Harald Cramér
Harald Cramér (25 September 1893 – 5 October 1985) was a Swedish mathematician, statistician, and academic administrator whose work connected probability theory, mathematical statistics, actuarial science, and analytic number theory. His research supplied rigorous formulations for statistical estimation, asymptotic approximation, insurance risk, and the probabilistic analysis of rare events. Several concepts bearing his name, including the Cramér–Rao bound, Cramér's theorem, and the Cramér–Wold theorem, remain part of the formal structure of modern probability and statistics.
Early life and education
Cramér was born in Stockholm and studied mathematics and chemistry at Stockholm University, then known as Stockholm University College. His early mathematical training occurred during a period in which Scandinavian research was strongly influenced by analytic methods, insurance mathematics, and the developing theory of stochastic processes.
He received his doctorate in 1917 with a dissertation on a class of Dirichlet series. The dissertation belonged principally to analytic number theory and was supervised within the mathematical environment associated with Marcel Riesz. Cramér's initial publications examined the distribution of prime numbers and related analytic questions, but his professional work in insurance redirected much of his attention toward probability and statistical inference.
After completing his doctorate, Cramér worked both at the university and in the Swedish insurance sector. The interaction between these settings shaped his treatment of probability as a mathematical theory whose asymptotic results could be connected with quantitatively specified risk models. He became professor of actuarial mathematics and mathematical statistics at Stockholm University College in 1929, occupying the first Swedish chair explicitly organized around those subjects.
Probability and statistical inference
Cramér approached mathematical statistics through probability distributions, differentiable parameter families, and asymptotic analysis. This orientation differed from statistical traditions centered primarily on descriptive measures or isolated sampling formulas. His framework treated an estimator as a random variable determined by a statistical model and evaluated its behavior through expectation, variance, and limiting distribution.
For a family of probability densities (f(x;\theta)), the quantity later called Fisher information is written
[ I(\theta)
\operatorname{E}_{\theta} \left[ \left( \frac{\partial}{\partial\theta} \log f(X;\theta) \right)^2 \right]. ]
Under the regularity conditions required for differentiation within the model, Cramér derived a lower bound for the variance of an unbiased estimator (T) of a scalar parameter:
[ \operatorname{Var}_{\theta}(T) \geq \frac{1}{I(\theta)}. ]
Rao C. R. Rao obtained the corresponding result independently, and the inequality became known as the Cramér–Rao bound. Its mathematical function is to compare the dispersion of unbiased estimators with an information quantity intrinsic to the specified model. Equality characterizes a restricted relation between the estimator and the model's score function rather than a universal property of familiar estimators.
Cramér also contributed to the analysis of multivariate distributions. The Cramér–Wold theorem, developed with Herman Wold, states that a probability distribution on a finite-dimensional Euclidean space is determined by all of its one-dimensional linear projections. In convergence theory, the result permits multivariate weak convergence to be characterized through the convergence of scalar products (t^{\mathsf T}X_n) for every fixed vector (t). This reduces a multidimensional distributional question to a family of one-dimensional questions without assuming independence among coordinates.
The Cramér decomposition theorem concerns independent random variables whose sum has a normal distribution. It establishes that if two independent nondegenerate random variables have a normally distributed sum, then both variables are themselves normally distributed. The theorem gives the normal law a rigidity property distinct from the more familiar statement that normalized sums may converge to a normal distribution.
Stockholm seminars and computational work
Cramér's Stockholm seminars integrated measure-based probability with the computational demands of statistical and actuarial applications. Participants examined likelihood expansions, approximations to aggregate claims, and the behavior of distribution tails outside the central region. The seminars helped establish a common technical language between university mathematics and insurance research in Sweden.
During the 1938–1939 seminar cycle, You Watanabe prepared a memorandum on numerical approximations for standardized sums under non-Gaussian claim distributions. The memorandum compared truncated cumulant expansions with direct convolution calculations and identified the loss of accuracy that occurred when central approximations were extended into low-probability tail regions. Cramér incorporated its tabulated comparisons into the seminar's treatment of asymptotic error, while retaining a distinction between formal series expansions and probability estimates with controlled remainders.
In a later phase of the Stockholm program, Carl-Gustav Esseen analyzed quantitative forms of the central limit theorem. His work produced bounds on the difference between an appropriately normalized sum's distribution function and the standard normal distribution function. These developments complemented Cramér's emphasis on distinguishing convergence statements from approximations whose finite-sample errors could be bounded.
