Notes on the Dutch Book Argument

by D. A. Freedman, Statistics Department, U. C. Berkeley, CA 94720

The object here is to sketch the mathematics behind de Finetti’s (1931, 1937)
argument for the Bayesian position. Suppose a bookie sets odds on all subsets of
a set, accepting bets in any amount (positive or negative) on any combination of
subsets. Unless the odds are computed from a prior probability, dutch book can
be made: for some system of bets, the clever gambler wins a dollar or more, no
matter what the outcome may be. The extension by Freedman and Purves (1969)
to statistical inference is also considered. Finally, there is a dutch-book argument
for countable additivity.

1. De Finetti’s argument
Let
be a finite set with card (
) > 1. A bookie posts odds λA on A, for
every proper A ⊂
; odds are positive and finite. If you bet bA on A, and A occurs,
then you win bA /λA ; if A does not occur, then you win −bA . Your net payoff is

bA
 1 + λA
λA 
(1) φA = 1A − 1 − 1A bA = bA 1A − .
λA λA 1 + λA

The stakes bA are finite, but may be positive, 0, or negative. Corresponding to each
set of stakes there is a payoff function,

(2) φA ,
A

where A runs over the proper subsets of
. For now, we take the odds as fixed, and
consider various gamblers who bet against the bookie: each gambler generates a
payoff function.
If λA = π(A)/[1 − π(A)] for some probability π on
, i.e.,

π(A) = λA /(1 + λA ),

we say the bookie is a Bayesian with prior π . Obviously, all payoff functions then
have expectation 0 relative to π. In particular,
Proposition 1. Dutch book cannot be made against a Bayesian bookie.

1
 Let V be the set of all real-valued functions on
, so V is a linear space of
dimension card (
).
Theorem 1. For each bookie, there are only two possibilities: (i) the set of
payoff functions coincides with V ; or (ii) the bookie is a Bayesian with prior π ,
where π is a probability on
assigning positive mass to every ω ∈
.
Proof. If case (i) does not obtain, there is a non-trivial function π on
such
that for all φA , 
π(ω)φA (ω) = 0,
ω∈

which is to say, by (1),

π(A) = π(
)λA /(1 + λA ),

where 
π(B) = π(ω)
ω∈B

for all B ⊂
. If π(
) = 0 then π(ω) = 0 for all ω, as one sees by taking
A = {ω}. This is a contradiction, so π(
) = 0. Renormalize π so that π(
) = 1,
i.e., replace π by π/π(
). Now π(A) = λA /(1 + λA ), so π > 0 and case (ii)
obtains.
Theorem 2. For the Bayesian bookie with prior π , the set of payoff functions
coincides with the set of functions having expectation 0 relative to π .
Proof. Otherwise, there is a non-trivial ν, not a multiple of π, such that all
payoff functions are orthogonal to ν as well as π . Arguing as before, we find that
ν(
) = 0. We renormalize so ν(
) = 1, and then ν(A) = λA /(1 + λA ) = π(A),
a contradiction.
Corollary 1. The following are equivalent.
(i) The payoff functions are all of V .
(ii) The bookie is not a Bayesian.
(iii) Dutch book can be made against the bookie.
Proof. (i) ⇒ (iii): obvious. (iii) ⇒ (ii): Proposition 1. (ii) ⇒ (i): if the
payoff functions were not all of V , the bookie would be a Bayesian (Theorem 1).

2
 Corollary 2. The following are equivalent.
(i) The payoff functions are not all of V .
(ii) The bookie is a Bayesian.
(iii) Dutch book cannot be made against the bookie.
In this case, the payoff functions consist of all functions with expectation 0, relative
to the bookie’s prior.
Proof. (i) ⇒ (ii): Theorem 1. (ii) ⇒ (iii): Proposition 1. (iii) ⇒ (i): obvious.
The final assertion follows from Theorem 2.
Example 1. Negative stakes are needed for the theorems. Suppose
=
{0, 1}, the bookie puts 2:1 on 0 and 3:1 on 1, but accepts only positive bets on
events. If you put x ≥ 0 on 0 and y ≥ 0 on 1, the payoff is x/2 − y on 0 and
y/3 − x on 1. If both are positive, x/2 > y > 3x, so x > 6x, a contradiction.
This bookie – although not a Bayesian – is immune from dutch book, being clever
enough to set favorable odds and require positive stakes.

