St. Petersburg Paradox¶

A gamble whose expected monetary value is infinite yet which real deciders will pay only a few coins to enter, exposing that a value function linear in money mishandles fat-tailed payoffs and must be replaced by a concave or bounded utility.

Core Idea¶

The St. Petersburg paradox is the observation, introduced by Nicolaus Bernoulli (1713) and analysed by Daniel Bernoulli (1738), that a specific gamble has infinite expected monetary value yet real decision-makers will pay only a modest finite sum to play it. The game: a fair coin is flipped repeatedly until tails first appears; the payoff is 2^n ducats where n is the number of flips including the final tails. The expected monetary value is the sum over all n of (½^n) × 2^n = 1 + 1 + 1 + … = infinity. Yet almost no one will pay more than a few ducats to enter. The paradox is the gap between what expected-value maximisation prescribes and what observed willingness-to-pay reveals.

Daniel Bernoulli's resolution is the foundation of modern utility theory: he proposed that the subjective value (utility) of money is not linear but concave — the marginal utility of each additional unit of wealth is smaller than the previous one, so the expected utility of the gamble is finite even though its expected monetary value is not. A decision-maker with logarithmic utility, for example, assigns finite expected utility to the gamble and will pay only a finite price for it, approximating observed behaviour. This concavity argument introduced the distinction between risk-neutral decision-makers (who maximise expected monetary value) and risk-averse ones (who maximise expected utility of a concave utility function), and directly prefigured the axiomatisation of expected utility theory by von Neumann and Morgenstern (1944). The paradox is also the historical entry point for recognising that expected-value maximisation is inadequate as a normative criterion when payoff distributions have divergent or fat tails, a lesson that recurs in decision problems with heavy-tailed payoffs such as catastrophic insurance and venture investment with power-law returns.

Structural Signature¶

Sig role-phrases:

the gamble structure — a coin flipped until tails, payoff 2^n for n flips: continuation probability declining geometrically while payoff grows proportionally
the divergent expected monetary value — the sum Σ (½^n)·2n = ∞ under linear-in-money scoring, a property of the distribution alone
the observed finite willingness-to-pay — real deciders paying only a few ducats, the behavioral fact the infinite EV contradicts
the linear-money/divergent-tail mismatch — the diagnosed pathology: a value function linear in money meeting a mean dominated by vanishingly probable enormous payoffs
the concave/bounded-utility repair — replacing linear money-value with a concave or bounded utility, rendering expected utility finite while expected money diverges
the risk-neutral/risk-averse split — the distinction the repair statably introduces, with utility curvature indexing risk aversion as a one-parameter family
the EV/EU wedge — the load-bearing separation of expected monetary value (distribution) from expected utility (preferences), so the decision rule cannot live in the distribution
the divergent-mean boundary — the scope limit: expected-value maximization is admissible only when the mean is well-behaved, inadmissible for fat-tailed or divergent expectations
the inverse rationalization move — recovering the utility function under which the observed price is rational (e.g. log utility), the program von Neumann–Morgenstern systematized

What It Is Not¶

Not a genuine logical paradox. It is a mismatch between a decision criterion (maximize expected money) and observed behavior, not a contradiction in the theory. It dissolves the moment value is allowed to be concave in money: the expected utility is finite while the expected money diverges. "Paradox" marks a surprise that resolves, not an antinomy.
Not uniquely resolved by concave utility. Diminishing marginal utility is the historical (Bernoulli) resolution, but bounded payoffs (no casino can pay an infinite prize), bounded utility, and probability weighting that compresses extreme tails each independently defuse the divergence. Reading concavity as the only answer mistakes one repair for the structural fact that any bound or sub-linear transform tames the tail.
Not evidence that real deciders are irrational. The modest finite price is not a failure to compute the mean; it reveals a concave (or bounded) value function and is fully rational under it. The paradox's enduring move is inverse — "under what utility is this willingness-to-pay rational?" — which turns an apparent anomaly into evidence about the shape of preferences, the program von Neumann–Morgenstern systematized.
Not a demonstration that expected-value reasoning is generally wrong. Expected-value maximization is inadmissible only when the payoff distribution's mean diverges or is fat-tailed. For well-behaved means it remains the right criterion; the lesson is to check whether the mean is even well-behaved before trusting an EV computation, not to abandon expectations wholesale.
Not a portable structural pattern. It is a named example and pedagogical anchor for expected_utility, risk_aversion, and heavy_tail (one of a family with Allais, Ellsberg, Newcomb, Pascal's Wager). The transferable content — "an unbounded linear-in-payoff criterion mishandles fat tails; bound or concavify" — belongs to those primes; Pascal's Mugging in AI alignment is that parent prime recurring, not "the St. Petersburg paradox" arriving.

