Day 54 · 2026.08.15

The Mathematics of Decisions & Gambling

Betting under uncertainty: from utility and risk to growth rates and regret
"The maximum exponential rate of growth of the gambler's capital is equal to the rate of transmission of information over the channel." — J. L. Kelly Jr., 1956

Expected Utility

Why a rational agent does not maximize the expected value of money
Decision Theory
Intuition

In 1738 Daniel Bernoulli posed a gamble: flip a coin until heads appears; if it first lands on flip $n$, you are paid $2^n$. The expected payoff is $\tfrac12{\cdot}2+\tfrac14{\cdot}4+\cdots=\infty$, yet nobody will pay a hundred for it. Nothing was miscalculated — the mistake is treating money as the goal itself. The first million moves you from renting to owning; the tenth adds a digit to a statement. Bernoulli's fix: first convert money into what it means to you — an internal scale called utility — and only then take the expectation. The curve flattens as it rises. It is concave.

wealth x utility u(x) risk premium u(E[X]) E[u(X)] x₁ E[X] x₂
Formal definition

von Neumann and Morgenstern (1944) went further than Bernoulli. Let $\succcurlyeq$ be your preference ordering over random outcomes (lotteries). If it satisfies four axioms — completeness, transitivity, continuity (between a better and a worse option there is always a probability mix exactly equivalent to the middle one), and independence (mixing the same third option into both sides leaves the ranking unchanged) — then a function $u$ must exist such that

$$L_1\succcurlyeq L_2\iff \sum_i p_i\,u(x_i)\ \ge\ \sum_j q_j\,u(y_j)$$

Here $x_i$ are outcomes, $p_i$ their probabilities, and $u$ maps outcomes onto a real-valued scale.

Why it is beautiful

Note the direction of the logic: nobody assumed you have a utility function. The premise is only that your choices are mutually consistent; the conclusion is that your behaviour must then look like maximizing some expectation. Such results are called representation theorems — the numbers are not inserted into the model, they are forced out of the order. And "risk aversion", a psychological word, collapses here into a single geometric statement: $u$ concave $\iff$ $E[u(X)]\le u(E[X])$. The chord always lies below the curve, and the gap is the insurance premium you are willing to pay.

Applications

An RLHF reward model does exactly this: humans supply only pairwise comparisons (A beats B), Bradley–Terry fits them into a single scalar score, and the policy maximizes its expectation — the vNM theorem is what licenses the step "a scalar score exists". Conversely, wherever the axioms break is where behavioural economics enters: in the Allais paradox people systematically violate independence, which is why prospect theory switched to reference points and probability weighting.

Essence + a question
Utility is not a psychological hypothesis about money; it is the mathematical shadow cast by the demand that choices be consistent.
Question: in multi-objective alignment we assume different values can be weighted into one scalar — which presupposes completeness and transitivity. Do human values actually satisfy them?

Coherent Risk Measures

What a legitimate yardstick for risk must satisfy
Convex Analysis
Intuition

Variance is the standard measure of risk, but it has an embarrassment: it penalizes gains just as heavily, and nobody is afraid of earning too much. Hence VaR, which ruled risk management in the 1990s — the line such that on 95% of days your loss stays below it.

The trouble is that it reports only where the threshold sits and says nothing about what lies beyond: two portfolios can share a VaR while one loses slightly more past that point and the other goes to zero. Worse, bundling two assets each with a small VaR can produce a combined VaR larger than their sum — diversification apparently raising the reading. The yardstick is broken.

VaR₉₅ where the cutoff is CVaR₉₅ how bad beyond it loss distribution loss → 5% tail
Formal definition

Artzner et al. (1999) asked the question backwards: what must a legitimate yardstick satisfy? Read $\rho(X)$ as "how much cash must be added before position $X$ is safe". The answer is four axioms: monotonicity, translation invariance $\rho(X+c)=\rho(X)-c$ (cash offsets exactly its own amount of risk), positive homogeneity, and subadditivity $\rho(X+Y)\le\rho(X)+\rho(Y)$. A measure satisfying all four is called coherent. VaR fails precisely on the last one, whereas conditional value at risk

$$\mathrm{CVaR}_\alpha(L)=E\big[\,L \mid L\ge \mathrm{VaR}_\alpha(L)\,\big]$$

— the mean loss inside the tail rather than the entrance to it — passes all four.

