Group Decision Making

Earlier decision models centered on a single decision maker — one player weighing actions against an uncertain system, or two players in direct opposition. Group decision making changes the question entirely: when many individuals each rank the same set of options by their own lights, how do we merge those individual rankings into a single decision that speaks for the collective?

The building block is a precise way to write down one individual’s ranking.

Rank mappings

Fix a finite set $A$ of candidates — the options everyone is ranking, whether that’s candidates in an election, teams in a sports league, or proposals up for a vote. One individual’s ranking is recorded as a function that hands each candidate its place.

A rank mapping on a finite candidate set $A$ is a function $r : A \to \mathbb{N}$ that is surjective onto a set of the form $\{1, \dots, k\} \subset \mathbb{N}$ . The value $r(x)$ is the place of candidate $x$ : $r(x) = 1$ marks the top candidate, $r(x) = 2$ the next, and so on down to the last place $k$ .

Concretely, this means the places run $1, 2, 3, \dots$ with no gaps: if some candidate sits in place $3$ , then places $1$ and $2$ are occupied too. The last place $k$ is however far the ranking reaches — as small as $k = 1$ when every candidate ties for first, as large as the number of candidates when no two share a place.

Think of a sports league table at the end of a season: $r(x)$ is the row candidate $x$ finishes in. A rank mapping need not be injective — two candidates can be assigned the same place, which is exactly a tie. Injectivity would forbid ties; surjectivity onto $\{1, \dots, k\}$ only forbids gaps. A small example shows how the two conditions interact.

With candidates $A = \{x, y, z\}$ , the assignment $r(x) = r(y) = 1$ , $r(z) = 2$ is a valid rank mapping: $x$ and $y$ tie for first, $z$ takes second, and the image $\{1, 2\}$ has no holes. But $r(x) = r(y) = 1$ , $r(z) = 3$ is not a rank mapping — place $2$ is empty, so $r$ is not surjective onto $\{1, 2, 3\}$ . A tie shifts the candidates below it up; it does not leave a gap behind.

Preference relations

A rank mapping carries more than its raw place numbers. For any two candidates it also says which one the individual prefers — and that comparison, stripped of the actual numbers, is the preference relation.

The preference relation $\varrho$ induced by a rank mapping $r$ on $A$ is defined by

x \varrho y \ :\Leftrightarrow\ r(x) < r(y)

The symbol $:\Leftrightarrow$ reads “is defined to hold exactly when”. So $x \varrho y$ holds precisely when $x$ has the smaller place number, and is read ” $x$ is better than $y$ ”.

The direction is worth pausing on: a smaller place number means a better placing, so ” $x$ is better than $y$ ” lines up with $r(x) < r(y)$ — first place beats second place. The preference relation discards the absolute place numbers and keeps only the order they impose on the candidates. The symbol $\varrho$ itself is a script variant of the Greek letter rho.

Many voters

So far $\varrho$ records just one individual’s ranking. The actual problem of group decision making involves many. Write $I = \{1, \dots, n\}$ for a set of voters, where each voter $i$ comes with their own rank mapping $r_i$ , and hence their own preference relation. The goal of the rest of this chapter is a method that turns the whole list of individual preferences into a single collective preference — and does so as “fairly” as possible.

Pinning down what “fair” should mean, and seeing how far it can be pushed, needs the language of relations: which structural properties a preference can have — and, crucially, which of those properties survive the act of aggregation.

Relations and their properties

This part is pure mathematical groundwork — a relation is a general object, worth setting up on its own before it touches preferences again.

A binary relation (or simply a relation) on a set $X$ is a subset $R \subseteq X \times X$ — a collection of ordered pairs of elements of $X$ . When the pair $(x, y)$ belongs to $R$ , write $x R y$ and say ” $x$ is related to $y$ ”.

The preference relation $\varrho$ from earlier is already an example of one: it is a relation on the candidate set $A$ , the set of pairs $(x, y)$ for which $x$ outranks $y$ . What the next definitions do is name the structural features any relation $R$ can have — and those features are exactly what will decide whether a relation behaves like a sensible preference.