Scope of Application¶

The St. Petersburg paradox lives across decision theory, economics, finance, and the philosophy of probability, as a diagnostic-and-repair and pedagogical anchor; it operates wherever an agent faces a heavy-tailed or divergent-expectation gamble. Its transferable content belongs to the parent primes expected_utility / risk_aversion / heavy_tail (Pascal's Mugging is that parent recurring, not the paradox arriving), so the genuine habitats are the in-domain contexts where the specific gamble organizes analysis.

Pedagogical anchor for utility theory — the canonical textbook entry point: every standard microeconomics and decision-theory course introduces concave utility and the expected-value/expected-utility wedge via St. Petersburg.
The risk-aversion debate — the paradox is the canonical motivation for adopting expected utility, rather than expected value, as the normative decision criterion, and for the risk-neutral/risk-averse distinction.
Tail-risk finance — lottery, venture-capital, and catastrophic-insurance problems with ill-behaved means are analyzed through St. Petersburg logic: divergent expectations forced through concave or bounded utility transforms.
Foundations of probability and rationality — the paradox features in philosophical debate over whether expected-value reasoning is normatively binding when expectations diverge (Hájek's "Vexing Expectations").
AI / decision-systems design — Pascal's-Mugging defenses against unbounded-expectation policies are modern St. Petersburg defenses, bounding utility so a tiny probability times an enormous claimed payoff cannot dominate.

Clarity¶

The paradox's clarifying force is to drive a wedge between two quantities that pre-Bernoulli reasoning treated as one: expected monetary value, a property of the payoff distribution alone, and expected utility, a property of the agent's preferences over outcomes. By exhibiting a gamble whose expected money is infinite while its accepted price is a few ducats, it makes legible that the decision criterion cannot live in the distribution — no amount of recomputing the mean rescues expected-value maximization — but must incorporate how the agent values wealth. That separation is the hinge on which risk-sensitive economics turns: once "what the gamble is worth in money" and "what it is worth to a decider" are held apart, the concavity move (diminishing marginal utility of wealth) renders the gamble's expected utility finite, and the distinction between a risk-neutral agent maximizing expected money and a risk-averse one maximizing expected utility of a concave function becomes statable. The paradox thereby lets a practitioner stop asking "what is the gamble's expected value?" and start asking the sharper "under what utility function is this finite willingness-to-pay rational?" — which is the question the von Neumann–Morgenstern axiomatization later answered systematically.

It also makes a specific structural failure mode visible: expected-value maximization is inadequate as a normative criterion precisely when the payoff distribution has a divergent or fat-tailed mean. The clarity is in localizing the breakdown to the tail. A practitioner who has absorbed the paradox knows that the trouble is not arithmetic error but the linear-in-money scoring rule meeting a distribution whose mean is dominated by vanishingly probable enormous payoffs — and so knows where to look in any heavy-tailed decision problem, from catastrophic insurance to power-law venture returns, for the same pathology. The sharp question it licenses there is not "what is the expected payoff?" but "is the mean even well-behaved, and if not, what bound or concave transform should replace it?"

Manages Complexity¶

The paradox compresses an open-ended class of decision problems — every gamble, lottery, insurance contract, or investment with a divergent or heavy-tailed payoff distribution — into a single diagnosable failure and a single repair. Confronting such problems one by one, an analyst might try to patch each with an ad hoc rule (cap the payoff here, ignore the rare catastrophe there, distrust the arithmetic somewhere else). The paradox collapses that case-by-case scramble to one structural recognition: the trouble is never an error in computing the mean but the linear-in-money scoring rule meeting a distribution whose mean is dominated by vanishingly probable enormous payoffs. Once that is the named pathology, the analyst tracks just two things — is the payoff distribution's mean well-behaved, and is the value function linear in money? — and the qualitative verdict (expected-value maximization is or is not adequate here) follows without re-litigating each gamble.