Why it is beautiful

Subadditivity looks like a technical condition, but it is the exact encoding of the common-sense demand that diversification never increase risk. The lovely part is its consequence: every coherent risk measure can be written as $\rho(X)=\sup_{Q\in\mathcal Q}E_Q[-X]$ — the worst-case expected loss over a family of candidate probabilities $\mathcal Q$. That proves two things to be one: "I am risk averse" and "I am unsure which probability applies, so I plan for the worst one" are equivalent. Four plain axioms, one pass through convex duality, and robustness and pessimism become the same object.

Applications

In 2016 the Basel framework replaced VaR with ES (i.e. CVaR) as the regulatory market-risk metric, citing subadditivity; Rockafellar and Uryasev showed CVaR optimization reduces to a linear program, making it computable. The same $\sup_Q$ appears in machine learning as distributionally robust optimization: instead of minimizing average loss on the training distribution, minimize the worst loss over a family of distributions near it, which buys resistance to distribution shift.

Essence + a question
The axioms of a risk measure decide whether it rewards or punishes diversification.
Question: how large should the family $\mathcal Q$ inside $\sup_Q$ be? Too small and it is not robust; too large and the optimal policy degenerates into doing nothing. Is there an objective answer?

The Kelly Criterion

What to maximize when wealth is multiplied rather than added
Information Theory
Intuition

You face a gamble with an edge, repeatable, and you choose what fraction of your bankroll to stake each round. Bet everything? One loss removes you permanently — ruin is an absorbing state, and no number of prior wins undoes it. Bet one percent? Safe, but too slow to matter. Somewhere in between lies an optimum.

The key is that wealth is multiplied, not added. Long-run performance of a multiplicative process is governed by the geometric mean, which is the arithmetic mean of $\log$. So what should be maximized is never the per-round expected return but the expected logarithmic growth rate.

g(f*) f* f*/2 2f* growth hits zero bet fraction f long-run growth g(f) gentle steep
Formal definition

Let the odds be $b{:}1$ (stake 1, win $b$), the win probability $p$, $q=1-p$, and $f$ the fraction of current wealth staked each round. After $n$ rounds $W_n=W_0(1+fb)^{k}(1-f)^{n-k}$ with $k$ wins. Take logs, divide by $n$, and the law of large numbers gives the long-run growth rate

$$g(f)=\lim_{n\to\infty}\tfrac1n\log\tfrac{W_n}{W_0}=p\log(1+fb)+q\log(1-f)\ \Longrightarrow\ f^*=\frac{pb-q}{b}=\frac{\text{edge}}{\text{odds}}$$

Note that $\log$ is not an assumed taste in utility; it is what you get from turning multiplication into addition so the law of large numbers can apply.

Why it is beautiful

Kelly derived this at Bell Labs in 1956, with Shannon down the hall, under the title A New Interpretation of Information Rate. For an even-money binary bet the optimal growth rate is $g^*=\log 2-H(p)$, where $H$ is Shannon entropy. The right-hand side is how much more you know than an ignorant bettor; the left-hand side is how fast your money can grow exponentially. What you can earn equals what you know beyond everyone else. A quantity invented to measure code length becomes, unmodified, the ceiling on wealth growth — information theory and gambling are one theorem read twice.

Applications

Thorp used it for blackjack and convertible-bond arbitrage, and carried Kelly into hedge-fund position sizing. In practice people stake only half-Kelly: as the figure shows, $g(f)$ is extremely flat near $f^*$ but plunges to the right, crossing zero at $2f^*$ — and $p$ is always an estimate. On a curve that is gentle to the left and steep to the right, estimation error should be allowed to fall leftwards.

Essence + a question
When wealth is multiplied rather than added, the rational target shifts from expectation to expected logarithm — what changed is not preference but time.
Question: a one-off major decision (changing careers, founding a company) has no "many rounds". What survives of Kelly's argument there?