A relation $R$ on $X$ is reflexive if every element is related to itself:

x R x \quad \text{for all } x \in X

A relation $R$ on $X$ is transitive if relatedness carries across a middle element:

x R y \ \text{ and }\ y R z \ \implies\ x R z

A relation $R$ on $X$ is symmetric if it always holds both ways at once:

x R y \ \implies\ y R x

A relation $R$ on $X$ is asymmetric if it never holds both ways at once:

x R y \ \implies\ \neg(y R x)

So $x R y$ and $y R x$ can never be true simultaneously.

A relation $R$ on $X$ is connex (also called total or complete) if any two elements are comparable in at least one direction:

\neg(x R y) \ \implies\ y R x

Equivalently, for every pair at least one of $x R y$ , $y R x$ holds — no two elements are left mutually unrelated. This is universal comparability.

A relation is a quasi-order (also called a preorder) if it is both reflexive and transitive.

These six are not a random checklist — different combinations carve out the familiar order structures, and which boxes a relation ticks is exactly what tells us whether it can serve as a preference.

The structure of preference relations

Run the preference relation $\varrho$ through that checklist. Recall $x \varrho y$ means $r(x) < r(y)$ , so $\varrho$ is built directly from the strict order $<$ on $\mathbb{N}$ and inherits its structure:

transitive: $r(x) < r(y)$ and $r(y) < r(z)$ force $r(x) < r(z)$ .
asymmetric: $r(x) < r(y)$ rules out $r(y) < r(x)$ .

It is not connex, though: when $x$ and $y$ are tied, $r(x) = r(y)$ , so neither $x \varrho y$ nor $y \varrho x$ holds — the two candidates are left mutually unrelated. Hold that thought; it is exactly the gap a second relation will fill below.

The transitive, asymmetric relations that arise this way are the ones we want to call preferences, so the whole collection of them deserves a name.

The set of preference relations on a candidate set $A$ , written $P_A$ , collects every preference relation generated by some rank mapping on $A$ :

P_A := \{\, \varrho \subseteq A \times A \mid \varrho \text{ is generated by a rank mapping } r \,\}

Every member of $P_A$ is transitive and asymmetric.

It is worth pausing on what kind of object $P_A$ is. A relation is itself a set — the ordered pairs $(x, y)$ that belong to it — so $P_A$ , whose elements are relations, is a set of sets: each element is one whole preference, packaged as its collection of pairs. The useful way to read it is as a menu: $P_A$ is every preference a single voter could possibly hold, and for that voter, having a preference is exactly singling out one element of $P_A$ . That framing carries the rest of the chapter — the aggregation methods to come are functions that take one such choice from each of the $n$ voters and hand back a single collective relation, itself an element of $P_A$ .

The same rank mapping $r$ also generates a second relation — the one that supplies the comparability $\varrho$ was missing.

From a rank mapping $r$ , define the relation $\varrho^*$ using the non-strict order $\le$ :

x \varrho^* y \ :\Leftrightarrow\ r(x) \le r(y)

Read $x \varrho^* y$ as ” $x$ is at least as good as $y$ ”. Because $\le$ compares any two places, $\varrho^*$ is connex, and it is also a quasi-order — reflexive and transitive — making it a connex preorder. The set of connex preorders on $A$ , written $P_A^*$ , collects every $\varrho^*$ obtained this way.

The strict relation $\varrho$ and the non-strict $\varrho^*$ are not independent — each one determines the other. Unfolding the definitions gives the conversion identity

x \varrho y \ \Leftrightarrow\ \neg(y \varrho^* x)

— ” $x$ is better than $y$ ” is the exact negation of ” $y$ is at least as good as $x$ ”. The upshot is that one and the same preference can be carried in three interchangeable forms: as the rank mapping $r$ , as the strict relation $\varrho \in P_A$ , or as the connex preorder $\varrho^* \in P_A^*$ . Later sections reach for whichever of the three is most convenient.

