Cramér's Theorem (Large Deviations)¶
An i.i.d.-sample-mean large-deviation principle whose exponential rate is the convex conjugate of the one-observation log moment-generating function.
Core Idea¶
Cramér's theorem describes the exponential scale of rare deviations of an average of independent, identically distributed (i.i.d.) real observations. In the sourced formulation, the one-observation log moment-generating function \(\Lambda(t)=\log E[e^{tX_1}]\) is finite in a neighborhood of zero. Its convex conjugate \(I(x)=\sup_{t\in\mathbb R}\{tx-\Lambda(t)\}\) is the rate function. The theorem says that the empirical means \(\overline X_n=n^{-1}\sum_{i=1}^nX_i\) satisfy a large-deviation principle: probabilities of sets of mean values have logarithmic lower and upper bounds controlled by \(I\). The one-observation law thus determines the leading exponential cost of a sustained deviation in the sample mean.[1]
The full assertion is more careful than the often-quoted formula \(P(\overline X_n\ge a)\approx e^{-nI(a)}\). For a Borel set \(A\subseteq\mathbb R\), Cramér's theorem bounds \(\liminf n^{-1}\log P(\overline X_n\in A)\) using \(\inf_{A^\circ}I\) and \(\limsup\) using \(\inf_{\overline A}I\). A single limit need not follow for an irregular set. For a threshold \(a\) above the mean, the simple logarithmic tail limit follows when \(I\) is finite and continuous around \(a\), so that the interior and closure give the same infimum.[1][2]
The theorem is not an exact finite-sample probability formula and does not determine the multiplicative prefactor. A Chernoff bound can produce a matching-looking upper exponent for finite \(n\); the theorem's lower bound is what establishes that the exponential rate is sharp in eligible cases. This difference matters even for a Gaussian average, whose tail has a factor of order \(n^{-1/2}\) in addition to its exponential term.[1][2][3]
Structural Signature¶
Sig role-phrases: i.i.d. empirical-mean carrier → light-tail log MGF → convex-conjugate rate → interior/closure event bounds → exponential-scale conclusion; optional tilt witness.
- i.i.d. empirical-mean carrier. The objects are real random variables \(X_1,X_2,\ldots\) with one common distribution and independence across observations. Their average is the changing random quantity. Dependence or a changing law requires a different result; merely having a sum is insufficient.[1]
- Light-tail hypothesis. The formulation used here assumes \(\Lambda\) finite on some open interval containing zero. That is a sufficient stated condition, not a claim that every version of every large-deviation theorem demands a moment-generating function finite on all of \(\mathbb R\).[1]
- Convex-conjugate rate. \(I(x)=\Lambda^*(x)\) transforms the single-summand law into a cost for a candidate average \(x\). Under the stated hypothesis, it is nonnegative, vanishes at \(E[X_1]\), and increases to the right of that mean. The relation is constitutive: substituting an unrelated penalty is not Cramér's rate.[1]
- Event-set regularization. Interior and closure are not cosmetic notation. An open-set lower bound and closed-set upper bound tell us what can be asserted for a general Borel event even when its boundary carries a different rate or its interior is empty.[1][2]
- Exponential-scale conclusion. The theorem compares \(n^{-1}\log P(\overline X_n\in A)\) as \(n\to\infty\). It identifies an exponent or bounds on one, not a finite-\(n\) probability, a central-limit approximation, or a universal prefactor.[1][3]
- Optional tilt witness. Exponential change of measure is a common way to prove a lower bound and interpret a rare average. An optimizing tilt may exist for an interior deviation, but its existence at every boundary point is not part of the theorem's recognition test.[2][4]
For a Borel \(A\), the operative form is
The convention \(\inf\varnothing=+\infty\) handles empty interiors. If the interior and closure infima agree, these bounds squeeze the limit to one rate; that equality is an additional event-specific check, not an automatic feature of the theorem.[1]
What It Is Not¶
It is not all of large-deviations theory. That field includes dependent processes, empirical measures, paths and many other carriers. Cramér's theorem is the specific i.i.d.-empirical-mean result. The live Large Deviations Theory node is a broader field-level neighbor, not an exact duplicate of the theorem.[1]
It is not a Chernoff bound. Exponential Markov inequality can yield \(P(\overline X_n\ge a)\le e^{-nI(a)}\) for an eligible right tail. This one-sided, finite-\(n\) certificate does not prove that the bound's exponent is attained asymptotically; Cramér's lower bound supplies the missing half when its conditions and event regularity permit.[1][2]
It is not the central limit theorem. A central-limit approximation describes fluctuations on the \(n^{-1/2}\) scale near the mean. Cramér's theorem concerns fixed deviations and their \(n\)-scale logarithmic probabilities. The central limit theorem may suggest a quadratic local shape but does not generally license a far-tail exponent.[2][3]
It is not a universal equality for every Borel set, nor a promise of a maximizing exponential-tilt parameter at every \(x\). Irregular sets can prevent a single limit; boundary points can frustrate an optimizer even though the convex conjugate remains defined.[1][2]
It is not Cramér's theorem on normal-distribution decomposition, Cramér's rule, or the Cramér–Rao bound. The parenthetical title identifies the large-deviation theorem uniquely.
