A discrete districting plan

The outcome of elections is strongly dependent on the districting choices, making thus possible (and frequent) the gerrymandering phenomenon, i.e. politicians suitably changing the shape of electoral districts in order to win the forthcoming elections. While so far the problem has been treated using continuous analysis tools, it has been recently pointed out that a more reality-adherent model would use the discrete geometry of graphs or network. Here we propose a parameter-dependent discrete model for choosing an"optimal"districting plan. We analyze several properties of the model and lay foundations for further analysis on the subject.


Introduction
The most ancient reference to the isoperimetric problem is known by the name of Dido's problem which has a political background. Legend has it that queen Dido of Carthage was given the chance to found a city on the area she would have been able to enclose with a given ox hide: she cut the hide in thin strips and proceeded to enclose a very large area. If one were to seize this opportunity and get the most out of it, what would be the best shape? In such a context, by best we mean with the greatest area given a length. Nowadays this problem is more commonly known through its dual formulation, i.e. to find the least perimeter enclosing a given area. It has been hypothesized for millennia that the best is given by the circle, yet a formal proof was not available until very recently. A first step toward the proof was made by Steiner in the 19th century who showed that if a solution existed, it had to be the circle [14]. It was in the 20th century that De Giorgi bridged the last gaps proving the full result via the theory of sets of finite perimeter and BV functions he developed in [7].
Whenever one has to minimize the (some notion of) perimeter with a constraint on the (some notion of) area, one refers to the problem as an isoperimetric-like. These problems have a wide number of practical applications ranging from physics such as the capillarity phenomenon [10]) to cristallography such as the shapes of equilibrium crystals [5]. In the latest years isoperimetric-like problems were looked at with interest from social sciences with the explicit aim to recognize gerrymandering phenomena in politics [8].
Most of the literature on gerrymandering discusses the current shapes of electoral districts and assignes to each different scores, all of isoperimetric nature whose goal is to measure the "compactness" of the shapes, whatever this may mean. The aim is twofold: compare districts via these scores and rank them from the most gerrymandered to the least; decide whether a district is gerrymandered or not.
In this paper we propose a discrete districting plan and discuss some properties we would like our model to possess. The underlying idea is that given some region Ω we want to partition it in N subregions {Ω k } N k=1 with the same population, i.e. (1) where f denotes the population density over Ω. A general approach to get the "most compact" shape is to minimize the (relative) perimeter of the partition, thus we would be led to minimize the functional 1 2 N k=1 P (Ω k ; Ω) , subject to (1), which is linked to the score proposed in [6]. Though, one might argue that it would be best to weigh the perimeter as well via the density population in order to make uneconomical to split highly populated regions. Indeed a high density might be an indicator of strong political, ethnical, religious, linguistic and so on ties, thus it would be fair to not split them. Then, the functional to be minimized would be where ∂ * Ω k denotes the reduced boundary. More general functionals to be minimized can be proposed such as where the weight g = g(x, ν) takes into account not only the point but as well the direction of the boundary at the point. This weight g might represent how big the flow of people and the exchange of informations at point x in direction ν is. Isoperimetric-like problems of this fashion have been treated in [12,13], as well as the regularity of minimizers in [11]. Continuous descriptions of the gerrymandering phenomenon suffer though from many issues as highlighted in [8] where the authors propose to make use of tools from discrete geometry and the theory of graphs to depict the scenario. Indeed, the situation may be very well described in the setting of discrete geometry as the total population is finite. The idea is to choose one of the levels of the census' units (for instance US: block, block's group, tract; Italy: comune, provincia, regione) and assign to each a vertex v i in a graph Γ. We shall then say that two vertexes, v i , v j , are adjacent if the corresponding units share (a positive amount of) boundary and denote their edge by e i,j . This latter hypothesis ensures that the resulting graph is planar. Then, the weight f denotes the population of a vertex, while the weight g is a measure of how well two adjacent vertexes are connected. Hence, the discrete functional equivalent to (3) is where {Γ k } is a N -partition of Γ and the sum is taken over the edges e i,j belonging to the cut set C({Γ k }), i.e. sloppily speaking the edges removed from Γ to obtain the partition {Γ k } (see Definition 2.2 for the formal details of partition and cut set). Clearly one can not impose a constraint analogous to (1) as in general this would prevent a solution to exist. Adding a penalization term of the form rather than a mass constraint, ensures the existence of solutions without completely dropping the request to have (almost) equally populated regions. This term represents the standard deviation of the populations in each subgraph with respect to the mean population of the whole graph, and one could use a general p-norm in place of the 2-norm. The functional we will look at is the convex combination of (4) and (5), i.e.
Depending on the choice of λ one gives more or less prominence to one of the two terms. When one looks purely at the perimeter energy (4), i.e. for λ = 1, the problem is closely related with the one known as minimum-Ncut, where one seeks to split the graph in at least N components rather than exactly N . The proposed functional clearly does not only have applications to politics. One can imagine many different scenarios: for instance the vertexes of the graph might represent computers with the weight f their powers, while edges represent direct connections and the weight g how many MB/s of dataflow these links grant. Then our problem would represent the need to break this computer network in smaller groups of similar power by cutting the slowest connections. In this latter case though it is very possible that the starting graph is not planar, possibly adding more algorithmic complexity. For sake of completeness, we recall that for fixed N the minimum-N -cut problem is polynomial time solvable (see [4,9] and the references therein).
The paper is organized as follows. In Section 2 we lay the notation and precisely define the functional. In Section 3 we discuss a series of properties one would like to have for such a problem. For each desired property we provide either a proof of the property or show a counterexample disproving it. In Section 4 we discuss the open problems we have not settled yet, which would be nice to explore, and define some further research directions we shall investigate in the future.