The repair compresses just as sharply. Rather than a bespoke fix per problem, the paradox installs one parametric lever: replace the linear money-value with a concave (or bounded) utility, and the divergent expected money becomes a finite expected utility. The entire space of "how much should one pay for a heavy-tailed gamble?" reduces to the shape of a single function — its curvature indexing risk aversion, the gap between risk-neutral (linear) and risk-averse (concave) decision-makers becoming a one-parameter family rather than a zoo of unrelated attitudes. An analyst facing catastrophic insurance, power-law venture returns, or a Pascal's-Mugging-style unbounded claim reads the same two-step off the paradox: check whether the mean is well-behaved, and if not, ask which concave transform or bound restores a finite, decidable value — a high-dimensional menu of tail-risk problems folded onto one diagnostic and one tunable utility curve.

Abstract Reasoning¶

The St. Petersburg paradox licenses a compact set of moves in decision theory under risk, all organized around the wedge between expected monetary value and expected utility, and the tail pathology that opens it.

Diagnostic (locate the breakdown in the linear-money rule meeting a divergent tail). The defining move is to attribute the gap between an infinite expected value and a modest willingness-to-pay not to arithmetic error but to a structural mismatch: a value function linear in money meeting a payoff distribution whose mean is dominated by vanishingly probable enormous payoffs. The signature inference runs from a finite accepted price for an infinite-EV gamble to the conclusion that the agent's value function cannot be linear in money — it must be concave or bounded, because only a sub-linear valuation renders the gamble's expected utility finite while its expected money diverges. The reasoning thus moves from observed behavior (small price) to a hidden property of preferences (diminishing marginal utility of wealth), and the diagnostic generalizes: in any heavy-tailed decision problem the analyst looks first at the tail, asking whether the mean is even well-behaved before trusting any expected-value computation.

Boundary-drawing (when expected-value maximization is an admissible criterion). The paradox draws a sharp boundary on the scope of the expected-value rule: it is adequate as a normative criterion only when the payoff distribution's mean is well-behaved, and is ruled inadmissible precisely when the mean diverges or is fat-tailed. The move "maximize expected monetary value" is out of bounds for a gamble whose expectation is dominated by rare extreme payoffs, and no amount of recomputing the mean rescues it — the criterion itself must be replaced. This boundary also separates two quantities the analyst must never conflate: expected monetary value (a property of the distribution alone) versus expected utility (a property of the agent's preferences) — holding them apart is what makes the criterion question decidable, since the decision rule cannot live in the distribution but must incorporate how the agent values wealth.

Interventionist / repair (restore a finite decidable value by concavifying or bounding). The paradox installs a single parametric repair: replace the linear money-value with a concave or bounded utility, and the divergent expected money becomes a finite expected utility with a determinate willingness-to-pay. The move is predicted to work in proportion to curvature — a more concave utility yields a lower finite price — so the gap between a risk-neutral (linear) and a risk-averse (concave) decision-maker becomes a one-parameter family indexed by the utility function's shape rather than a collection of unrelated attitudes. The interventionist content is therefore a tunable lever: choose the concave transform or bound that restores a finite value, and the curvature of that choice is the agent's risk aversion. The analyst facing catastrophic insurance, power-law venture returns, or an unbounded-claim problem applies the same repair and reads the implied risk attitude off the transform required to tame the tail.

Inverse / rationalization reasoning (recover the preference that makes observed behavior rational). The paradox licenses an inverse move characteristic of utility theory: rather than asking "what is the gamble worth?", ask "under what utility function is this finite willingness-to-pay rational?" — inferring the agent's preferences from their accepted price. A logarithmic utility, for instance, assigns the gamble finite expected utility consistent with paying only a few units, so the observed behavior reveals a concave value function rather than contradicting rationality. This reasoning — from choice back to the preference structure that justifies it — is exactly the program the von Neumann–Morgenstern axiomatization systematized, and the paradox is its historical entry point: it converts a behavioral anomaly into evidence about the shape of the value function.

