For a formal scientific paper that does the same stuff as this web page,
see paper #98 here.
Return to main page
We begin with some proofs that (essentially) range voting satisfies both AFB ("avoids favorite betrayal") and ICC ("immune to candidate cloning") but that no voting method based on rank-order ballots can satisfy both. Later we give a long list of theorems, each of which exhibits an interesting set of criteria which no voting method based on rank-order ballots can satisfy – but in each case, range voting satisfies them all. In some of these theorems (which we shall note), rank-equalities are permitted in the rank-order ballots.
Warren D. Smith's proof that (for pure rank ballots) ICC & AFB incompatible (essentially): Theorem: These four criteria, for a voting system which inputs pure-rank-order-ballots and outputs the name of a winner (or a set of co-equal winners to be chosen among by random tiebreak), are incompatible: a1. AFB = avoids favorite betrayal a2. ICC = immune to candidate cloning a3. no vetoer = There does not exist a voter who can single-handedly prevent a candidate of her choice from winning, regardless of how the other voters vote. a4. neutrality = symmetry under candidate renaming = permuting the candidate names on the ballot rankings permutes their winning probabilities in the same way. Equivalent Theorem: These six criteria, for a single-winner voting system based on pure-rank-order-ballots, are incompatible: a1. AFB = avoids favorite betrayal a2. ICC = immune to candidate cloning a3. reduces to simple majority vote in 2-candidate case. As we shall discuss below, due to Campbell-Kelly 2003 this criterion can be replaced by the demand that a "vetoer" does not exist (i.e. no voter exists, who can singlehandedly prevent any candidate X he chooses from winning, regardless of all the other votes) and the demand that the system be deterministic (chance is not employed except where required by symmetry, i.e in the case of true ties; or, better, we can regard any system obeying a5 as being 100% deterministic but simply outputting tied-winner sets; this interpretation is also ok), and the demand it be based on rank-order ballots. And this replacement is desirable since you get a stronger theorem. But leaving a3 as is, is convenient for the purposes of the proof below. a4. neutrality = symmetry under candidate renaming = permuting the candidate names on the ballot rankings permutes their winning probabilities in the same way a5. method is deterministic aside from tiebreaks which (if any) are random [this axiom can be subsumed into the definition of "voting system". It is simplest, but is not actually necessary, for the tiebreaking to give equally likely win probabilities to all co-winners; it actually will suffice for us if the probabilities are fixed and positive for each co-winner in each scenario.] a6. adding a new candidate to the election whom all voters unanimously rank unique-bottom, does not change the winner. Actually, a6 can be dropped too if we are using Campbell-Kelly 2003 to replace a3 with "no vetoer." That is because we shall only use a6 in a 3-candidate situation with an odd number of voters where removal of the always-ranked-last candidate yields 2-candidate simple majority vote; and those situations were already covered by the CK 2003 theorem. Notes about criterion 3 (simple majority rule in 2-candidate case): Actually, this is a consequence of determinism, neutrality, anonymity and positive responsiveness [voting for A top breaks an A-containing perfect tie in favor of A, plus monotonicity] and hence really is NOT NEEDED per se [thanks to Forest Simmons for pointing that out] May's theorem on 2-candidate elections: A group decision function meets 1, 2, and 3 below if and only if it is the simple majority method: 1* It is symmetric under permuting the voters. (anonymity) 2* Reversing each preference reverses the group preference. (neutrality) 3* If the group decision was 0 or 1 and a voter raises a vote from -1 to 0 or 1 or from 0 to 1, then the group decision is 1. (positive responsiveness) [Kenneth O. May: A set of independent necessary and sufficient conditions for simple majority decision, Econometrica 20,4 (1952) 680-684. Mark Fey: May's theorem with an infinite population, Social Choice and Welfare 23,2 (2004) 275-293. T.N.Tideman: A majority-rule characterization with multiple extensions, Social Choice & Welfare 3 (1986) 12-30.] Also, even better, simple majority rule in 2-candidate case is a consequence of AFB (actually