The model
Let us consider a graph The elements {v i } i = V (Γ) represent the vertexes of the graph, while those of E(Γ) the edges, i.e. the connections e k i,j from vertex v i to vertex v j available in the graph. We shall suppose that • Γ is finite, i.e. |V (Γ)| < +∞; • Γ is simple, i.e. there is at most one edge e i,j connecting the vertex v i to v j and there are no loops e i,i ; • Γ is undirected, i.e. e i,j is identified with e j,i ; • Γ is connected, i.e. given any two vertexes v, w there exists a sequence of vertexes {v k } n k=1 such that v 1 = v, v n = w and e k,k+1 ∈ E(Γ) for all k = 1, . . . , n − 1. We endow the graph Γ with two different weights, one acting on the vertexes and one on the edges. Specifically, we set f : V (Γ) → (0, +∞) and g : E(Γ) → (0, +∞). Graph with weights are occasionally referred to as networks. More details and basic notions of graphs can be found in the monographs [3,15]. Given these choices, one could represent the weighted edges as the symmetric traceless square matrix A ∈ M(|V (Γ)|), where a i,j = g(e i,j ) if e i,j ∈ E(Γ) and a i,j = 0 otherwise.
The graph amounts to the region which we want to divide into electoral districts; the vertex v i represents a town or neighbourhood in the region (f (v i ) being the population of v i ) and the edge e i,j represents a direct connection between two nearby towns/neighbourhoods. The weight g(e i,j ) is a measure of how good the connection between the two towns is: the greater g(e i,j ), the better they are connected, e.g. g(e i,j ) represents the number of people that can go from one town to the other in a fixed amount of time.
We denote by the total population, or mass, of a graph.
Moreover, we define its boundary ∂Γ k as In our model we are interested in pairwise disjoint and connected partitions of the graph. For sake of brevity, in the following we shall only say N -partition, truly referring to a pairwise disjoint, connected N -partition accordingly to the next definition.
We define the cut set of the partition (or boundary of the partition) as Given N , we would like to find a N -partition of Γ such that all subgraphs Γ k have the same total weight of the vertex, i.e. the same mass In our electoral interpretation this means dividing the region Γ in N electoral districts with the same population, while minimizing the total weight of the cut set to ensure that the districts are as "compact" as possible. We shall refer to the above as to the cut or perimeter energy of the partition. This of course is not possible since in general the set of N -partitions with all subgraphs with the same weight is empty. Thus, we want to allow the possibility for the weight of the subgraphs to differ from the arithmetic mean, at a cost, increasing with the difference from the mean. More precisely, we add the penalization term which we shall refer to as the deviation energy of the partition. Given λ ∈ [0, 1], we define the energy functional which is a convex combination of the cut energy (6) and of the deviation energy (7). We are interested in minimizing F λ among all N -partitions of Γ. Notice that the minimization problem is invariant under the action of a uniform dilation of the weights f and g, i.e. taking as weights θf and θg, with θ > 0, in place of f and g does not change the minimizers. We shall call any partition minimizing (8) a minimal or optimal N -partition (relative to some λ).
It is worth noticing that for λ = 1 the only energy we consider is the cut energy, i.e. the problem is similar to the minimal N-cut problem and there is no request on the mass of the districts to be near to the mean value. Remark 2.3. Existence of minimizers is trivial and follows straightforwardly from the finiteness of the graph Γ. Moreover, notice the following fact. GivenΓ = V (Γ); E(Γ) a subgraph belonging to a partition of Γ, one can suppose wlog that This is because the first term of the functional F λ , i.e. the cut energy, is defined on the cut set C({Γ k }) which is contained but not necessarily equal to E(Γ) \ ∪ k E(Γ k ). Thus, the functional does not detect any internal changes in a subgraphΓ, i.e. consideringΓ orΓ = (V (Γ), E(Γ) \ {e}) is the same, provided that removing e does not disconnectΓ. This means that the energy of a partition {Γ k } is actually a function of the partition of the vertexes {V (Γ k )} and does not depend on the edges.
Regarding uniqueness, note that fixed a N -partition its energy F λ is affine linear in λ. Hence, it easily follows a uniqueness theorem.
Proof. Given {Γ k } minimal in λ 1 and λ 2 (λ 1 < λ 2 ) we have that Notice now that the energy of any given partition is affine linear in λ and thus determined by its value in any two given points, or equivalently F λ is a linear combination of F λ 1 and F λ 2 . Hence by (9), the energy of any partition {Γ k } is greater than or equal to the energy of . Since the energy is affine linear, inequalities (9) paired with this last equality necessarily imply that the two partitions always have the same energy. The converse is readily achieved by using again the linearity w.r.t. λ and the hypothesis that at least one of the inequalities in (9) is strict.
Corollary 2.5. If there are no partitions that have the same energy in any two points, then F λ has a unique minimizer up to a finite set of threshold values.