Knowledge Transfer¶

Within decision theory, economics, finance, and the philosophy of probability the St. Petersburg paradox transfers as mechanism — more precisely, as a diagnostic-and-repair and a pedagogical anchor — across the settings that share its tail pathology. The same recognition (a linear-in-money scoring rule meeting a distribution whose mean is dominated by vanishingly probable enormous payoffs) and the same repair (replace the linear value with a concave or bounded utility, whose curvature is the risk aversion) carry intact into the textbook introduction of utility theory, the risk-aversion debate (the canonical motivation for expected utility over expected value as the normative criterion), tail-risk finance (lottery, venture capital, catastrophic insurance — ill-behaved means forced through utility transforms), foundational debates about whether expected-value reasoning is binding when expectations diverge (Hájek's "Vexing Expectations"), and AI decision-systems design (Pascal's Mugging defenses against unbounded-expectation policies). The diagnostics carry with the vocabulary — the expected-monetary-value/expected-utility wedge, the divergent-mean boundary on the EV criterion, the concave/bounded repair, the inverse "under what utility is this price rational?" move — wherever an agent faces a heavy-tailed or divergent-expectation gamble. The von Neumann–Morgenstern axiomatization is the systematized form of the same inverse reasoning the paradox introduced.

Beyond decision theory the honest reading is the shared-abstract-mechanism case (B): the paradox is one named example and pedagogical entry point, not itself a portable structural pattern, and what travels is the cluster of primes it instantiates. The transferable content — "an unbounded linear-in-payoff criterion mishandles fat tails, requiring a concave or bounded transform" — is the content of expected_utility and risk_aversion (with prospect_theory as the behavioral refinement, and concavity and fat_tail/heavy_tail as the structural relations). Those primes are what genuinely recur across substrates, and the cross-domain lesson should be carried by them. The clearest case is Pascal's Mugging in AI alignment: it is structurally the same problem (a tiny probability times an enormous claimed payoff yielding unbounded expectation under linear scoring), but its solution is just the expected-utility/bounded-utility prime applied in a new substrate — so the recurrence there is the parent prime as a co-instance, not "the St. Petersburg paradox" arriving.

The home-bound cargo is essentially the specific gamble and its pedagogical framing: the coin-until-tails structure with geometrically declining continuation probability and proportionally growing payoff, the ducat denomination, the historical Bernoulli resolution, and the paradox's role as the textbook anchor for utility theory. Strip the coin/ducat vocabulary and the paradox dissolves into "a distribution with a divergent mean is not well-handled by linear value functions; bound or concavify" — a methodological lesson, not a structural object, and one already housed in the expected-utility family. So invoking "the St. Petersburg paradox" in another substrate is best understood as pointing to that family via its most famous illustration; the honest move is to name expected_utility / risk_aversion (or heavy_tail) for the actual content, reserving the paradox as the teaching example. It sits in a family of named decision-theory anomalies (Allais, Ellsberg, Newcomb, Pascal's Wager) that share this disposition — pedagogical anchors for the same underlying primes rather than primes themselves. One discipline travels usefully wherever a heavy-tailed decision problem appears and is the paradox's sharpest lesson: look at the tail first — ask whether the mean is even well-behaved before trusting any expected-value computation, and if it is not, replace the criterion (concavify or bound) rather than recompute it. That habit generalizes to any divergent-expectation problem, which is exactly why the example endures. Mechanism (as diagnostic-and-repair) within decision theory, parent-prime (expected_utility / risk_aversion / heavy_tail) recurrence beyond — the profile Structural Core vs. Domain Accent makes precise.

Examples¶

Canonical¶

The defining computation is the gamble itself against Bernoulli's log-utility resolution. A fair coin is flipped until the first tails on flip n, paying 2ⁿ ducats. The expected monetary value is Σₙ (½ⁿ)·2ⁿ = 1 + 1 + 1 + ⋯ = ∞, so any risk-neutral, money-linear agent should pay any finite entry fee — yet people offer only a few coins. Daniel Bernoulli (1738) proposed valuing the payoff by its logarithm. The expected utility is then Σₙ (½ⁿ)·ln(2ⁿ) = ln 2 · Σₙ n/2ⁿ = ln 2 · 2 = 2 ln 2 ≈ 1.386. The certainty-equivalent price C satisfies ln C = 2 ln 2, giving C = 4 ducats — a small, finite figure close to what people actually pay. The infinite money-expectation collapses to a modest utility-value.