Collective choice function

The pieces are now in place: $n$ voters, each holding some preference relation from $P_A$ . What is still missing is the actual goal of group decision making — a rule that reads off everyone’s preferences and returns one preference for the collective.

A collective choice function is a function

K : P_A^n = \underbrace{P_A \times P_A \times \dots \times P_A}_{n \text{ copies}} \to P_A

It takes a tuple $(\varrho_1, \dots, \varrho_n)$ of $n$ individual preference relations — one per voter — and returns a single collective preference relation $K(\varrho_1, \dots, \varrho_n) \in P_A$ .

Recall that $P_A$ is the menu of preferences a single voter can hold. The domain $P_A^n$ stacks $n$ copies of that menu, one slot per voter; writing it as a Cartesian product — $n$ copies multiplied together — makes the slots independent, so an element of $P_A^n$ is any way of filling every slot, one full snapshot of how the electorate has chosen. $K$ collapses that whole snapshot down to a single relation, and because the codomain is $P_A$ again, the collective verdict comes out as the same kind of object as the individual inputs.

That codomain is not a throwaway detail. The signature $K : P_A^n \to P_A$ already bakes in two of the democratic basic rules, before $K$ computes anything. Rule 1: $K$ is defined on all of $P_A^n$ — every voter is free to submit any preference, no ballot combination ruled out in advance. Rule 2: the output lands back in $P_A$ — the collective verdict is itself a valid preference relation. These two, together with three further rules that pin down what makes a $K$ genuinely fair, are collected and stated in full in the democratic basic rules.

Majority decision

With the collective choice function framed, the most natural first attempt at building one is majority rule: settle every pairwise contest by counting votes. Each voter $i$ submits their individual preference $\varrho_i \in P_A$ , and the method compares candidates two at a time.

The majority decision (also called the Condorcet method) builds a collective preference by pairwise majority vote. For candidates $x, y$ , let

N(x, y) = \text{number of voters } i \text{ with } x \varrho_i y

be the count of voters who individually rank $x$ above $y$ . The majority relation $\varrho$ is then defined by

x \varrho y \ :\Leftrightarrow\ N(x, y) > N(y, x)

— the collective prefers $x$ to $y$ exactly when more voters rank it that way than the other way.

Note the two roles $\varrho$ now plays: each $\varrho_i$ is one voter’s individual preference, while the unsubscripted $\varrho$ is the collective relation the method outputs. Majority decision is a concrete attempt at a collective choice function — feed it the tuple $(\varrho_1, \dots, \varrho_n)$ , get back the majority relation $\varrho$ . The only question is whether that $\varrho$ is a legal output: does it land in $P_A$ ?

It does not always. The majority relation can fail to be transitive.

The voting paradox (also called the Condorcet paradox) is the phenomenon that the majority relation can fail to be transitive: the pairwise majority votes can form a cycle — $x$ beats $y$ , $y$ beats $z$ , yet $z$ beats $x$ — even when every individual voter’s preference is a perfectly transitive ranking.

Take three candidates $A = \{x, y, z\}$ and three voters $I = \{1, 2, 3\}$ with the rank mappings:

Voter	$r_i(x)$	$r_i(y)$	$r_i(z)$
$i = 1$	$1$	$2$	$3$
$i = 2$	$3$	$1$	$2$
$i = 3$	$2$	$3$	$1$

Each row is one voter; the three columns hold that voter’s rank mapping $r_i$ — the place $r_i(x), r_i(y), r_i(z)$ they assign to each candidate. A smaller number is a better place, so voter $1$ ‘s row $1, 2, 3$ reads ” $x$ first, $y$ second, $z$ last”. Every individual ranking here is a clean, transitive preference.