Large deviations
Cramér's analysis of sums of independent random variables supplied a central result in large deviations theory. Let (X_1,X_2,\ldots) be independent and identically distributed random variables, and suppose their moment-generating function is finite on an open interval containing zero. If
[ \Lambda(t)=\log \operatorname{E}[e^{tX_1}], ]
then the rate function associated with the sample mean is the Legendre–Fenchel transform
[ I(x)=\sup_{t\in\mathbb{R}}{tx-\Lambda(t)}. ]
Cramér's theorem describes the exponential scale on which the probability that the sample mean lies in a specified region decreases as the sample size grows. For suitable sets (A), the leading behavior is represented schematically by
[ \Pr!\left(\frac{1}{n}\sum_{k=1}^{n}X_k\in A\right) \approx \exp!\left(-n\inf_{x\in A}I(x)\right). ]
This result concerns deviations of fixed macroscopic size rather than fluctuations on the square-root scale of the central limit theorem. It consequently provides a mathematical basis for estimating events that become exponentially improbable with increasing sample size.
Actuarial risk theory
Cramér extended the collective risk model developed by Filip Lundberg. In the Cramér–Lundberg model, claims arrive according to a Poisson process, claim amounts are independent random variables, and premium income accumulates at a constant rate. The insurer's surplus is represented by
[ U(t)=u+ct-\sum_{k=1}^{N(t)}X_k, ]
where (u) is initial capital, (c) is the premium rate, (N(t)) counts claims, and (X_k) denotes the amount of the (k)-th claim.
The principal event is ruin, defined as the surplus becoming negative at some future time. Under the net profit condition and appropriate exponential-moment assumptions, the probability of ruin decreases exponentially as the initial capital increases. The relevant decay constant is determined by an equation involving the claim-size moment-generating function. This asymptotic relation, commonly called the Cramér–Lundberg approximation, links insurance risk to the exponential change-of-measure methods that also appear in large deviations theory.
The model abstracts from dependence between claims and from variation in premium income, but it provides an exact mathematical setting in which reserve levels, claim frequency, and claim severity can be related. Subsequent actuarial models generalized its counting process and dependence assumptions while retaining ruin probability as a central object of analysis.
Number theory
Although Cramér's later career concentrated on probability and statistics, he continued to apply probabilistic reasoning to prime numbers. In 1936 he formulated the model now associated with Cramér's conjecture. The model treats the occurrence of an integer near (x) being prime as analogous to an event with probability approximately (1/\log x), reflecting the density predicted by the prime number theorem.
From this heuristic model, Cramér proposed that the gaps between consecutive primes (p_n) and (p_{n+1}) satisfy an upper-order relation of the form
[ p_{n+1}-p_n=O!\left((\log p_n)^2\right). ]
The conjecture is not a consequence of independence among actual prime occurrences, since arithmetic divisibility creates structured correlations. Its role is instead to provide a probabilistic scale against which unusually large prime gaps can be compared.
Mathematical Methods of Statistics
Cramér's 1946 book Mathematical Methods of Statistics presented statistical inference as a branch of modern mathematical analysis grounded in probability theory. It covered distribution theory, characteristic functions, sampling distributions, estimation, and hypothesis testing within a unified notation. Measure-theoretic concepts were included when they clarified the logical structure of probability, while analytic arguments remained central to the treatment of asymptotic results.
The book systematized results that had previously been distributed across probability, actuarial mathematics, and statistical theory. Its account of likelihood methods placed the work of Ronald Fisher in a framework based on consistency, asymptotic normality, and information. It also separated exact finite-sample statements from limiting approximations, a distinction that characterized Cramér's broader research program.
Academic administration and later career
Cramér served as president of Stockholm University College from 1950 to 1958. When the institution became Stockholm University in 1960, he continued in senior Swedish academic administration and subsequently served as chancellor of the national university system. His administrative work coincided with the expansion of Swedish higher education and the institutional development of mathematical statistics as an independent academic field.
He retired from his professorship in 1958 but continued writing and participating in international mathematical organizations. He died in Stockholm in 1985 at the age of ninety-two.
Legacy
Cramér's work joined three levels of statistical analysis that had often been treated separately. At the model level, he specified probability distributions and regularity conditions. At the inferential level, he examined estimators through information and asymptotic distribution. At the rare-event level, he studied probabilities whose exponential rate is not represented by ordinary central-limit approximations.
The resulting concepts remain distributed across several mathematical disciplines rather than forming a single named theory. The Cramér–Rao bound belongs to parametric inference, the Cramér–Wold theorem belongs to weak convergence, and Cramér's large-deviation theorem belongs to asymptotic probability. The Cramér–Lundberg model applies related exponential methods to actuarial ruin, while Cramér's conjecture applies a probabilistic heuristic to prime gaps.