Supplement to Epistemic Utility Arguments for Probabilism

Proof of Theorem 7

We wish to prove the following theorem (Joyce 1998):

Theorem 7. Suppose G satisfies:

  • Strongly Non-Trivial,
  • Proposition-Wise Continuity,
  • Unbounded,
  • Truth-Directed,
  • Strong Convexity,
  • Symmetry, and
  • Dominating Compromise.

Then:

  1. For every non-probabilistic b in BP, there is a probabilistic p in P that strongly dominates it.
  2. For every probabilistic p in P, there is no credence function b in B that weakly dominates it.

We do this by proving the following two lemmas:

Lemma 1 (de Finetti) P is the convex hull V+ of V. That is, P is the smallest set that (i) contains each valuation function v in V and (ii) contains λb + (1-λ)b whenever it contains b and b. In other words, if U has properties (i) and (ii), then V+U.

Lemma 2 (Joyce) Suppose G satisfies

  • Strongly Non-Trivial,
  • Proposition-Wise Continuity,
  • Unbounded,
  • Truth-Directed,
  • Strong Convexity,
  • Symmetry, and
  • Dominating Compromise.

Then:

  1. If bBV+, then there is p in V+ such that G(b, v) > G(p, v) for all v in V.
  2. If pV+, then there is no b in B such that G(p, v) ≥ G(b, v) for all v in V.

Proof of Lemma 1

We prove this in two stages: first we prove V+P and then we prove PV+.

  1. V+P. To prove this, we need only show that (i) VP and (ii) P contains λb + (1-λ)b whenever it contains b and b. It is straightforward to verify that every valuation function is a probability function. After all, the classical truth value of a disjunction AB of mutually exclusive propositions A and B is obtained by adding together the truth values of the disjuncts A and B. (ii) is straightforward arithmetic.
  2. PV+. To prove this, we must show that each p in P is a convex combination of the elements of V. That is, for each p in P, there is 0 ≤ λv ≤ 1 for each v in V such that Σ λv = 1 and p = Σ λvv. If this is true, then each probability function must belong to the convex hull V+ of V. To prove this, we use the notation ΣA to denote the sum over all worlds v that make A true. Σ without a subscript denotes the sum over all worlds v in V. Suppose p in P. And suppose A is a proposition in F. Then note that A is equivalent to the disjunction of all the atomic propositions v of F that entail A; and these atoms correspond to the valuation functions that make A true. Thus, since probability functions are finitely additive, for each proposition A in F,
    p(A)
      =  ΣA p(v)
      =  ΣA 0.p(v) + ΣA 1.p(v)
      =  Σp(v)v(A).

    Thus, p = Σp(v)v, as required.

This completes our proof.

Proof of Lemma 2

Suppose G satisfies Strongly Non-Trivial, Proposition-Wise Continuity, Unbounded, Truth-Directed, Strong Convexity, Symmetry, and Dominating Compromise. We begin by extending G from a measure of the distance between a credence function b and a valuation function v to a measure of the distance between any two credence functions b and b. We define D : B × B → ℜ in the following way:

D(b, c) := G(v + (bc), v)

for v in V. Since G is Truth-Directed, we have G(b, v) = G(c, v), if |b(A) − v(A)| = |c(A) − v(A)| for all A in F. From this, we can derive the following facts about our new distance measure D:

  • The definition of D does not depend on the valuation v in V used in its definition. That is, G(v + (bc), v) = G(v′ + (bc), v′) for all v, v′ in V.
  • D extends G. That is, for b in B and v in V, G(b, v) = D(b, v).
  • D is symmetric. That is, for all b, c in B, D(b, c) = D(c, b).

With the definition of D in hand, we can prove Lemma 2(1) and Lemma 2(2).

We begin with Lemma 2(1). We wish to show that each b in BV+ is strongly dominated by some p in V+. Thus, we let b be a credence function in BV+. And we consider the function D(b, •) : V+ → ℜ. This function takes a member of V+ and gives its distance from b. Now, it follows from Proposition-Wise Continuity that D(b, •) is continuous with respect to the Euclidean metric. Thus, we have that D(b, •) is a real-valued, continuous function on a closed, bounded subset of ℜn. The Extreme Value Theorem says that, for any such function, there is at least one member of its domain at which it takes a minimum value. That is, there is a p in V+ such that D(b, p) ≤ D(b, c), for all c in V+. In fact, by Strong Convexity and Symmetry, we can show that this minimum must be unique. If it weren't, we would have D(b, p), D(b, p′) ≤ D(b, c) for all c in V+. But then by Strong Convexity and Symmetry, we would have D(b, ½p + ½p′) < D(b, p), D(b, p′), which gives a contradiction, since ½p + ½p′ is in V+. In sum: for each b in BV+ there is a unique closest member of V+. We denote it p. Our next task is to show that, for all v in V, D(p, v) < D(b, v), and thus G(p, v) < G(b, v). That is, we must show that b is further than p from each valuation function v.