Discussion of (desirable) properties
We analize here several properties we would like our districting model to have. We are interested in: (i) Stability of the division in districts when a new town/neighbourhood is built or abandoned; (ii) Stability of the division in districts when a new road connecting different cities is built (or an existing connection is destroyed); (iii) Possibility to force certain adjacent cities to be in the same district (up to suitably modifying the connection between the two); (iv) Possibility to force a city to form a district on its own (up to suitably modifying the connections between that city and its neighbors); (v) Stability of the model at a multi-layer districting, i.e. does creating "super"-districts, say N , and then splitting them, say in j "sub"districts each, yield the same result as directly creating jN districts? Our model does not grant properties (i), (ii), (v), as we show by counterexamples. Properties (iii) and (iv) hold true, under suitable conditions. The fact that properties (i), (ii) and (v) do not hold can be: either a weakness of the model deriving from the high degree of freedom of weights and may possibly be circumvented by modifying the form of the deviation energy; or an unavoidable drawback present in any possible model of districting. The latter is far from being unrealistic and it is actually something that often happens in social choices, e.g.
• Arrow in [1] proved that, if there are at least 3 choices, no electoral system satisfies at once Pareto's property (if all voters prefer X over Y , X is group-preferred to Y ), independence from irrelevant choices (the group preference between X and Y only depends on the single preferences between X and Y ) and there is no dictator (no single voter possesses the power to always determine the group's preference). • Balinski and Young in [2] proved that, if there are at least 3 parties, no apportionment system simultaneously follows the quota rule (if a party fair share is between n and n + 1 it gets assigned either n or n + 1 seats), avoids the Alabama paradox (if the total number of seats is increased, no party's number of seats decreases) and avoids the population paradox (if party A gets more votes and party B gets fewer, no seat will be transferred from A to B).