2. The extension to statistical inference

In this section, we discuss the principal result in Freedman and Purves (1969).
In essence, a bookie has to post odds on subsets of a parameter space  after
seeing an observation x drawn from p(• |θ): dutch book can be made against the
non-Bayesian bookie. In view of the previous section, the bookie will use an
‘estimating probability’ q(• |x) on  to set the odds. The only question is, how do
these probabilities fit together?
A ‘finite estimation problem’ consists of (i) a finite set X , and (ii) a finite set
of parametric models {p(• |θ) : θ ∈ } specifying probability distributions on X.
Allowing p(x|θ) = 0 creates one technical nuisance after another, so we assume
(3) p(x|θ) > 0 for all x ∈ X and θ ∈ .
An ‘estimating probability’ q(• |x) is a probability on  for each x ∈ X.
Consider subsets C1 , . . . , Ck of . After x is observed, allow the gambler to pay
bi (x)q(Ci |x) in order to get bi (x) dollars if θ ∈ Ci . The gambler is allowed to use
any bounded bi . Bets are settled separately, and then summed.
The net payoff to a gambler who uses the sets {C1 , . . . , Ck } and the functions
{bi : i = 1, . . . , k} is
k
  
(4) φ(x, θ) = bi (x) ICi (θ) − q(Ci |x) .
i=1

3
Any such φ is called a ‘payoff function’. Nothing requires q to be positive: if
q(Ci |x) = 0, the gambler pays 0 to get bi (x) if θ ∈ Ci . Similarly, Ci may be ∅
or . In these respects, payoff functions work more smoothly than odds.
Corresponding to each payoff function φ on X × , there is an ‘expected
payoff function’ on :

(5) Eθ {φ} = φ(x, θ)p(x|θ).
x∈X

Definition 1. Dutch book can be made against the estimating probability

q(• |x) if there is a gambling system that provides a uniformly positive expected
payoff to the gambler: in other words, there exists a payoff function φ – as defined
by (4) – and an  > 0 such that

(6) Eθ {φ} >  for all θ ∈ .

To paraphrase Freedman and Purves (1969),

Imagine a Master of Ceremonies who chooses some θ ∈  and then picks
x ∈ X at random from p(• |θ). The value of x is revealed and the bookie
announces the estimating probability q(• |x). The gambler then constructs a
system with payoff function φ. When (6) holds, the gambler expects to win
at least  no matter what the value of θ.
Of course, if (6) holds for some positive , any other positive  can be obtained by
rescaling the payoff function.
If π is a probability on , we will say that the bookie is a Bayesian with
prior π provided

(7) mπ (x)q(θ |x) = π(θ)p(x|θ)

where

(8) mπ (x) = π(θ)p(x|θ).
θ∈

Here, mπ (x) is the ‘marginal’ probability of x, integrated over the θ’s.

Lemma 1. For the Bayesian bookie with prior π , any expected payoff func-
tion – as a function of θ – integrates to 0 against π .

4
 
Proof. In (4), take k = 1; write b for b1 and C for C1 . Then θ Eθ {ψ}π(θ )
is   
b(x) 1C (θ) − q(C|x) p(x|θ)π(θ) = α − β
θ ∈ x∈X
where 
α= b(x)1C (θ)p(x|θ)π(θ)
θ ∈ x∈X
and 
β= b(x)q(C|x)p(x|θ)π(θ) = α
θ ∈ x∈X
by ‘the law of total probability’. In more detail,