Mapped back: The coin-until-tails, pay-2ⁿ structure is the gamble structure; the diverging sum Σ (½ⁿ)·2ⁿ = ∞ is the divergent expected monetary value. Substituting ln for linear money is the concave/bounded-utility repair, the finite 2 ln 2 versus infinite money is the EV/EU wedge made numeric, and solving for C = 4 is the inverse rationalization move — recovering the utility (log) under which the observed small price is rational.

Applied / In Practice¶

Nick Bostrom's "Pascal's Mugging" (2009) is a working St. Petersburg problem inside AI decision-system design. A mugger tells an expected-utility-maximizing agent: give me five dollars and I will, with tiny probability, produce an astronomically large number of units of value. Because the claimed payoff can be made to grow faster than the agent's assigned probability shrinks, a linear-in-value agent computes an unbounded expected utility and hands over the money — and can be repeatedly hijacked by any such unfalsifiable enormous claim. The engineering response deployed in AI-safety work is precisely Bernoulli's: bound or sufficiently concavify the agent's utility function so that no tiny-probability claim, however large, can dominate the expectation, restoring a finite, decidable value.

Mapped back: The tiny probability multiplied by an astronomically large claimed payoff is exactly the linear-money/divergent-tail mismatch — a mean dominated by a vanishingly probable enormous outcome. Bounding the agent's utility is the concave/bounded-utility repair, and the recognition that unbounded expected-value maximization is the wrong criterion here rather than a computation to be redone is the divergent-mean boundary applied in a new substrate.

Structural Tensions¶

T1: Concave utility as the resolution versus one repair among several (curvature is not the only tail-tamer). Bernoulli's concavity is the historical and canonical fix — log utility turns the infinite money-expectation into a finite 2 ln 2 ≈ 1.386 and a certainty-equivalent of 4 ducats — and the paradox is standardly taught as the motivation for concave utility. But the divergence is defused equally by a bounded payoff, a bounded utility, or probability weighting that compresses the extreme tail; any sub-linear or bounded transform works. The tension is that presenting concavity as the answer mistakes one repair for the structural fact, and it obscures that concavity alone is a weaker medicine than bounding — a value function that is concave but unbounded can be re-broken by a payoff schedule engineered to outrun it, whereas a hard bound cannot. The pedagogical convenience of "just use log utility" hides that what actually guarantees a finite value is a ceiling, of which concavity is only one soft version. Diagnostic: Is the finite value secured by genuine boundedness of the valuation, or merely by concavity that a faster-growing tail could still overwhelm?

T2: Revealing rational preferences versus rationalizing any behavior (the inverse move's power and its emptiness). The paradox's most durable contribution is the inverse read: the modest accepted price is not a failure to compute the mean but reveals a concave value function, so the behavior is rational under the utility it implies — the program von Neumann-Morgenstern systematized. This rescues rationality and converts an anomaly into evidence about preferences. But the same move is nearly unfalsifiable: for almost any observed willingness-to-pay there exists some utility function under which it is optimal, so "recover the preference that makes the choice rational" can absorb any behavior whatever. The tension is that the inverse inference is simultaneously the paradox's deepest insight and a standing invitation to circularity — the utility is inferred from the choice it is then said to explain. Discipline requires the utility function be constrained independently (by other choices) rather than fitted to the one datum. Diagnostic: Is the concave utility identified from independent evidence about the agent's preferences, or reverse-engineered solely to rationalize the price in question?

T3: Replacing the criterion versus over-throwing expectations (the divergent-mean boundary cuts both ways). The paradox's boundary lesson is exact: expected-value maximization is inadmissible only when the payoff distribution's mean diverges or is fat-tailed, and for well-behaved means it remains the correct criterion. This licenses two opposite errors. One is the original mistake — trusting the linear-money mean when the tail makes it meaningless. The other is over-learning the paradox into a general suspicion of expected-value reasoning, abandoning a criterion that is perfectly sound wherever the mean is finite and thin-tailed. The tension is that the correction ("check the tail before trusting the mean") is not a repudiation of expectations but a scoped one, and mistaking the scope in either direction — applying EV to a divergent tail, or refusing EV for a benign one — is a failure the paradox is meant to prevent. Diagnostic: Has the mean actually been shown ill-behaved here, or is the expected-value criterion being discarded reflexively for a distribution whose mean is in fact well-defined?