Now tally the pairwise contests:

$x$ vs $y$ — voters $1$ and $3$ rank $x$ above $y$ , so $N(x, y) = 2 > 1$ and $x \varrho y$ .
$y$ vs $z$ — voters $1$ and $2$ rank $y$ above $z$ , so $N(y, z) = 2 > 1$ and $y \varrho z$ .
$z$ vs $x$ — voters $2$ and $3$ rank $z$ above $x$ , so $N(z, x) = 2 > 1$ and $z \varrho x$ .

The three results $x \varrho y$ , $y \varrho z$ , $z \varrho x$ close into a cycle — a rock-paper-scissors with no top candidate. And the cycle breaks transitivity. If $\varrho$ were transitive, $x \varrho y$ and $y \varrho z$ would force $x \varrho z$ ; chaining that with $z \varrho x$ would force $x \varrho x$ — the collective strictly preferring $x$ to itself. No relation in $P_A$ can do that: every member of $P_A$ is asymmetric, so it never holds between an element and itself. Hence $\varrho$ is not transitive, and $\varrho \notin P_A$ .

This is a direct failure of Rule 2 — the requirement that $K$ ‘s output land back in $P_A$ . The majority decision can return a relation that is not a valid preference at all, so it is not a collective choice function. The unsettling part is the contrast: every individual voter in the example is perfectly reasonable, each $\varrho_i$ a clean transitive ranking, and still the majority of those reasonable preferences is not reasonable.

Be precise about what has and has not been shown. At one level the result is flatly definitive: the Condorcet method — this one specific rule, pairwise majority counting — is genuinely broken as a collective choice function, since its output can land outside $P_A$ . That much is settled. But the result reaches no further than that single rule. It does not show that some other aggregation method must fail the same way, and it does not show that real-world majority voting, or democracy, is “unreasonable”. The rhetorical question “are majority decisions unreasonable?” gets its bite by trading on a word — “majority decision” is both the name of this one formal rule and the loose everyday idea of letting the majority decide. Only the formal rule has been put to the test. Reading its failure as a verdict on the broader idea, let alone on reality, is the step too far.

Rank addition

A collective choice function is just any rule of the right shape; the real work is finding one that is also fair. The Condorcet method was the first candidate, and it failed Rule 2. Rank addition is the next candidate — not a variant of the Condorcet method but a structurally different rule. Where the Condorcet method compares candidates pairwise, rank addition gives each candidate a single global score.

Rank addition builds a collective preference by adding up the individual rank mappings. Each candidate’s total score is the sum of the places the voters gave it, and the relation $\varrho$ it generates is defined by

x \varrho y \ :\Leftrightarrow\ \sum_{i=1}^{n} r_i(x) < \sum_{i=1}^{n} r_i(y)

The collective prefers $x$ to $y$ exactly when $x$ ‘s total is the smaller — the voters’ places for $x$ add up to less than their places for $y$ .

The sum $\sum_i r_i$ assigns each candidate a number, but it is not itself a rank mapping: its values need not form a gap-free block $\{1, \dots, k\}$ — in the example below the totals come out as $3, 3, 6$ , whose image $\{3, 6\}$ skips almost everything. Even so, the relation $\varrho$ it generates still lands in $P_A$ .

Why the relation is still a valid preference

A preference relation records only the order of the candidates, not the particular numbers behind it. So even though $\sum_i r_i$ is not a rank mapping, the ordering it produces can always be reproduced by one: take the distinct totals, sort them, and assign place $1$ to the lowest group, place $2$ to the next, and so on — the totals $3, 3, 6$ collapse to the genuine rank mapping $1, 1, 2$ , which generates the very same $\varrho$ . This is what surjectivity onto $\{1, \dots, k\}$ was always for. It was never needed to make the ordering valid — it is needed to make the rank mapping unique. Without the no-gaps requirement, $3, 3, 6$ and $1, 1, 2$ and $5, 5, 9$ would all serve equally well as rank mappings for the same preference; pinning the places to a gap-free block singles out exactly one.

