A calculating machine with an abyss in one cell

Infinity-Nothing

Pascal's wager is usually remembered as a shortcut to belief. The more interesting object is the machinery underneath it: an early, influential decision under uncertainty. Pascal tries finite prizes before the infinite one. The finite arithmetic works. Then one symbol enters the matrix and its power to rank strategies comes apart.

choice / world
reward exists
reward absent
wager
+3gain
-1cost
do not wager
0baseline
0baseline

EV(wager) = +1 life; break-even p* = 25%

At even odds: one half times 3 minus one half times 1 equals +1. In Pascal's finite sequence, this is the point where he says you would have to play.

q = 0.000000001; mixed EV = infinity

Under this simple infinite-utility matrix, any positive q inherits infinite expectation. So pure wagering and even a fantastically unlikely route to wagering both read infinity. Standard expected utility no longer orders them. Hájek's 2003 objection has replies; the current SEP describes arguments that it is formal rather than substantive.

Rival column added. Diderot's 1746 challenge was brief and blunt (in the standard English rendering): “An Imam could reason just as well this way.” Once incompatible rewards enter, the original two-state dominance claim needs further premises. This many-gods objection also has published replies; the button displays an objection, not a verdict.

The fragment is Infini-rien, numbered Lafuma 418, Sellier 680, and Brunschvicg 233. The numbering matters because Pensées is a posthumous 1670 collection of notes, not Pascal's finished treatise. Hacking 1972 finds three arguments in the passage. The Stanford Encyclopedia presents the nearby forms as superdominance, expectation, and generalized expectations. Hacking 1975 calls the wager “the first well-understood contribution to decision theory,” a calibrated historical description, not a claim that Pascal invented decision theory or probability.

The finite machine survives

Turn infinity off and the threshold is ordinary algebra: p* = S / (G + S). With one life staked, gains of one, two, and three lives at even odds produce expected values of 0, +0.5, and +1. That is the finite sequence in the fragment. Infinity is a limit here, not a large number: as finite G grows, the threshold approaches zero but never reaches it.

The finite cap: Pascal's mugging territory

G = 1,000,000; p* = 0.0000999999%

Bostrom's 2009 Pascal's mugging asks how expected-value reasoning should react when an implausible claim is paired with a sufficiently vast finite payoff. This slider does not solve that problem. It exposes the pressure: no matter how tiny a positive probability you choose, a large enough finite prize can dominate. Proposed solutions exist and remain debated.

The same finite skeleton, with real-world caveats

Insurance uses expected loss in pricing and risk analysis, but buying insurance is not explained by expected value alone. Risk aversion, premiums, exclusions, solvency, regulation, and the declining marginal value of money matter. This abstract example is deliberately not a real quote.

Expected loss = $1,000; expected net benefit of policy = +$100

All four inputs are free choices. The calculation is p x covered loss - premium. It omits risk preference and every policy detail, so it demonstrates the finite arithmetic only.

The check

Live now: p = 0.5, G = 3, S = 1. EV = 0.5 x 3 - 0.5 x 1 = +1; p* = 1 / (3 + 1) = 0.25.

EV = pG - (1-p)S p* = S/(G+S) finite mixed EV = q x EV if G is infinite and p,q are positive: pG = q(pG) = infinity

Proven by the bench: the finite expected values and break-even threshold follow from the selected inputs. Under the displayed extended-real convention, positive probability times positive infinity is infinity, so the simple matrix ties every positive-path strategy.

Assumed or disputed: the probability of the reward, the utility of salvation, whether utility may be infinite, whether belief can be chosen, the cost S, the available religions, and the completeness of the state space. Pascal says the this-life loss is nothing; the slider does not assume that. Mixed-strategy and many-gods objections have published replies.

Sources and honest apparatus

Could not independently inspect a freely accessible scan of Hacking 1972 or Hájek 2003 during this build. Their bibliographic details and the claims attributed to them were cross-checked through the current Stanford Encyclopedia entry. The Pascal wording varies by translation, so only the short decision phrases needed here are paraphrased.