3.1.
Adding or removing a vertex. We here briefly discuss what can happen whenever a vertex is removed to a graph (resp. added). In our practical example of districting this could correspond to a city being abandoned (resp. to a new city being built).
Example 3.1. We consider the graph in Figure 1, where the grayed-out vertex is the new/removed vertex and the dashed edges are the edges connecting it to the other vertexes, and look at its minimal 2-partitions. We call Γ the whole graph andΓ its subgraph without the grayed-out vertex and the dashed edges. We start by analyzing the graph without the grayed-out vertex,Γ. Notice that for λ = 1 the minimal 2-cut can be obtained by cutting edges with a total weight of 2 (i.e. any two edges), while for λ = 0 there are two partitions with null energy (cutting either the two horizontal edges, or the two vertical edges). As the partitions obtained by removing two parallel edges are minimal both for λ = 0 and λ = 1 it follows from Theorem 2.4 that they are minimal for all values of λ and as well unique for λ ∈ (0, 1) as no other minimizer has the same energy on both the extrema of the interval. As for the graph with the grayed-out vertex, Γ, there are three different partitions to be considered, up to obvious symmetries: • D, the partition of least deviation energy, i.e. the one for which one vertex with weight 9 is in a district with the new vertex: the cut energy is 5, while the deviation energy is null • C, the partition of least cut energy, i.e. the one for which one vertex with weight 9 is a district on itself: the cut energy is 3, while the deviation energy is 2 • M , the one for which one vertex with weight 9 is in a district with a vertex of weight 1: the cut energy is 4, while the deviation energy is The energy of the other possible 2-partitions is strictly controlled from below by one of the above away from the extremal points, hence they can be a minimal 2-partition only for λ = 0, 1. Thus the optimal partition is: Hence, whenever 0 ≤ λ < 2 − √ 2 the minimal 2-partitions for Γ andΓ are significantly different.

3.2.
Adding or removing an edge. We here briefly discuss what can happen whenever an edge is removed to a graph (resp. added). In our practical example of districting this could correspond to a bridge collapsing (resp. to a new road being built).
Example 3.2. We consider the graph in Figure 2, where the dashed edge is the new/removed edge, and look at its minimal 3-partitions. If we call Γ the whole graph and e the dashed edge, we are then looking, respectively, at Γ and its subgraphΓ = V (Γ); E(Γ) \ {e} .
We start by analyzing the graph with the dashed edge, Γ. Notice that for λ = 1 the minimal 3-cut can be obtained by cutting edges with a total weight of 4 (the three right edges, the three left ones or the four horizontal ones), while for λ = 0 there are partitions with zero energy (cutting the 4 horizontal edges, or the left and center vertical edges and the two horizontal right edges, or the right and center vertical edges and the two horizontal left edges). As the partition obtained by removing the 4 horizontal edges is minimal both for λ = 0 and λ = 1 it follows from Theorem 2.4 that it is minimal for all values of λ and as well unique for λ ∈ (0, 1) as no other minimizer has the same energy on both the extrema of the interval. This 3-partition though induces a 4-partition for the subgraphΓ thus it can not be a minimizer of F λ onΓ for any λ! In particular the minimal solutions relative toΓ are the following: • for 0 ≤ λ ≤ 3 − √ 6 it is obtained by cutting the two horizontal left edges (F λ = (2 − √ 6)λ + √ 6); • for 3 − √ 6 ≤ λ ≤ 1 it is obtained by cutting the center vertical edge and the two horizontal right edges (F λ = 4λ).

3.3.
Forcing two vertexes in the same component. Let Γ be a graph and v 1 , v 2 be any two neighbouring vertexes. There are obviously graphs where it is impossible to have v 1 and v 2 in the same district (e.g. if there are N vertexes and we are looking for a N -partition). Yet, if there are Npartitions where these two vertexes are in the same district, one can force the minimal partition to be one of them, by a suitable modification of g on the edge e 1,2 . More precisely, the following result holds. Let us suppose Γ N,− = ∅, otherwise the claim is trivial. By hypothesis one has as well Γ N,+ = ∅. The λ-energy of any N -partition of Γ N,+ does not depend on the value of g(e 1,2 ). As the number of partitions is finite, we can let K λ be the minimum of those energies.
Notice now that the λ-energy of any N -partition {Γ k } ∈ Γ N,− can be expressed as where D λ ({Γ k }) does not depend on the value of g(e 1,2 ). Again in virtue of the finiteness of the partitions, let D λ be the minimum of D λ ({Γ k }), for {Γ k } ∈ Γ N,− . Thus, by modifying g such that g(e 1,2 ) > λ −1 (K λ − D λ ) one gets that the minimal partition belongs to Γ N,+ .
Notice that the above proof only works for a single pair and there is no clear way to force multiple pairs to stick together as the functional F λ displays a non-local behaviour with respect to the weights. 1 1 2M ε 1 Figure 3. Splitting two vertexes is not always possible by simply modifying the weight of their common edge.
Remark 3.4. Conversely, if one were to try to split two adjacent vertexes, provided that there exists a suitable partition, would find as a condition g(e 1,2 ) < λ −1 (K λ − D λ ), which is impossible to achieve if the RHS is non positive. There are indeed very easy cases for which this exact behaviour occurs. Take for instance the graph Γ in Figure 3, where M > 1. For any λ < √ 2(1 + √ 2) −1 the leftmost vertexes will stick together no matter how big M or how small ε is.