strategyproofness, but this is the same as AFB in the 2-candidate rank-equality-forbidden case with odd#voters, and "odd" is the only parity we shall need) and non-dictatorship (which is implied by the nonexistence of a vetoer) and determinism: [Donald E. Campbell & Jerry S.Kelly: A strategy-proofness characterization of majority rule, Economic Theory 22 (2003) 557-568] Both the May and Campbell-Kelly results above have versions that work even if equal rankings are allowed in ballots, incidentally. Also, majority-rule is a consequence of Neutrality, Anonymity, Pareto criterion (if all voters say A>B then B cannot win), odd number of voters, and Independence of Irrelevant Alternatives [E.S.Maskin: Majority rule, social welfare functions, and game forms. In: K.Basu, P.K.Pattanaik, K.Suzumura (eds.) Choice, welfare, and development. Oxford Clarendon Press 1995] a result which was improved by [D.E. Campbell & J.S. Kelly: A simple characterization of majority rule, Economic Theory 15,3 (2000) 689-700]. Two more characterizations, both of which strike me as somewhat silly, are [G.Asan & M.R.Sanver: Another characterization of the majority rule, Economics Letters 75,3 (2002) 409-413.] [G.J.Woeginger: A new characterization of the majority rule, Economics Letters 81,1 (2003) 89-94.] In particular, the Asan-Sanver result actually seems a trivial corollary of the far stronger Smith+Young theorem we'll mention later which was proved over 20 years previously. A&S say Neutrality, Anonymity, Pareto (if all voters say A>B or A=B with at least some saying A>B, then B cannot win), and a partition-consistency property (if both subdistricts say A wins or ties with at least one saying A wins, then A wins in the combined country). More careful definitions of criteria 1 & 2: Clone immunity for the purposes of this proof (and it also is I think common usage) is this. If clones of C are added to the election, that does not affect the winner (except perhaps up to replacement of the winner by a clone). Here "clones" have to be contiguous in all rank-orders (for rank-order voting systems). There can be slight preferences among the clones e.g. some voter prefers C4>C1>C2>C3 but these preferences are assumed to have FAR SMALLER strength than C versus a non-clone of C, e.g. far smaller strength than any comparisons like D>C or C>G or (for that matter) A>B. Therefore range voters will always vote clones almost-equal, to within ε, say, where we will allow ourselves to take the limit ε→0. That is the definition used in Tideman's book, and he invented the whole clone-immunity concept. Under this definition range voting is clone-immune, and so is Schulze-beatpath voting, and so is IRV, but plurality and Copeland and approval voting are not clone-immune (and Tideman's book agrees with all these statements). [Mike Ossipoff with elegant wording: "A clone-set is a set of candidates between whom no one has voted any other candidate(s)."] (Weak) Favorite betrayal: AFB is the criterion that it is never strategically forced for any voter to rank his true favorite, strictly below topmost. (A "strategically forced" move means, if you don't make that move, the election result comes out worse, in terms of expected utility, from your point of view.) Really this all depends on the candidate-utility numbers for that voter, which is something I (badly) did not fully explain in the proof below. To fill in the gaps, one should point out at various places how to construct utility numbers to make it clear that voter-betrayal decisions were indeed strategically forced, otherwise the utility would be worse. But dreaming up appropriate utility numbers for the voter in question, is usually a triviality. Notes about the strategy the proof shall use: The proof will work by demonstrating that, if a single-winner voting system based on pure-rank-order-ballots satisfying axioms 1-6 did exist, then we could deduce a logical contradiction. If any of these properties 1-6 are violated in any election situation, then they are violated by that voting system, period; and we are done. In the proof, a lot of little election situations will be considered, and if in any one of those situations, there is an ICC violation or an AFB violation (with some utility values and some way to alter some vote to betray), then game over and proof done. We are allowed to use the assumption there is no useful way to betray a favorite, or use the assumption ICC is true, or any other axiom, to create new election situations and to deduce the winners in