T4: An unphysical idealization versus a real recurring structure (the infinite prize that cannot exist). The divergence depends on an idealized infinite payoff schedule that no casino, with a finite bankroll, could ever honor — bounded payoffs alone dissolve the paradox before any utility argument is needed. On one reading the "paradox" is therefore an artifact of taking a mathematical limit more seriously than any realizable game warrants. Yet the structure it isolates — a value dominated by vanishingly probable enormous outcomes — recurs in genuinely heavy-tailed decision problems (catastrophic insurance, power-law venture returns, Pascal's Mugging) where tails are finite but fat enough that naive means mislead. The tension is that the paradox earns its keep precisely by exaggerating to the point of impossibility, and the analyst must decide whether an invoked St. Petersburg structure is a real heavy tail worth respecting or an idealization already tamed by the finiteness of the world. Diagnostic: Is the troubling tail a genuine heavy-tailed feature of the actual payoff distribution, or an unrealizable infinite idealization that any real-world bound already removes?

T5: Autonomy versus reduction (a named teaching example or the expected-utility, risk-aversion, and heavy-tail primes it illustrates). The St. Petersburg paradox is a specific, historically anchored construction — coin-until-tails, 2ⁿ ducats, Bernoulli's log-utility resolution — and a canonical pedagogical entry point for utility theory, sitting in a family of named anomalies (Allais, Ellsberg, Newcomb, Pascal's Wager). But it is not itself a portable structural pattern: strip the coin-and-ducat vocabulary and what remains — "a linear-in-payoff criterion mishandles a fat tail; bound or concavify" — is the content of expected_utility, risk_aversion, and heavy_tail, which are what genuinely recur across substrates. Pascal's Mugging in AI alignment is that parent prime arriving in a new substrate, not "the St. Petersburg paradox" traveling. The tension is between a vivid example that earns its permanent place in the curriculum and the recognition that its transferable cargo already belongs to more general primes it merely illustrates. Diagnostic: Resolve toward expected_utility / risk_aversion / heavy_tail when carrying the lesson to another substrate; toward the named paradox when teaching or diagnosing the EV/EU wedge on this specific gamble in situ.

Structural–Framed Character¶

The St. Petersburg paradox sits in the mixed region of the spectrum, with a profile specific to its being a named example rather than a mechanism. Its evaluative_weight is nil: the paradox names a mismatch between a decision criterion and observed behavior, and the entry is emphatic that it is not evidence of irrationality — it renders no verdict, only exposes where a scoring rule breaks. It is human_practice_bound only weakly: it concerns agents with preferences over gambles, but its load-bearing core — that Σ (½ⁿ)·2ⁿ diverges under linear-in-money value, tamed by a concave transform — is a mathematical fact independent of any observer, and the divergent mean is discovered, not stipulated. Its institutional_origin is moderate and confined to the packaging: the specific coin-until-tails/ducat construction and its role as the textbook anchor for utility theory are artifacts of decision theory's pedagogical tradition (Bernoulli, von Neumann–Morgenstern), while the underlying tail pathology is not. On vocab_travels the coin/ducat gamble stays home, and on import_vs_recognize the entry is unusually explicit that the paradox is not itself a portable structural pattern: what recurs across substrates (Pascal's Mugging in AI alignment) is the parent prime arriving as a co-instance, not "the St. Petersburg paradox" traveling.

The portable skeleton is an unbounded linear-in-payoff criterion mishandles a fat tail and must be bounded or concavified — the content of the parent primes expected_utility, risk_aversion, and heavy_tail/fat_tail (with concavity as the structural relation and prospect_theory as the behavioral refinement). That cluster is what the St. Petersburg paradox illustrates and instantiates, and it is what carries the cross-domain lesson; the paradox itself is one named example in a family of decision-theory anomalies (Allais, Ellsberg, Newcomb, Pascal's Wager), each a pedagogical anchor for the same primes rather than a prime in its own right. Its distinctive cargo — the coin-until-tails structure, the ducat denomination, the historical Bernoulli resolution — stays home. Its character: an evaluatively-neutral, mathematically-cored decision-theory teaching example whose transferable content belongs wholly to its expected-utility/risk-aversion/heavy-tail parents, mixed — structural in that borrowed core and domain-specific as the particular constructed gamble that names it.