With that settled, rank addition clears the first two of the democratic basic rules. It is defined on every input — any voter may submit any $\varrho_i \in P_A$ and the sum is always computable (Rule 1) — and its output always lands in $P_A$ (Rule 2). Unlike the Condorcet method, rank addition genuinely is a collective choice function.

But clearing Rules 1 and 2 is not the end of the story. Rank addition has a subtler flaw, one that surfaces only when two different ballots are compared side by side.

Two voters rank three candidates $A = \{x, y, z\}$ , and rank addition is run on two separate ballots.

First ballot. The voters submit $(\varrho_1, \varrho_2)$ , which rank addition aggregates into the collective relation $\varrho$ :

Voter	$r_i(x)$	$r_i(y)$	$r_i(z)$
$i = 1$	$1$	$2$	$3$
$i = 2$	$2$	$1$	$3$
$\sum r_i$	$3$	$3$	$6$

The totals are $3, 3, 6$ , so $\varrho$ has $x$ and $y$ tied, both ahead of $z$ .

Second ballot. Voter $1$ submits the exact same ranking as before. Voter $2$ also keeps $y$ in first place, but swaps the places of $x$ and $z$ — $x$ drops from second to last, $z$ rises from last to second:

Voter	$r_i(x)$	$r_i(y)$	$r_i(z)$
$i = 1$	$1$	$2$	$3$
$i = 2$	$3$	$1$	$2$
$\sum r_i$	$4$	$3$	$5$

Now the totals are $4, 3, 5$ , so $\varrho'$ has $y$ strictly ahead of $x$ , and $x$ ahead of $z$ .

Compare the two ballots on the single pair $x$ versus $y$ . Voter $1$ ranks $x$ above $y$ in both. Voter $2$ ranks $y$ above $x$ in both. Not one voter changed their mind about $x$ versus $y$ . Yet the collective verdict on that pair moved — a tie in the first ballot, $y$ strictly winning in the second. The only thing that changed was voter $2$ lifting $z$ past $x$ , which pushed $x$ ‘s total from $3$ to $4$ and broke the tie.

This is a failure of Independence of Irrelevant Alternatives (IIA). The principle says the collective ranking of $x$ against $y$ should depend only on how the voters rank $x$ against $y$ — for that one comparison, candidate $z$ is an irrelevant alternative, and moving it around should not matter. Rank addition violates it: because each candidate’s score is built from its place relative to all the others, shuffling an irrelevant $z$ leaks into the $x$ -versus- $y$ result. Its precise form is one of the democratic basic rules introduced further on.

Dictatorships

Rank addition cleared Rules 1 and 2 — so are those two rules enough to call a collective choice function fair? They are not, and the cleanest way to see it is with two collective choice functions that pass both rules effortlessly and are still, obviously, the opposite of fair.

An external dictator is the collective choice function that ignores every voter and always returns one fixed preference relation $\varrho_E \in P_A$ , chosen by someone outside the electorate:

K^E_{\varrho_E}(\varrho_1, \dots, \varrho_n) := \varrho_E

The output never depends on the inputs $\varrho_i$ at all.

An internal dictator is the collective choice function that fixes one distinguished voter $d$ and always returns that voter’s preference, ignoring everyone else:

K^D_d(\varrho_1, \dots, \varrho_n) := \varrho_d

The output depends on $\varrho_d$ alone — the preferences $\varrho_i$ of all other voters ( $i \neq d$ ) have no effect on it.

Both are genuine collective choice functions. The external dictator is a constant function — defined on all of $P_A^n$ (Rule 1), and always returning the same fixed $\varrho_E \in P_A$ (Rule 2). The internal dictator is a projection — it hands back one coordinate of the input tuple: defined on all of $P_A^n$ (Rule 1), and returning $\varrho_d$ , which, being a voter’s own preference, is in $P_A$ by construction (Rule 2). Each one produces a “reasonable” result on every input, in the exact sense Rules 1 and 2 ask for: a well-formed preference relation, always.

