, 2015.
  • [10] P. Cisek and J. F. Kalaska, “Neural mechanisms for interacting with a world full of action choices,” Annual review of neuroscience, vol. 33, pp. 269–298, 2010.
  • [11] P. Cisek, “Cortical mechanisms of action selection: the affordance competition hypothesis,” Philosophical Transactions of the Royal Society B: Biological Sciences, vol. 362, no. 1485, pp. 1585–1599, 2007.
  • [12] A. P. Duchon, L. P. Kaelbling, and W. H. Warren, “Ecological robotics,” Adaptive Behavior, vol. 6, no. 3-4, pp. 473–507, 1998.
  • [13] P. Zech, S. Haller, S. R. Lakani, B. Ridge, E. Ugur, and J. Piater, “Computational models of affordance in robotics: a taxonomy and systematic classification,” Adaptive Behavior, vol. 25, no. 5, pp. 235–271, 2017.
  • [14] P. B. Reverdy and D. E. Koditschek, “A dynamical system for prioritizing and coordinating motivations,” SIAM Journal on Applied Dynamical Systems, vol. 17, no. 2, pp. 1683–1715, 2018.
  • [15] P. Reverdy, V. Vasilopoulos, and D. E. Koditschek, “Motivation dynamics for autonomous composition of navigation tasks,” Submitted, 2020.
  • [16] M. Golubitsky and D. Schaeffer, Singularities and Groups in Bifurcation Theory, ser. Applied Mathematical Sciences.   Springer, 1985, vol. 51.
  • [17] M. Golubitsky and I. Stewart, The symmetry perspective: from equilibrium to chaos in phase space and physical space.   Springer Science & Business Media, 2003, vol. 200.
  • [18] P. B. Reverdy, “Two paths to finding the pitchfork bifurcation in motivation dynamics,” in Proc. IEEE Conf. Decision and Control, 2019, pp. 8030–8035.
  • [19] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of agents with switching topology and time-delays,” IEEE Transactions on automatic control, vol. 49, no. 9, pp. 1520–1533, 2004.
  • [20] M. Golubitsky, I. Stewart, and D. G. Schaeffer, Singularities and groups in bifurcation theory: Vol. II, ser. Applied Mathematical Sciences.   New York: Springer-Verlag, 1988, no. 69.
  • [21] W. Dicks and M. J. Dunwoody, Groups acting on graphs.   Cambridge University Press, 1989, vol. 17.
  • [22] A. V. Aho, J. E. Hopcroft, and J. D. Ullman, The design and analysis of computer algorithms.   Addison-Wesley, 1974.
  • [23] M. Hazewinkel. Binary tree. Encyclopedia of Mathematics. [Online]. Available: {http://www.encyclopediaofmath.org/index.php?title=Binary_tree&oldid=31607}
  • [24] P. B. Reverdy. (2020, Mar.) preverdy/binary-tree-decisions: First release. [Online]. Available: https://doi.org/10.5281/zenodo.3698325
  • [25] H. Kress-Gazit, M. Lahijanian, and V. Raman, “Synthesis for Robots: Guarantees and Feedback for Robot Behavior,” Annual Review of Control, Robotics, and Autonomous Systems, vol. 1, no. 1, pp. 211–236, 2018. [Online]. Available: https://doi.org/10.1146/annurev-control-060117-104838