Structural Core vs. Domain Accent¶

This section decides why the St. Petersburg paradox is a domain-specific abstraction and not a prime, and carries the case for its domain-specificity — so it is worth being exact about what could lift and what cannot.

What is skeletal (could lift toward a cross-domain prime). Strip the coin and the ducats away and a thin relational structure survives: a decision criterion linear in payoff scores a distribution whose mean is dominated by vanishingly probable enormous outcomes, so the criterion diverges while a real decider's valuation stays finite — and bounding or concavifying the value function restores a finite, decidable price. The portable pieces are abstract — a value function linear in the payoff, a fat- or divergent-tailed payoff distribution, the wedge between expected payoff (a property of the distribution) and expected value-to-the-agent (a property of preferences), and a sub-linear transform whose curvature indexes risk attitude. That is the expected-utility repair of a fat-tailed gamble. The skeleton is genuinely substrate-portable — it recurs in catastrophic insurance, power-law venture returns, and Pascal's Mugging in AI alignment, and is exactly the content of the parents expected_utility, risk_aversion, and heavy_tail/fat_tail (with concavity the structural relation and prospect_theory the behavioral refinement) — but it is the core the entry shares, not what makes it distinctively the St. Petersburg paradox.

What is domain-bound. What makes the concept the St. Petersburg paradox in particular is a specific constructed gamble and its pedagogical framing, and — unusually — this domain accent is a named example rather than a worked mechanism. The coin-flipped-until-tails structure with geometrically declining continuation probability and proportionally growing 2^n payoff; the ducat denomination; the specific diverging sum Σ (½^n)·2n = ∞; the historical Bernoulli log-utility resolution yielding a certainty-equivalent of 4 ducats; and the paradox's institutional role as the canonical textbook entry point for utility theory and the risk-neutral/risk-averse distinction — these are the particular illustration, not a transportable structure. The decisive test: remove the coin-and-ducat construction and the "paradox" does not become a looser version of itself — it dissolves into the bare methodological lesson "a distribution with a divergent mean is mishandled by a linear value function; bound or concavify," which is not a structural object at all but a statement already housed in the expected-utility family. The specific gamble is what earns the name; nothing under it travels as the gamble.

Why this does not clear the prime bar. A prime is a relational structure whose vocabulary travels and whose transfer is recognition of the same mechanism, not analogy. The St. Petersburg paradox's reach is bimodal, and the entry is unusually candid about which mode is which. Within decision theory, economics, finance, and the philosophy of probability the paradox travels as a diagnostic-and-repair and a pedagogical anchor — the EV/EU wedge, the divergent-mean boundary on the expected-value criterion, the concave/bounded repair, and the inverse "under what utility is this price rational?" move all carry intact into the textbook introduction of utility theory, the risk-aversion debate, tail-risk finance, and Hájek-style foundational debate, because each supplies an agent facing a heavy-tailed gamble. Beyond decision theory it does not travel as itself at all: when the same structure appears in AI alignment as Pascal's Mugging, what has arrived is the parent prime as a co-instance — a tiny probability times an enormous payoff diverging under linear scoring, repaired by bounding utility — not "the St. Petersburg paradox," which stays home as the coin-and-ducat teaching example. When the bare structural lesson is needed cross-domain, it is already carried, in more general form, by the primes the paradox merely illustrates: expected_utility and risk_aversion supply the concave-valuation repair, and heavy_tail supplies the fat-tailed-mean pathology that voids the linear criterion. The cross-domain reach belongs to those parents; the named paradox is one constructed gamble in a family of decision-theory anomalies (Allais, Ellsberg, Newcomb, Pascal's Wager) that should stay home as illustrations.

Relationships to Other Abstractions¶

Current abstraction St. Petersburg Paradox Domain-specific

Parents (4) — more general patterns this builds on

St. Petersburg Paradox is part of Expected Utility Prime

The St. Petersburg Paradox contains the probability-weighted utility repair that replaces linear money value with a concave or bounded valuation of outcomes.
St. Petersburg Paradox is part of Expected Value Prime

The St. Petersburg Paradox contains the raw monetary expectation whose divergent sum creates the conflict between the gamble's formal price and observed willingness to pay.
St. Petersburg Paradox is part of Heavy-Tailed Distributions Prime

The St. Petersburg gamble contains a heavy payoff tail whose rare exponentially large prizes dominate and ultimately destroy the ordinary monetary mean.
St. Petersburg Paradox is part of, typical Risk Aversion Prime

The canonical Bernoulli resolution usually contains concave utility and its certainty preference, although bounded or otherwise sublinear utility can tame the tail without that exact interpretation.