And yet neither is remotely democratic. The external dictator listens to nobody; the internal dictator listens to one person and discards the other $n - 1$ votes. They are the proof that Rules 1 and 2 together only guarantee the output is well-formed — they say nothing about whether the procedure is fair. A collective choice function can satisfy both rules and still be a dictatorship.

So “reasonable” is not the same as “desirable”, and the gap between them has to be closed explicitly. What is missing is a description of what democratic even means — pinned down not by intuition but as an explicit list of requirements that every fair collective choice function should meet. Those are the democratic basic rules; Rules 1 and 2 are only the first two.

The democratic basic rules

Rules 1 and 2 came for free with the signature of a collective choice function, and on their own they were not enough. The Condorcet method broke Rule 2; rank addition cleared both yet still misbehaved; and the dictators of the previous section cleared both while being the very opposite of democratic. The fix is to write down, in full, what a fair collective choice function should do — five democratic basic rules, the first two of which we have already met. Each of the five is a constraint on the collective choice function $K$ itself, so its definition is repeated here for reference:

[embedded: def-collective-choice-function]

Rule 1: Universal domain

$K$ must be defined on the whole of $P_A^n$ — every tuple of individual preferences is a legal input, and $K$ returns an answer for each. In function terms, $K$ is a total function: one defined on the entirety of its domain, with no input left unhandled. The democratic content is individual freedom — any voter may submit any preference $\varrho_i \in P_A$ , no combination of ballots is ruled out in advance, and nobody is handed extra weight (no “you have two votes”).

The word total here is unrelated to the total of connex relations. There, “total” described a relation in which every pair of elements is comparable; here it describes a function defined everywhere on its domain. Same word, two different ideas.

Rule 2: Collective rationality

The result of $K$ must always lie in $P_A$ — whatever the ballots, $K(\varrho_1, \dots, \varrho_n)$ is itself a genuine preference relation: transitive, asymmetric, a real ranking and not a cycle. The collective verdict is held to exactly the standard of well-formedness demanded of each individual voter. This is the rule the Condorcet method broke.

Rule 3: Pareto condition

Rules 1 and 2 only constrain the shape of $K$ ; the remaining three constrain its behavior. The first of them is the Pareto condition.

The Pareto condition (Rule 3) requires unanimity to carry through: if every voter ranks $x$ above $y$ , the collective must too. With $\varrho = K(\varrho_1, \dots, \varrho_n)$ ,

\big(\forall i \in \{1, \dots, n\} : x \varrho_i y\big) \ \Rightarrow\ x \varrho y

This is the supremacy of the collective — a unanimous electorate can always enforce its agreed ranking. If literally everyone prefers $x$ to $y$ , no rule worth the name returns anything else.

Rule 4: Independence of irrelevant alternatives

Rule 4 is the one rank addition violated.

Independence of irrelevant alternatives (IIA, Rule 4) requires the collective ranking of any two candidates to depend only on how the voters rank those two. Take any two ballots $(\varrho_1, \dots, \varrho_n)$ and $(\varrho'_1, \dots, \varrho'_n)$ , with collective results $\varrho = K(\varrho_1, \dots, \varrho_n)$ and $\varrho' = K(\varrho'_1, \dots, \varrho'_n)$ . Then for every pair $x, y \in A$ ,

\big(\forall i \in \{1, \dots, n\} : x \varrho_i y \Leftrightarrow x \varrho'_i y\big) \ \Rightarrow\ \big(x \varrho y \Leftrightarrow x \varrho' y\big)

If every voter ranks $x$ against $y$ the same way in both ballots, the collective must rank $x$ against $y$ the same way in both — however the ballots differ on other candidates.

For the $x$ -versus- $y$ question those other candidates are irrelevant alternatives, and shuffling them must not leak into the verdict. Rank addition is the cautionary case: its two-ballot comparison flipped the collective ranking of $x$ and $y$ purely by relocating an irrelevant $z$ , with no voter changing their mind about the pair.

