Yamagishi Complexity Torus LabYamagishi Lab

[Paper 629] A Game with Infinite Expectation Has an Expectation of About $41$ Once the Bank Is Limited to $10^{12}$

Abstract

A coin is tossed until heads, and $2^k$ is paid if the first head comes on toss $k$ (the St. Petersburg game). Its expectation is infinite. What becomes of that infinity when something is made finite is counted. No new theorem or law is claimed.

First, the expectation is infinite——the first head comes on toss $k$ with probability $1/2^k$ and pays $2^k$, so every $k$ adds exactly $1$ to the expectation. The terms never run out, so the sum grows without bound (Section 2).

Second, limiting the bank to $B$ brings the expectation to about $\log_2 B$——if only $B$ can be paid once $2^k$ exceeds $B$, the expectation is $K+B/2^K$ (with $K$ the largest integer for which $2^K$ does not exceed $B$), and even with $B=10^{12}$ it is $\mathbf{40.818989}$. Up to toss $39$ each adds $1$, and everything beyond adds $1.818989$ together. Each thousandfold increase of the bank adds only $9.954224$ to $9.957367$ (Section 3).

Third, measured by logarithmic utility, the certain equivalent is $4$——measuring not the amount but the expectation of its natural logarithm (Bernoulli), the value is $2\ln 2=1.386294$, and the certain amount of equal worth is exactly $4$ (Section 4).

Fourth, and this is the core. Making only the number of plays finite, without a cap, the average does not settle——the average prize over $n$ plays was simulated and the median of $41$ repetitions taken: $8.62$ for $n=100$ and $\mathbf{22.295930}$ for $n=10^6$. Multiplying $n$ by $10^4$ raised the average by $13.675930$, differing by $0.388218$ from $\log_2 10^4=13.287712$. The average keeps growing like the logarithm of $n$ (Section 5).

Fifth, the separator is what is made finite——a finite bank brings the expectation to about $\log_2 B$, logarithmic utility gives a certain equivalent of $4$, and a finite number of plays gives a finite average that keeps growing with $n$. The infinity remains in a different form depending on where it is cut (Section 6).

Keywords: St Petersburg paradox, expected value, logarithmic utility, bankroll limit, Daniel Bernoulli.

Scope (what is not claimed)

No new theorem or law is claimed——that this game has infinite expectation, that logarithmic utility makes it finite (Daniel Bernoulli, $1738$), and that the average over plays grows like the logarithm of $n$ (Feller, $1945$) are all known. Gambling is not recommended. The merits of utility theory are not discussed——logarithmic utility is only computed as one way of making the game finite. The simulation uses one random seed——the deviation of the median from $\log_2 n$ ($+1.976144$ to $+3.312216$) moves with the seed. Relation to earlier papers: Paper 298 showed that a doubling strategy winning $99.9999\%$ of the time still has expectation exactly $0$, because of the fence of capital——here the same fence of capital shrinks an infinite expectation to about $\log_2 B$. What is added is matching the capped expectation by two routes (the sum by definition and $K+B/2^K$), a simulation varying the number of plays from $10^2$ to $10^6$, and placing the separator on what is made finite.

1. Notation and the Question

The game: a coin is tossed until heads. If the first head comes on toss $k$, $2^k$ is paid. The probability that the first head comes on toss $k$ is $1/2^k$.

Question: when the expectation is infinite, what remains if the bank, the utility or the number of plays is made finite?

2. The Expectation Is Infinite

Observation 2.1 the term for toss $k$ is the product of the probability $1/2^k$ and the amount $2^k$, exactly $1$. Every $k$ adds $1$ and the terms never run out, so the expectation grows without bound. Large amounts are rare, but each halving of the probability comes with a doubling of the amount, so what is added does not shrink.

3. Limiting the Bank

Computation 3.1: with only $B$ paid once $2^k$ exceeds $B$, the expectation was found by two routes, the sum by definition $\sum_k \min(2^k,B)/2^k$ and the formula $K+B/2^K$ (with $K$ the largest integer for which $2^K\le B$), and they agreed.

bank $B$expectation$\log_2 B$difference$B/2^K$
$10^3$$10.953125$$9.965784$$0.987341$$1.953125$
$10^6$$20.907349$$19.931569$$0.975780$$1.907349$
$10^9$$30.862645$$29.897353$$0.965292$$1.862645$
$10^{12}$$\mathbf{40.818989}$$39.863137$$0.955852$$1.818989$
$10^{15}$$50.776357$$49.828921$$0.947435$$1.776357$

Observation 3.2 up to toss $K$ each adds $1$, and everything beyond is cut to $B$, adding $B/2^K$ together (at least $1$ and less than $2$). For $10^{12}$ this is $39$ from the first $39$ tosses and $1.818989$ from the rest——each thousandfold increase of the bank adds only $9.954224$ to $9.957367$, close to $\log_2 1000=9.965784$. The infinite expectation shrinks to the logarithm of the bank.