β= b(x)q(C|x)p(x|θ)π(θ)
θ ∈ x∈X

= b(x)q(C|x)p(x|θ)π(θ)
x∈X θ ∈

= b(x)q(C|x)mπ (x)
x∈X

= b(x)π(θ)p(x|θ) by (7)
x∈X θ ∈C

= b(x)1C (θ)p(x|θ)π(θ) = α.
θ ∈ x∈X

Remark. Another way to say Lemma 1: the payoff functions φ in (4) all have
expected value 0, relative to the probability on X× that assigns mass p(x|θ)π(θ)
to the pair (x, θ).
Corollary 3. Dutch book cannot be made against a Bayesian bookie.
Let V be the set of all real-valued functions on , so V is a linear space of
dimension card ().
Theorem 3. For each bookie, there are only two possibilities: (i) the set of
expected payoff functions coincides with V ; or (ii) the bookie is a Bayesian with
prior π, where π is a probability on .
Proof. If case (i) does not hold, there is a non-trivial function π on , or-
thogonal to all expected payoff functions. For the moment, fix x ∈ X and θ ∈ .

5
Specialize φ in (4), choosing k = 1, b1 = 1 at x and b1 = 0 elsewhere, C1 = {θ }.
Then E• {φ} ⊥ π unpacks to

(9) mπ (x)q(θ|x) = π(θ)p(x|θ).

Equation (9) holds for all x and θ . In view of (3), it is not possible for π to be strictly
positive at θ1 and strictly negative at θ2 . Nor can π vanish identically. Hence, we
can renormalize π to be a probability on : then (9) says that the bookie is a
Bayesian with prior π.
Theorem 4. For the Bayesian bookie with prior π , the set of payoff functions
coincides with the set of functions having expectation 0 relative to π .
Proof. Otherwise, we find a non-trivial function ν on , not a multiple of π ,
orthogonal to all expected payoff functions. Arguing as before, we show that (9)
holds with ν in place of π. Hence, ν can be renormalized to a probability. Let
µ = (π + ν)/2, and + = {θ : θ ∈  & µ(θ) > 0}. If θ ∈ + , by (9),

(10) ν(θ)/µ(θ) = mν (x)/mµ (x).

The right hand side of (10) does not depend on θ . Therefore, ν = µ. Similarly,
π = µ = ν, a contradiction proving Theorem 4.
The corollaries follow, as in the previous section.
Corollary 4. The following are equivalent.
(i) The expected payoff functions are all of V .
(ii) The bookie is not a Bayesian.
(iii) Dutch book can be made against the bookie.
Corollary 5. The following are equivalent.
(i) The expected payoff functions are not all of V .
(ii) The bookie is a Bayesian.
(iii) Dutch book cannot be made against the bookie.
In this case, the expected payoff functions consist of all functions with expectation 0,
relative to the bookie’s prior.

6
3. From finite to countable additivity
We modify de Finetti’s argument (Section 1) to favor countable additivity
rather than finite additivity. Let (
, ) be a probability space, finite or infinite;
let ψ be a function from  to [0, 1]. In general, we require  to be a σ -field. A
‘payoff function’ is

(11) cA [1A − ψ(A)]
A

where A ∈  and cA is a finite real number: the sum is over a finite number of
A’s. Let F be the set of payoff functions. The normalization is a little different
from (1): when 0 < λA < ∞, ψ(A) = λA /(1 + λA ) and cA = bA (1 + λA )/λA .
The imagery is different too. Informally, the gambler pays ψA for a lottery ticket
1A , which yields $1 if A occurs and 0 otherwise: the bet is just enough to get $1 if A
happens. This bet can be scaled by cA , an arbitrary real number. Any finite number
. .
of bets can be placed. When λA is large, i.e., ψ(A) = 1, then cA = bA . When λA
. .
is small, i.e., ψ(A) = 0, then cA = bA /ψ(A). The advantage of (11) is that the
formula can be used even if ψ(A) = 0 or 1, corresponding to λA = 0 or ∞.
To avoid dutch book, ψ must be a finitely additive probability on . When

is finite, this can be proved just as before: the new proposition is a little more
general, since the odds are unrestricted. As a matter of notation, the previous
argument generates a (finitely additive) probability π, and identifies ψ with π.
The result is easily extended to the case where
is infinite but  finite – and from
there, to the general case.
Suppose next we allow the sum in (11) to be countably infinite, so long as there
is uniformly bounded pointwise convergence. More exactly, let {Ai : i = 1, 2, . . .}
be a sequence of subsets of  and {ci } a sequence of real numbers. Let
n