Hierarchy paths (16) — routes to 5 parentless roots

St. Petersburg Paradox → Expected Utility → Expected Value → Aggregation → Micro Macro Linkage

Show alternative paths (15)

Not to Be Confused With¶

Pascal's Wager. The other famous infinite-payoff argument in decision theory — that even a tiny probability of God's existence, multiplied by an infinite reward, dominates the expected value of belief. It shares the "infinity in the expectation" motif but runs the opposite way: the Wager wields an unbounded expectation to prescribe an action (believe), whereas St. Petersburg exposes an unbounded expectation as a defect of the linear-money criterion to be repaired. Tell: is infinity being used as a reason to act (Pascal's Wager), or diagnosed as a pathology that voids the expected-value rule (St. Petersburg)?
Pascal's Mugging. The AI-alignment problem in which a tiny probability times an astronomically large claimed payoff yields unbounded expected utility, hijacking a linear-value agent. It is structurally the same problem as St. Petersburg in a new substrate — but that is precisely the parent prime (expected_utility / heavy_tail) recurring, not "the St. Petersburg paradox" arriving; Mugging has no coin-until-tails gamble, only the shared divergent-tail-under-linear-scoring pathology and the same bounded-utility repair. Tell: is there the specific coin-until-tails ducat gamble (St. Petersburg), or an unbounded-claim decision problem in another substrate sharing only the fat-tail structure (Pascal's Mugging, the parent recurring)?
Allais and Ellsberg paradoxes. Sibling decision-theory anomalies that cut the other way: they expose systematic violations of the expected-utility axioms themselves (the independence axiom in Allais, ambiguity-aversion in Ellsberg), challenging expected utility as a description of choice. St. Petersburg instead motivates expected utility, showing why linear-money expected value must be replaced by concave expected utility. Tell: does the puzzle motivate adopting concave utility over linear money (St. Petersburg), or reveal that even expected-utility theory is violated by real choosers (Allais/Ellsberg)?
Prospect theory. The behavioral refinement (Kahneman–Tversky) that replaces a utility-over-wealth curve with a value function over gains and losses plus nonlinear probability weighting. It is a downstream descendant of the lineage St. Petersburg opened, and its compression of extreme-tail probabilities is one of the mechanisms that independently defuses the divergence — but it is a full descriptive theory of risky choice, not the single gamble-and-repair the paradox names. Tell: are you naming the specific divergent-EV gamble and its concave-utility resolution (St. Petersburg), or a general reference-dependent model of choice with probability weighting (prospect theory)?
A genuine logical paradox (e.g. the two-envelope paradox). A true antinomy is a contradiction within the theory that resists resolution. St. Petersburg is not that: it is a mismatch between a decision criterion and observed behavior that dissolves the moment value is allowed to be concave — expected utility is finite while expected money diverges, with no contradiction left standing. Tell: does the puzzle survive as a self-contradiction after the value function is allowed to be nonlinear (genuine paradox), or evaporate once concave or bounded utility is admitted (St. Petersburg, a resolvable surprise)?
The parent primes it illustrates (expected_utility, risk_aversion, heavy_tail). The substrate-general content — an unbounded linear-in-payoff criterion mishandles a fat tail and must be bounded or concavified — that St. Petersburg is the canonical teaching example of, not itself a portable pattern. When the structure appears in catastrophic insurance, power-law venture returns, or AI alignment, the transferable lesson rides with these parents; the coin-and-ducat gamble is the illustration, not the mechanism. Tell: strip the coin-until-tails ducat construction — if what remains is bare "linear value mishandles a divergent mean; concavify or bound," you are using the parents, not the named paradox. (Treated fully in Knowledge Transfer and Structural Core vs. Domain Accent.)

Neighborhood in Abstraction Space¶

St. Petersburg Paradox sits in a moderately populated region (60^th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.

Family — Choice Paradoxes & Collective Decision-Making (14 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-07-12