those new situations. (If those deductions were wrong, the proof would be complete because a contradiction would have been found.) Then in these new situations, we can aim to complete the proof by finding some other contradiction. Incidentally, if we alter election 1 to get election 2 and make deductions about elections 1 and 2 using made-up utility values and reasoning about favorite-betrayal, then it is not necessary that the utility values in the two elections be consistent. That is because the voting system acts based solely on the votes and works in ignorance of honest utility values. Proof of the Equivalent Theorem: (We've already discussed why, by results of previous authors, the two theorems are equivalent.) The proof will work by demonstrating that, if a single-winner voting system based on pure-rank-order-ballots satisfying axioms 1-6 did exist, then we could deduce a logical contradiction. So suppose such a system existed. Consider these 3 votes: A>B>C C>A>B B>C>A. By symmetry axiom 4 this is a perfect 3-way tie with win probabilities 1/3, 1/3, 1/3. However we shall argue under axioms 1-3 that A must win, which is a contradiction that establishes the proof. [Really, we shall go through all the 7 possible winners ABC-tie, BC, AC, AB, and A, B, C proving none of these 7 possibiities are allowed by our axioms, which is the sought contradiction. However, we focus on the two cases ABC and A because they remain after the other 5 cases are eliminated, and then we can prove "both are true" which is the requisite contradiction that establishes the proof.] So: If A does not win, then B or C does (or some sort of tie; we'll consider the cases below). If B wins (or if AB tie), then the C>A>B voter would betray C to vote A>C>B getting A>B>C A>C>B B>C>A and then {B,C} is a clone set and hence by axioms 2 and 3 then A must win and hence the betrayal worked and hence we get a contradiction with axiom 1. (In slo-mo that is: the votes are really A>BC, A>BC, BC>A in a 2-candidate election, which A wins by axiom 3, and when we clone BC into two candidates B and C, A still must win by axiom 2.) This betrayal would be a utility improvement from the point of view of that voter if her vote really is "C>A>>B", i.e. if her utility for B is greater than her average utility for {A,B,C}. To see that this betrayal was strategically forced, we also have to note that the alternate dishonest vote (which is not a C-betrayal) C>B>A, would not work since B still would win [here's how we know that: A>B>C C>B>A B>C>A now {B,C} is a clone set, so by axioms 2 and 3 we know B or C must win; but winner here must be B and not C (and not BC tie) because if it were C or BC tie then the A>B>>C-voter could betray: B>C>A causing B>C>A C>B>A B>C>A in which case B must win by axiom 6.] If C wins, or if BC tie, or if ABC tie, or if AC tie, then the A>B>C-voter (whom for this purpose we assume feels A>B>>C) can betray A to vote B>A>C getting B>A>C C>A>B B>C>A whereupon {A,C} is a clone set and hence by axioms 2 and 3 then B must win and hence the betrayal worked (assuming this voter had utilities such that B was valued above the mean utility of A,B,C) and hence we get a contradiction with axiom 1. [The alternate dishonest vote which is not an A-betrayal, A>C>B, would not work since C still would win when A>C>B C>A>B B>C>A because {A,C} is a winning clone set and if A wins (or AC tie) then B>C>>A voter betrays: C>A>B to make C win A>C>B C>A>B C>A>B by axiom 6.] Q.E.D. Remark 1. This theorem is not claiming the 6 criteria are good or bad (although they happen to all sound pretty good to me), and not claiming anybody necessarily should accept or reject them. It simply is claiming that it is logically unachievable to satisfy all of them at once by a rank-order voting system. But (normalized) range voting does satisfy all of them at once. [It is very nice when you can prove every voting system cannot do something - even voting systems nobody has ever invented yet.] Remark 2. Antiplurality voting obeys all 6 axioms except for #2 (and #6). You can make a version of antiplurality voting that obeys #6 by making a last-place-vote count -1, a second-last-vote count -ε, a third-last-vote count -ε², etc in the limit ε→0+, highest score wins. I think there are also an infinite number of other rank-ballot systems avoiding favorite-betrayal, e.g. weighted positional systems depending on the last C-2 rankings only, where C is the number of candidates. Approval Voting obeys all 6 axioms except for #2 (provided we are sufficiently generous about axioms 3 and 6; there is no doubt these are satisfied "in practice"). Remark 3. Schulze beatpaths voting obeys all 6 axioms except for #1. So does Instant Runoff Voting (at least if we are sufficiently generous about the random-tiebreaking axiom 5). Remark 4. Chris Benham suggests the proof might be less confusing (fewer worries about ties etc.) if we modify the initial example to replace the three voters with three identically voting equal-sized large factions and then adding one fickle bullet-voter. 