Suppose vV. If p = v, then by Strongly Non-Trivial, we have D(p, v) < D(b, v), and we're done. Thus, we suppose that pv. Let

R := {λp + (1-λ)v : -∞ < λ < ∞}

Thus, R is the straight line that passes through p and v and passes to infinity in both directions.

We now prove that there is some m that lies on the line R such that

  1. D(m, v) ≥ D(p, v)
  2. D(b, v D(m, v)

If we can show this, then we can put (1) and (2) together to give D(p, v) < D(b, v), as required. Consider the function D(b, •) on R. Again, this is a continuous, real-valued function on R. And, by Unbounded, D tends to infinity as we move away from v towards p and beyond. But since p is closest to b of all members of V+, it is closer to b than v is. That is, D(b, p) < D(b, v). Thus, the Intermediate Value Theorem gives us that there is a credence function k in R that lies beyond p on the line from v through p to infinity that is exactly the same distance from b and v is. That is, D(b, k) = D(b, v). Then, by Symmetry and Strong Convexity, ½k + ½v is the unique minimum of D(b, •) on the segment kv of this line. Let m be ½k + ½v. So m is the credence function that lies on the segment kv that is closest to b. We now prove that m satisfies (1) and (2) from above:

  • If m = p, then certainly D(m, v) ≥ D(p, v). Thus, suppose mp. First, we show that p lies on the line segment mv. Suppose not. Then m must lie on the line segment pv. But this line segment lies entirely inside V+. Thus, m is in V+ and it is closer to b than p is. But recall that p is the credence function in V+ that is closest to b. So we have a contradiction. Thus, p lies on the line segment mv. Thus, by Truth-Directed, we have that m is at least as far from v as p is. That is, D(m, v) ≥ D(p, v). Thus, (1).
  • By the definition of k, we have D(b, k) = D(b, v). A little calculation shows that, by Truth-Directed, we have D(b, k) = D(b, 2bk). Thus, we have D(b, v) = D(b, 2bk). And thus by Strong Convexity and Symmetry, it follows that D(b, •) obtains a unique minimum on the line segment v(2b-k) at ½(vk) + b. Thus, in particular, D(b, v) > D(b, ½(v - k) + b). But, again by Truth-Directed, from the definition of m we have D(b, ½(wk) + b) = D(m, v). Thus, (2).

As already noted, putting (1) and (2) together gives us Lemma 2(1).

Next, we turn to Lemma 2(2). This time we must show that no credence function in V+ is weakly dominated by any other credence function in B. In fact, we show something a little stronger. We show that, for any p in V+ and any distinct b in B, there is v in V such that G(p, v) < G(b, v). That is, by moving from a credence function in V+ to any alternative credence function, our agent will become less accurate at some world. Suppose, for a reductio, that p in V+ and b in B and, for all v in V, G(b, v) ≤ G(p, v). Thus, D(b, v) ≤ D(p, v). Then define the following subset AV+ as follows:

A := {c in V+ : D(b, c) ≤ D(p, c)}

Thus, A is the set of credence functions that are either (i) equidistant from b and p or (ii) closer to b than to p.

We will show that A is a convex set that contains all valuation functions v in V, but does not contain p. Thus, we have that A is a convex set that includes V but is a proper subset of V+. This contradicts our assumption that V+ is the convex hull of V. This gives us our conclusion.
  1. VA, by the assumption from which we are trying to derive a contradiction.
  2. pA, since by Non-Triviality, we have D(b, p) > D(p, p). Thus, AV+.
  3. A is convex, since by Dominating Compromise, we have: for all b, b′, c, c′ in B, if D(b, c) ≤ D(b′, c) and D(b, c′) ≤ D(b′, c′), then for all 0 ≤ λ ≤ 1,
    D(b, λc + (1-λ)c′) ≤ D(b′, λc + (1-λ)c′)

This gives us our contradiction from which we infer Lemma 2(2).

Copyright © 2011 by
Richard Pettigrew <Richard.Pettigrew@bris.ac.uk>

Open access to the SEP is made possible by a world-wide funding initiative.
Please Read How You Can Help Keep the Encyclopedia Free


The SEP would like to congratulate the National Endowment for the Humanities on its 50th anniversary and express our indebtedness for the five generous grants it awarded our project from 1997 to 2007. Readers who have benefited from the SEP are encouraged to examine the NEH’s anniversary page and, if inspired to do so, send a testimonial to neh50@neh.gov.