3.4.
Forcing a vertex to be isolated. By a straightforward application of the pigeonhole principle we can observe the following fact. Suppose that a graph Γ has vertex connectivity of 1 about the vertexv, i.e., the sub-graphΓ obtained by removing from Γ the vertex v and all related edges is disconnected. IfΓ has k = N + l connected components, l ∈ N, then in any N -partition of Γ at least l + 1 of these components belong to the subgraph containing v. In such a situation thus there is no way to force v to be "isolated" i.e. to form a district by its own.
It is then of interest, assuming the necessary assumption that there exist N -partitions containing the singleton vertex Γ v := {v}; ∅ , to know if one can force the vertex to be isolated by modifying the weights of the edges in ∂Γ v . We are able to prove that this is possible for values of λ near to 1 limitedly to 2-partitions. More precisely, the following theorem holds. Proof. Start noticing that if {Γ v , Γ c v } is a 2-partition, then it is the unique 2-partition containing the subgraph Γ v (up to removing edges of Γ c v which do not disconnect it as noted in Remark 2.
while for any 2-partition {Γ k } not containing Γ v one has Our claim corresponds to proving that F λ ({Γ k })−F λ ({Γ v , Γ c v }) ≥ 0 for some λ near to 1 and for an appropriately small perimeter of ∂Γ v , i.e. ∂Γv g(e).  Figure 4. A graph where the optimal 4-partition is not a 2-refining of the optimal 2-partition.
We have As for any 2-partition not containing Γ v one has that C({Γ k }) \ ∂Γ v = ∅ the claim follows for λ close to 1 and ε << 1, i.e. the sum of the weights of the edges in ∂Γ v .
Remark 3.6. Theorem 3.5 is in agreement with the idea that assigning a zero weight to an edge means that the edge is missing. Notice that the hypothesis can be equivalently reformulated by suitably modifying all weights not in ∂Γ v and taking each one of them greater than some M >> 1.
The result is somewhat weak as it holds only for 2-partitions. It would be desirable to extend it to N -partitions but it is unclear how to do it, as in the energy expression for a partition containing Γ v would appear the additional term λ C({Γ k })\∂Γv g(e) on which we have no control.
It is as well unclear if fixed any λ > 0, by suitably modifying the perimeter of ∂Γ v , one can force the vertex to be isolated, even in the 2-partition case. Clearly, one would expect that as λ → 0, necessarily ∂Γv g(e) → 0 as well.