33: A>B>C 33: C>A>B 33: B>C>A 1: ? (A or B or C) In this modified version of "Election 1", we could assume some perturbed symmetry axiom that the lone truncator must determine the winner. This should lead to a slightly different theorem statement with a somewhat simpler proof. Remark 5. Range voting obeys all 6 axioms if all range votes are "normalized" so voters (obeying the recommendations for voting on the front page) always give the best candidate the top score and the worst the bottom score in a 2-candidate election [i.e. in practice with voters who are not idiots]. But with possible-idiot voters, range fails axiom #3 (which does not bother me). CAVEAT: with automatic normalization after the votes are cast, range voting violates axiom 6. Thanks to Markus Schulze for pointing that out. That's embarrassing if our goal is to find a set of criteria range voting obeys but rank-ballot methods fail. Here are some ways to stop being embarrassed: FIX #1: We can either trust range voters not to be idiots in 2-candidate elections (to make axiom 3 hold), or we may, e.g. rephrase axiom 6 as 6'. adding a new candidate to the election whom all voters unanimously rank ε below their previous bottommost does not change the winner in the limit ε→0+. FIX #2 (the one we used in the "Theorem" at top): we can discard axiom 3 by use of characterizations of simple majority rule as described at top (e.g. Campbell-Kelly), then we are free to use unnormalized range voting without having to "trust" voters not to be stupid. This fix is excellent since discarding axioms is always a fine thing. Note that then we do not even need to assume anonymity in the Theorem, aside from (with Campbell-Kelly) a requirement of non-dictatorship. Then we have indeed proven a sense in which range is superior to EVERY pure-rank-ballot voting method, and using two of the most important voting criteria AFB and ICC. Remaining Open question: what happens if we permit rank order votes to have EQUALITIES in them? Are ICC and AFB still incompatible or do they become compatible? I.e. I do not presently know if this theorem can be extended to permit equalities in vote-rankings. (I thank Forest W. Simmons for inspiring me to work on this some more.) Warren D Smith
A day or 2 later, Forest W. Simmons came with his own proof of (his own version of) my same theorem. His theorem also gives a nice set of criteria range voting satisfies but rank-order methods cannot. Here it is [repost from EM with a few edits by WDS to improve clarity]:
Here's a slightly simpler approach for a slightly weaker result. I shall show that (in the case of pure ordinal ballots) you cannot have all three of Monotonicity, Clone Independence, Pareto, and the Strong FBC. To be very careful we explicitly list the assumptions: 1. Strictly ranked ordinal ballots. 2. Neutrality (permuting the candidate names on the ballot rankings permutes their winning probabilities in the same way.) 3. Anonymity (Permuting the ballots does not affect the outcome. [WDS: did not need this.]) 4. Determinism (Chance is not employed except where required by symmetry, i.e in the case of true ties.) 5. Monotonicity - if a voter interchanges two adjacent candidates in her rank order (converting A>B to B>A, say) that cannot reduce A's winning chances. [WDS's proof did not need to assume this] 6. Majority wins when there are only two candidates. (Actually, this property is a consequence of properties one through five above. [WDS: Or of property 8, determinism, and non-dictatorship, by Campbell-Kelly 2003. Also, axioms 1-5 need to be supplemented with positve responsiveness, i.e. voting A top will actually make A win if it was an exact A-involving tie, in May's theorem, BUT ties cannot occur with an odd number of voters, which is the only parity we shall need, so this supplementation is not necessary for us.] But we mention it explicitly for convenience.) 7. ICC. The probabilities assigned by the method to members of a clone set must add up to the probability that the method would give to a single alternative replacing the clone set on each ballot. 