Multi-Armed Bandits

Putting a price tag on ignorance
Sequential Decision
Intuition

A row of slot machines with different, unknown payoff distributions. Every pull does two things at once: it earns this round's money, and it buys a little information about which machine is better. Information cannot be bought separately — it arrives only as a by-product of betting, and that is what separates this from ordinary optimization. Always pull the currently best arm and you may never find the truly good machine; try them all in turn and most pulls are wasted on arms already known to be bad. The tension cannot be removed, only priced.

payoff estimate 400 pulls 9 pulls ← picked now 120 pulls 30 pulls highest bound dot = mean, bar = confidence radius
Formal definition

With $K$ arms, arm $a$ has expected payoff $\mu_a$ and the best is $\mu^*$. What measures an algorithm is not total reward but regret — the gap against knowing the answer from the start: $R_T=T\mu^*-E\big[\sum_{t=1}^{T}\mu_{a_t}\big]$. At each step UCB selects

$$a_t=\arg\max_a\ \hat\mu_a+\sqrt{\frac{2\ln t}{n_a}}$$

where $\hat\mu_a$ is the current estimate (exploitation) and the square root is a confidence radius: the fewer the pulls ($n_a$ small) the wider it is, and it creeps upward with $t$. This puts a price tag on ignorance and lets long-unverified options resurface on their own. In the figure the arm selected is the one whose mean is not highest but whose upper bound is.

Why it is beautiful

Lai and Robbins (1985) proved that no reasonable algorithm escapes $R_T=\Omega(\log T)$, and UCB attains $O(\log T)$ — the bounds meet in order. This means "optimism in the face of uncertainty" is not an engineering heuristic but a provably optimal principle.

More surprising still is Gittins (1979): under infinite-horizon discounting the optimal policy decomposes — compute for each machine an index depending only on its own history, and pull the largest. A $K$-dimensional coupled dynamic program collapses into $K$ one-dimensional problems. And Thompson had already proposed, in 1933, sampling an arm according to its posterior probability of being best; that too is optimal in order. The frequentist confidence bound and the Bayesian posterior sample arrive at the same frontier from opposite directions.

Applications

A/B testing uses it to keep tilting traffic toward the winner instead of waiting for the experiment to end; recommender cold start, adaptive allocation in clinical trials, and hyperparameter search (Hyperband) are bandits outright. The most famous instance places UCB at the nodes of a search tree: UCT and Monte Carlo tree search, the skeleton of AlphaGo. Preference-data collection for LLMs is the same shape — candidates are arms, and the labelling budget is the number of pulls.

Essence + a question
The cost of exploration can be priced exactly, and the optimal price tag is called optimism.
Question: a failure detector in a distributed system also weighs "wait a little longer" against "declare it dead". Can that be written as a bandit problem, and how would regret be defined there?

Going Deeper

Kelly maximizes log wealth, yet vNM says utility may be any concave function. Is Kelly quietly prescribing a preference?
Samuelson argued this for decades. The substance of the disagreement is time average versus ensemble average: vNM ranks preferences over a single decision, while Kelly describes the almost-sure behaviour of one trajectory over time. For a multiplicative process the two are not equal (non-ergodic), because the law of large numbers acts on $\log W$, not on $W$. So Kelly does not assume "$u=\log$ happens to be right": if the goal is to make that trajectory grow at the maximal rate, $\log$ is derived. Ole Peters' ergodicity economics pushes this further, arguing that many alleged behavioural biases are correct responses to non-ergodicity.
Which of the four expected-utility axioms do real people break first?
Independence. Allais (1953) built two pairs of lotteries on which most people contradict themselves, driven by an extra taste for certainty. The deeper attack is Ellsberg (1961): people prefer betting on an urn of known composition over one of unknown composition — what is disliked is not risk but uncertainty about the probability itself. The repair is exactly concept two: replace a single probability with a family $\mathcal Q$ and take the worst case, and ambiguity aversion becomes a coherent risk measure.
If the payoff distributions drift over time, does the $\log T$ regret bound survive?
No, and the algorithm must actively forget. In the stationary case $\sqrt{2\ln t/n_a}$ relies on "more samples, better estimates"; once the distribution can move, old samples are actively harmful. Sliding-window or discounted UCB rewrite regret in terms of the number of changes, with the typical order degrading to $T^{2/3}$. The price is having to answer "how old is evidence before it expires" — a parameter with no free optimum, encoding your prior about how fast the world changes. Tuning timeout thresholds in distributed systems is the same difficulty.
Is there a single thread running through all four concepts?
Yes, ordered by what you are allowed to know. Expected utility assumes the probabilities are known and the decision happens once; risk measures concede that the probability itself is uncertain and take the worst over a family $\mathcal Q$; Kelly assumes known probabilities but unlimited repetition, so the target shifts from expectation to growth rate; bandits must learn the probabilities while acting, with the price quantified as regret. The four are grid points of one problem along two axes — information and number of repetitions — and dropping either axis makes the advice wrong along the other.