3.5.
Refining the number of districts. Given j ∈ N, we shall say that a jN -partition One might wonder if some "refining" property holds. For instance, it would be desirable that any (or at least one) minimal jN -partition is a refining of a minimal N -partition. This is exactly the situation where you have two different kinds of elections (e.g. Italian and European elections) and one would like to form "super"districts for one election just by gluing together some districts of the other election.
Clearly, a necessary condition to allow such a situation is that there exists an optimal N -partition such that each of its subgraph has at least j vertexes. Even if this necessary condition is satisfied, this refining property can fail, as the following example shows.  Figure 5. A graph where the optimal 4-partition is a 2refining of the optimal 2-partition, but does not induce the optimal 2-partitions on its components. Example 3.7. Consider the graph Γ of Figure 4. The unique optimal 2-cut is obtained by removing the two edges with weight 1. This same partition has zero deviation energy, hence by Theorem 2.4 it is an optimal 2-partition for every λ and the unique one for λ = 1.
Removing the four edges with weights 2 and 4, gives a 4-partition D with zero deviation energy and cut energy 12λ. Thus, any 4-partition whose cut set contains an edge of weight 10 cannot be optimal.
Hence, up to symmetries, there is only one possibly optimal 4-partition which is a 2-refining of the minimal 2-partition: the one in which are cut one weight 2 and one weight 4 edges. This partition R has λ-energy F λ (R) = 8λ + 2(1 − λ) , and it is never minimal. Indeed, either D or C, the partition whose cut energy is 6λ, is the optimal one. More precisely, thus, C is the optimal 4-partition, withλ = √ 6/(6 + √ 6) being the threshold where they are both minimizers.
Even if we assume that a minimal jN -partition is a j-refining of a minimal N -partition, then {Γ i } kj i=(k−1)j+1 may not be a minimal-j-partition of Γ k , as shown by the following example.
Example 3.8. Consider the graph Γ in Figure 5, and fix 1 λ = 1 2 . Let us consider first the 2-partitions. The optimal 2-cut (λ = 1), C 2 , is obtained by removing the two edges of weight 4. It has energy and for λ = 1 2 , F 1/2 (C 2 ) = 4 + 3 √ 2 < 17 2 . Hence, all other 2-partitions whose cut energy is greater than 17 cannot be an optimal 2-partition (w.r.t. 1 The choice λ = 1 2 has nothing special. For any λ ∈ (0, 1) the weights of the graph Γ can be suitably chosen to have the same behavior of the shown example. the choice λ = 1 2 ). This leaves only two possible partitions with cut energy 16. Both of them have deviation energy greater than that of C 2 . Hence, C 2 is the optimal 2-partition of Γ w.r.t. the choice λ = 1 2 . Consider now the 4-partitions. The optimal 4-cut (λ = 1) is obtained by removing all edges but one with weight 16. Among the 2 possible choices, let C 4 be that with smaller deviation energy, i.e. the case where the 14 and 4 weight vertexes are grouped together. This 4-partition has energy and for λ = 1 2 , F 1/2 (C 4 ) = 18 + √ 17. There is only another 4-partition with smaller deviation energy, D 4 , in which the 10 and 4 weight vertexes form a district. Its energy is F λ (D 4 ) = 40λ + 6(1 − λ) , and for λ = 1 2 is 23, strictly more than that of C 4 . Hence, C 4 is the optimal 4-partition for λ = 1 2 . Observe that C 4 is a 2-refining of C 2 , yet it does not define minimal 2partitions for each of the subgraphs of C 2 . Consider indeed the district with 3 vertexes of C 2 . There are two ways to partition it in 2 subgraphs, one induced by C 4 , denoted byĈ 4 , and one induced by D 4 , denoted byD 4 . Their energies are Relatively to λ = 1 2 , one has F 1/2 (Ĉ 4 ) = 6 + 2 √ 2 > 8 = F 1/2 (D 4 ) .
Thus, even though C 4 is a 2-refining of C 2 it does not induce minimal-2partitions of the subgraphs of C 2 .

Conclusions
The model we propose here is very general and it lays the foundations for future work. On the one hand, we point out that the main results we proved, Theorem 2.4, 3.3 and 3.5 would still hold if we were to replace the deviation energy with a general p-norm, independently from the choices of the weights. Further, they would hold even if we were to take more general penalizations terms. It is truly interesting that as a practical consequence of Theorem 3.3 and of Theorem 3.5 (specifically with the equivalent hypothesis highlighted in Remark 3.6) the best course of action for politicians to achieve some particular partition would be to improve some connections.
On the other hand, all "desired" properties we checked to fail, do so because the two components of the energies do not interact. This is because the weights f and g can be arbitrarily chosen. It is possible that by forcing some constitutive relation between the edge and the vertex weights these could be ensured.

Open problems.
There are some open problems we are interested in studying in the future. Among those • generalizing the result of Theorem 3.3 trying to force multiple distinct pairs together; • generalizing the result of Theorem 3.3 trying to force n vertexes together; • generalizing the result of Theorem 3.5 to N -partitions.
• trying to either find a modification of the model for which the properties (i), (ii) and especially (v) hold or prove a general impossibility result in the spirit of Arrow [1] and Balinski-Young [2]. Clearly there are some necessary hypotheses to be made. Moreover we would expect that forcing multiple vertexes together would require a stringent structure of the subgraph they form, e.g. to be a cycle or wheel graph. It is very possible that these properties do not hold for the fully general model, but we do not have any counterexample to exhibit at the current stage.
Finally, we plan to impose some constraints on the choices of f and g by pairing them via some suitable equation. This would imply a more rigid structure to the energy possibly leading to stronger theorems. For instance a possible choice would be to ask f (v) := ∂Γv g(e) , being Γ v the subgraph consisting of the lone vertex v. This choice would mean, in the politics' application, that each district ensures a flow-in/out of people equal to the number of its citizens.

Declarations of Interest
Declarations of interest: none