8. Strong FBC. To each strategy that allows a faction to achieve a certain expected level of utility, there is another strategy that allows that faction to rank favorite over all other candidates without sacrificing that expected level of utility. The proof: Suppose that we have three voters with ranked preferences of 1 A>B>>C 1 B>>C>A 1 C>>A>B Their nearest expression in the pure ordinal ballots of our method would be 1 A>B>C 1 B>C>A 1 C>A>B . Properties one thru four demand that each alternative be given equal probability P(A)=P(B)=P(C)= (1/3) . The A>B>C voter's expectation is below their utility of their compromise B, while the other two voters' compromises have less utility than their expected utilities. So a sure win by B would be an improvement for the first two voters. The first voter can accomplish this unilaterally by burying favorite: 1 B>C>A 1 B>C>A 1 C>A>B Properties 6, and 7 (applied to the clone set {C, A}) make B the winner with 100 percent probability under our method with this ballot set. Now property 8 says the first voter can do as well without moving A downward from the top. With that constraint all she can do is switch B and C: 1 A>C>B 1 B>C>A 1 C>A>B For future reference we will call this ballot set S. Here the {A,C} clone set has a majority over B, so by property 6, all of the probability goes to {A,C}. And since there is no symmetry that would require chance, either A or C gets 100 percent of the probability by our axiom of determinism. We know that C is not the one with 100 percent probability because the FBC guarantees that our result will be no worse than B. So our method gives the win to A, given this ballot set. Now here's the tricky part: We change this last ballot set S into the following one S' by two different means: (1) raising A over C on the middle ballot. OR (2) swapping A with C on every ballot and then swapping the first and last ballot. 1 A>C>B 1 B>A>C 1 C>A>B Monotonicity ensures that A is still the winner, since all we did was raise A relative to C on one ballot (at least from the first point of view). But from the point of view of the second way of getting from S to S', neutrality and anonymity require that the win switch from A to C. This contradiction shows that the stated conditions are incompatible. Q.E.D.
Then, I am presently unable to settle the question of whether AFB and ICC are incompatible. However, I can prove ICC is incompatible with an inequivalent version of AFB (which I do not preferentially endorse) call it AFB':
AFB' = "Raising favorite to top rank must not decrease expected utility."AFB' is easier to work with than AFB, but I regard it as of less interest. AFB'⇒AFB but the reverse implication does not hold. Here are the details:
Theorem: AFB', ICC, neutrality, the assumption that in a perfect 3-way tie you break ties randomly with all tiers getting nonzero win probabilities, and finally, reduction to simple majority vote in the 2-candidate case (which assumption again can be replaced by simpler ones), are logically incompatible in any single-winner election method based on rank-order ballots with rank-equalities permitted. Proof: Begin with the election 3 A=B>C 3 C=A>B 3 B=C>A 2 A>C>B 2 B>A>C 2 C>B>A By symmetry this is a perfect 3-way tie. Now suppose the three A=B>C voters all betray their co-favorite B to get 3 A>B>C 3 C=A>B 3 B=C>A 2 A>C>B 2 B>A>C 2 C>B>A and then the B>A>C voters (regarded as B>A>>C), all betray their solo favorite B, to get 3 A>B>C 3 A=C>B 3 B=C>A 2 A>C>B 2 A>B>C 2 C>B>A and finally the A=C>B voters betray their co-favorite C to get 3 A>B>C * 3 A>C>B * 3 B=C>A 2 A>C>B 2 A>B>C * 2 C>B>A where the *s indicate dishonest votes. A wins in this scenario with 100% probability (by ICC and 2-candidate majority using the clone set {B,C}). So the net effect of these 3+2+3=8 betrayal decisions, was to cause a result all 8 of the betrayer-voters prefer. The betrayals worked. Now consider these 8 betrayers changing their votes to the betrayal-vote one by one. At the start-end of the 8-betrayal chain, ABC tie. At the finish-end, A wins (a result all betrayer-types prefer). So at some point in the chain, there must have been a beneficial (to that betrayer) election-result-change. This proves a single betrayal must work in some election situation. However... it remains possible that some non-betraying dishonest vote also works. But with the weakened definition AFB' of AFB that is not an issue. Q.E.D.