Rule 5: No-dictator condition

The remaining rule rules out a single voter dominating the outcome.

The no-dictator condition (Rule 5) forbids any single voter from always getting their way. There must be no voter $d$ such that, for every ballot and every pair of candidates, that voter’s strict preference is automatically the collective’s:

\neg\, \exists\, d \in \{1, \dots, n\} : \forall (\varrho_1, \dots, \varrho_n) \in P_A^n : \forall (x, y) \in A \times A : x \varrho_d y \Rightarrow x \varrho y

Such a $d$ would be a dictator — and the internal dictator of the previous section is exactly one. Rule 5 says no such voter exists.

Rule 5 does not stop a voter from often agreeing with the collective; it stops any one voter from always prevailing, on every pair and across every possible ballot.

With all five rules on the table, the two foils from the previous section fall into place. The internal dictator satisfies Rules 1–4 and trips only on Rule 5 — it is the dictator that rule names. The external dictator is shut out by Rule 3 instead: a single fixed ranking $\varrho_E$ cannot honor every unanimous electorate, since voters can be unanimous on any pair in either direction and one fixed relation cannot agree with all of them. Each foil is killed by exactly one rule.

A fair collective choice function, then, is one that satisfies all five democratic basic rules at once. Whether such a function can exist at all is the question the next section settles.

Arrow’s impossibility theorem

Every method examined so far — the Condorcet method, rank addition, and the two dictatorships — breaks exactly one of the five democratic basic rules, and a different rule each time. The pattern invites optimism: with rules this reasonable, surely a cleverer method clears all five at once.

There is even a sliver of encouragement. The Condorcet method’s one failure — the voting paradox — needs a cycle, and a cycle needs at least three candidates to run through. With only two candidates on the ballot there is no room for one: the majority relation is automatically transitive, lands in $P_A$ , and Rule 2 holds. For a two-candidate election, the majority decision satisfies all five rules.

This two-candidate escape has been put forward as a theoretical argument for the Anglo-American two-party system — hold the field to two candidates and majority rule is immune to the paradox. It is more provocation than proof, but a striking thought all the same.

The escape is narrow, though. The moment a third candidate joins, the cycle is possible again — and “three or more candidates” is exactly the setting the central result speaks to.

Arrow’s impossibility theorem (Kenneth Arrow, 1951). If there are more than two candidates ( $|A| > 2$ ) and more than one voter ( $n > 1$ ), then no collective choice function satisfies all five democratic basic rules at once.

Rules 1 and 2 come built into what a collective choice function is, so the theorem’s real force lands on the other three: the Pareto condition, independence of irrelevant alternatives, and the no-dictator condition cannot all hold at once. It is not a verdict on any one method but on all of them — the chapter-long search, method by method, was never going to succeed. The five rules, each entirely reasonable alone, are jointly unsatisfiable: in the precise sense they give to the word, there is no fair collective choice function.

A weakness of the model, or of reality?

An impossibility result this sweeping forces the question that, in modeling, is always the important one. Does Arrow’s theorem expose a weakness of reality — is fair collective decision-making, in any robust sense, genuinely unattainable? Or does it expose a weakness of the model — is the impossibility an artifact of how the problem was set up?

The second reading deserves real weight, because every step in this chapter rested on a modeling choice that could have gone differently. The five rules are not laws of nature; they are a proposal for what “fair” should mean, and a different proposal might be jointly satisfiable. And beneath the rules sits an assumption easy to wave past: that each individual’s preference is a connex preorder — a ranking in which every pair of candidates is comparable. Is that what a real preference looks like? Someone faced with two options they genuinely cannot compare may simply have no answer, and a connex preorder leaves no room for “I cannot compare these”. If the model’s picture of an individual preference is already too rigid, the impossibility it generates may say more about the model than about democracy.

None of this softens the theorem — Arrow’s result is exact and proven. What stays open is what it means, and the honest answer is the one this chapter has come back to more than once: a model’s limitation is a fact about the model, until reality is shown to share it.