4. Logarithmic Utility

Computation 4.1: the expectation of the natural logarithm of the amount is $\sum_k k\ln2/2^k=2\ln2=1.386294$. The certain amount of equal worth is $e^{2\ln 2}=4.000000$.

Observation 4.2 measured by the logarithm, the term for toss $k$ becomes $k\ln2/2^k$, and the halving of probability outpaces the doubling of the amount. So the sum is finite——what was made finite is not the amount paid but the way it is measured.

5. Limiting the Number of Plays

Computation 5.1: the average prize over $n$ plays without a cap was simulated, repeated $41$ times, and the median taken.

$n$median of the average$\log_2 n$difference
$10^2$$8.620000$$6.643856$$+1.976144$
$10^3$$13.278000$$9.965784$$+3.312216$
$10^4$$15.764000$$13.287712$$+2.476288$
$10^5$$19.714700$$16.609640$$+3.105060$
$10^6$$\mathbf{22.295930}$$19.931569$$+2.364361$

From $n=10^2$ to $10^6$ the median of the average rose by $13.675930$, differing by $0.388218$ from $\log_2 10^4=13.287712$.

Observation 5.2 this is the core: making the number of plays finite makes the average finite, but it keeps growing as the number increases, with no value to settle at. The more plays, the more likely a rare large amount is met, so the average grows like the logarithm of $n$——reading “over many plays the average settles at the expectation” belongs to the case where the expectation is finite.

6. The Separator

Observation 6.1 the separator is what is made finite. With the bank, the expectation is about $\log_2 B$; with the utility, the certain equivalent is $4$; with the number of plays, the average keeps growing like $\log_2 n$——the same infinity remains, depending on where it is cut, as the logarithm of the bank, a single constant, or the logarithm of the number of plays.

Observation 6.2 beside Paper 298. In Paper 298, a doubling strategy winning $99.9999\%$ of the time had expectation exactly $0$ because of the fence of capital——here the same fence of capital shrinks an infinite expectation to about $\log_2 B$. In both, capital rather than probability decides the answer.

Observation 6.3 what is not said. What one should pay for this game is not stated, the merits of utility theory are not discussed, and gambling is not recommended——what is shown is the number that remains under each of three ways of making the game finite.

7. Roll Call and Closing

  1. every $k$ adds $1$ (Observation 2.1): the expectation is infinite.
  2. $\mathbf{40.818989}$ with a bank of $10^{12}$ (Computation 3.1): two routes agree.
  3. $0.947435$ to $0.987341$ above $\log_2 B$ (Computation 3.1): shrunk to the logarithm of the bank.
  4. about $9.95$ more per thousandfold (Observation 3.2): close to $\log_2 1000$.
  5. $4$ by logarithmic utility (Computation 4.1): the exponential of $2\ln2$.
  6. the measure was made finite (Observation 4.2): not the amount paid.
  7. $\mathbf{22.295930}$ for $n=10^6$ (Computation 5.1): median of the average.
  8. $13.675930$ more for $10^4$ times the plays (Computation 5.1): $0.388218$ from $\log_2 10^4$.
  9. the average does not settle (Observation 5.2): the core.
  10. the fence of capital decides (Observation 6.2): Paper 298.

Closing: The game of tossing until heads and receiving $2^k$ has infinite expectation, since every $k$ adds $1$. Limiting the bank to $B$ leaves an expectation less than $1$ above $\log_2 B$: $40.818989$ even with $10^{12}$. Measured by logarithmic utility the certain equivalent is $4$. Limiting only the number of plays makes the average finite, but it keeps growing with the number and does not settle——multiplying $n$ by $10^4$ raised the median of the average by $13.675930$. The separator is what is made finite. Placed among the earlier papers——the fence of capital that made the expectation $0$ in Paper 298 here shrinks the infinity to the logarithm of the bank. To be honest——neither what one should pay for this game nor the merits of utility theory are entered.

References

  1. Bernoulli, D. (1954). Exposition of a new theory on the measurement of risk. Econometrica.
  2. Feller, W. (1945). Note on the law of large numbers and “fair” games. The Annals of Mathematical Statistics.
  3. Yamagishi, Y. (2026). A Strategy That Wins $99.9999\%$ of the Time Has Expectation Exactly $0$ [Paper 298]. Zenodo. 10.5281/zenodo.22171857

Reference

Yamagishi, Y. (2026). A Game with Infinite Expectation Has an Expectation of About $41$ Once the Bank Is Limited to $10^{12}$ [Paper 629]. YOI: https://www.yamagishiyuuki.com/yoi/629
DOI: 10.5281/zenodo.23014552
ORCID: 0009-0006-8587-3450

PDF (English) · PDF (Japanese) · Date proof of the English PDF (OpenTimestamps .ots) · SHA256SUMS

The record page for this paper (in Japanese, with the PDFs in each language and the date proof): YOI 629. This paper is a preprint and has not been peer reviewed.