Ingenious! That settles it for practical purposes, because voters are not sophisticated enough to take advantage of AFB when AFB' fails. [WDS: I am not as happy as Forest. But we shall let him continue...]
AFB says that you can remain loyal to favorite if you are clever enough. AFB' says that remaining loyal to favorite will never mess things up, even if you are not too clever. And that's what people want.
Roughly speaking the two main reasons that people are interested in replacing Plurality with something else are
Finally I note that the faction sizes 3,3,3,2,2,2 could more simply have been 2,2,2,1,1,1 because 2+2+1 is greater than half of 2+2+2+1+1+1.
Yes. The property-sets mentioned above are particularly appealing because AFB and ICC seem very important properties to make democracy work. But other property sets also work and may have some appeal.
Theorem. These five properties are incompatible in a single-winner voting system based on pure-rank-order ballots: 1. partition-consistency. (That is, if X wins in district 1 and in district 2, then X must win in the combined 2-district country.) 2. AFB 3. "responsiveness at top" - raising a candidate from second-top to top in your vote (by swapping those two positions) in some election situations actually helps him. 4. anonymity 5. neutrality Proof sketch. The hard part was done by [John H. Smith: Aggregation of preferences with variable electorate, Econometrica 41,6 (1973) 1027-1041. H. Peyton Young: Social choice scoring functions, SIAM J. Appl. Math. 28,4 (1975) 824-838.] These authors independently showed that any rank-ballot system obeying axioms 1, 4, and 5 had to be a "composition of weighted positional systems." But then that is readily seen to be incompatible with properties 2 and 3 (e.g. by considering a suitable cyclic-tie situation). Q.E.D. Theorem. These four properties are incompatible in a single-winner voting system based on pure-rank-order ballots: 1. three-candidate semi-honesty. (That is, in a 3-candidate election, it is never strategically forced to dishonestly vote as though X>Y when your honest view is Y>X.) 2. determinism. 3. no dictator. 4. unanimously top-ranked candidate must win. Proof sketch. This is an immediate consequence of the Gibbard-Satterthwaite strategyproofness theorem. Q.E.D. Theorem. These three properties are incompatible in a single-winner voting system based on pure-rank-order ballots: 1. no dictator. 2. If every voter prefers A to B then so does the group. (This implicitly assumes the voting system outputs not only a winner but also a rank-order of all finishers. You can get a rank order even from a system without one by ranking the winner first, then delete that candidate from all ballots and ask who would have won then - rank him "second" - and so on.) 3. "independence of irrelevant alternatives": The relative positions of A and B in the group ranking depend on their relative positions in the individual rankings, but do not depend on the individual rankings of any "irrelevant alternative," i.e. other candidate, C; to word it more precisely, we shall demand that if C is deleted from all ballots, then whether A finishes ahead of or behind B, is unaffected. Proof sketch. This is an immediate consequence of K.Arrow's impossibility theorem. Q.E.D. Note. Range voting obeys all the properties in all of the above theorems (at least if simple majority in the 2-candidate case is dealt with appropriately, cf. remark 5 above). Thus each one demonstrates a way in which range voting is superior to all rank-ballot voting methods.
Another theorem: Every single-winner voting system based on rank-order ballots (with equalities either forbidden or permitted – both work) must suffer from at least one of the following paradoxes:
Proof: This claim was made in P.C.Fishburn & S.J.Brams: Paradoxes of Preferential Voting, Mathematics Magazine 56, 4 (Sept. 1983) 207-214. Actually, just criteria 1 and 2 alone are incompatible, and that is true whether or not rank-equalities are permitted; a proof, due to Markus Schulze and really dating back to Herve Moulin, is found in my paper #79 here or on this page. Q.E.D.
Range voting, however, avoids all of these paradoxes (at least, as they are worded here).