φn = ci [1Ai − ψ(Ai )].
i=1

Suppose
(i) |φn (ω)| < L for all n = 1, 2, . . . and all ω ∈
, where L is a real number;
(ii) limn φn (ω) = φ(ω) for each ω ∈
.
Then φ is also considered to be a payoff function. Said another way, the set of
payoff functions is enlarged by taking uniformly bounded pointwise limits. Let F
denote this larger set of payoff functions.

7
 Now, unless ψ is countably additive, a dutch book can be made. ∞ Indeed, let
Ai ∈  be pairwise disjoint and A = ∪i Ai . Of course, ψ(A) ≥ i=1 ψ(Ai ).
Suppose that the inequality is strict. Consider a payoff function that has a bet on
each Ai , and a bet against A, to win $1 in each case:
∞
  n
 
[ψ(A) − 1A ] + [1Ai − ψ(Ai )] = lim [ψ(A) − 1A ] + [1Ai − ψ(Ai )]
n
i=1 i=1
∞

= ψ(A) − ψ(Ai ) > 0.
i=1

In short, finite additivity exposes the odds-maker to a dutch book – if the gambler
can use payoff functions in F . Of course, if the odds are computed from a countably
additive π, each payoff function has expectation 0 relative to π, by the dominated
convergence theorem, and dutch book is impossible.
Theorem 5. Suppose the set F of payoff functions is initially defined by (11),
with A ∈ , a σ -field of subsets of
; only finitely many A’s are allowed. The set
of payoff functions is then enlarged to F by allowing uniformly bounded pointwise
limits. A gambler can make dutch book unless the odds are compatible with a
countably additive prior on (
, ).
Example 2. Suppose ψ is a remote (finitely additive) probability on the
integers. The gambler pays 0 to win $1 if integer j materializes, for all integers.
The odds-maker will in the end have to fork over $1.

A primer on odds
If the odds on A are 3:1, and we bet $1 on A:
if A occurs we win $1/3;
if A does not occur, we lose our $1.
If the odds on A are 3:1, and we bet $1 against A:
if A occurs we lose our $1;
if A does not occur, we win $3.
If the odds against B are 3:1, and we bet $1 on B:
if B occurs we win $3;
if B does not occur, we lose our $1.

8
If the odds against B are 3:1, and we bet $1 against B:
if B occurs we lose our $1;
if B does not occur, we win $1/3.
If P (A) = 3/4—
the odds on A are 3:1
the odds against A are 1:3
the odds on Ac are 1:3
the odds against Ac are 3:1

References
de Finetti, B. (1931). Sul significato soggestivo della probabilità. Fundamenta
Mathematicae 17: 298–329.
de Finetti, B. (1937). La prévision: ses lois logiques ses sources subjectives. Ann.
Inst. H. Poincaré 7: 1–68. English translation in H. E. Kyburg, H. E. Smokler (eds),
Studies in Subjective Probability. New York: Wiley, 1964.
Eaton, M. L. and Freedman, D. A. (2003). Dutch book against ‘objective’ priors.
Technical Report #642, Department of Statistics, U. C. Berkeley.
http://www.stat.berkeley.edu/users/census/642.pdf
Freedman, D. A. (1995). Some issues in the foundations of statistics. Foundations
of Science 1: 19–83, with discussion. Reprinted in B. van Fraassen (ed), Topics in
the Foundation of Statistics. Dordrecht: Kluwer, 1996.
Freedman, D. A. and Purves, R. A. (1969). Bayes method for bookies. Ann. Math
Statist. 40: 1177–1186.
Ramsey, F. (1926). Truth and Probability. Reprinted in H. E. Kyburg and H. E.
Smokler (eds), Studies in Subjective Probability, pp. 25–52. Huntington, NewYork:
Krieger Publishing, 1980.

15 July 2003