Scope of Application¶
The theorem applies directly when observations are i.i.d., real-valued, and satisfy an appropriate exponential-moment condition. Binary trials and Gaussian measurements are two unlike mathematical carriers. In both, the same construction starts from a one-observation log MGF and ends with the rate for an empirical mean. A reader must still specify the event whose probability is being assessed, because the rate of a set is the infimum of \(I\) over appropriate interior or closure, not necessarily \(I\) at a casually chosen point.[1][2]
For upper tails above the mean, a particularly useful special case has \(a>E[X_1]\) and \(I\) finite and continuous near \(a\). Then the closed half-line \([a,\infty)\) and its open interior have the same infimum \(I(a)\), yielding \(\lim_{n\to\infty}n^{-1}\log P(\overline X_n\ge a)=-I(a)\). The finite/continuous qualification should travel with this shorthand. Other intervals or complex sets require their own boundary check.[2]
The seed proposed insurance ruin and communications/queue overflow. Sums of independent claims, noises or arrivals can sometimes fit this theorem at a fixed horizon after verifying the law and threshold; the domain words alone are not proof. Ultimate ruin, which asks whether a reserve or net-loss process ever crosses a barrier, changes the event and often the asymptotic variable. It is not the fixed-\(n\) sample-mean claim here, and no ruin exponent is asserted by this entry. Dependence, heavy tails, adaptive policies, or moving thresholds likewise require separate hypotheses or theorems.[1][4]
Clarity¶
A responsible Cramér claim states five items: (1) the i.i.d. summand law; (2) the exponential-moment domain around zero; (3) the log MGF and its convex conjugate; (4) the event \(A\) for the empirical mean; and (5) whether only LDP bounds or an exact logarithmic limit follows. This checklist prevents a true-looking rate formula from being applied to the wrong event.[1][2]
The symbol \(\sim\) is especially hazardous here. In ordinary asymptotics \(P_n\sim e^{-nI}\) can suggest a ratio tending to one. Cramér's theorem generally establishes the weaker logarithmic equivalence \(n^{-1}\log P_n\to-I\) when a limit exists. Goodman’s Gaussian calculation explicitly retains a subexponential prefactor, showing why the two readings differ.[3]
The log MGF should also not be replaced by a sample histogram without an estimation argument. The theorem is a statement about a specified probability law. Estimating that law from finite data adds statistical uncertainty not covered by the theorem itself.
Manages Complexity¶
The result compresses a sequence of increasingly rare events into one deterministic function \(I\). Instead of deriving a new far-tail calculation for every sample size, one transforms the common summand law once and then asks which \(x\) in the target set is least costly. Convexity and the minimum at the mean organize comparison across thresholds.[1]
That compression discards important information. An exponent does not report a probability's prefactor, does not provide a uniform finite-\(n\) error, and does not tell whether an event with a complicated boundary achieves one clean rate. The full open/closed form preserves those limits explicitly rather than hiding them behind an attractive \(e^{-nI}\) slogan.[1][2][3]
The distinction between theorem and computational recipe is also useful. Convex conjugation is the rate-building operation; exponential tilting may help derive or estimate a rare event; a Chernoff calculation supplies a finite-\(n\) upper bound. The theorem relates these ingredients to a two-sided logarithmic asymptotic statement under its hypotheses.[1][2]
Abstract Reasoning¶
Begin with the carrier: define \(S_n\) and \(\overline X_n\) and verify independent identical summands. Check that \(E[e^{tX_1}]\) is finite for \(t\) in a neighborhood of zero. Construct \(\Lambda\) from the one-observation law, then take \(I=\Lambda^*\); do not infer a rate solely from a normal approximation or from a fitted curve.[1]
Next type the target event. For a general Borel set, write its interior and closure and take the corresponding rate infima. If they differ, retain separate lower and upper bounds. For a threshold above the mean, test the rate's finiteness and continuity around the threshold before condensing the result into \(-I(a)\). This is the step omitted by the seed's unconditional tail statement.[1][2]
Finally distinguish the information level. If only an exponential Markov/Chernoff argument has been performed, report an upper bound. If the full theorem's conditions hold, report the LDP; if the event is a continuity set for the rate, report the resulting logarithmic limit. A precise finite-\(n\) probability or prefactor requires further analysis.[1][2][3]
Knowledge Transfer¶
The theorem transfers literally between Bernoulli counts and Gaussian averages because both are i.i.d. empirical means with finite log MGFs. The same roles are present, but their rates differ: binary relative entropy for the former and a quadratic function for the latter. The transferable content is the construction and bounds, not the numerical rate or a single universal tail formula.[1][3]
Many fields can instantiate the theorem by mapping one modeled observation to a summand and a fixed-horizon average to \(\overline X_n\). The mapping must preserve independence, common law, moment condition and event geometry. A dependence model, all-time supremum, or heavy-tail mechanism may admit a large-deviation analysis, but borrowing the Cramér label without a new proof would be an analogy rather than transfer.[1][2]
At a higher level, the live Convex Conjugate entry captures the transform used to build \(I\), and the prime Probability entry captures the broader measure-and-event substrate. Neither carries the theorem's i.i.d.-mean and asymptotic-bound commitments; that residual keeps the entry domain-specific.
Examples¶
Bernoulli success proportions¶
Let \(X_i\) be independent Bernoulli\((p)\) observations with $0<p<1$ and let \(\overline X_n\) be the observed success fraction. The log MGF is \(\Lambda(t)=\log(1-p+pe^t)\), finite for every real \(t\). Its convex conjugate on $0<q<1$ is \(I(q)=q\log(q/p)+(1-q)\log((1-q)/(1-p))\), the binary relative-entropy expression given in Seppäläinen’s Exercise A.9(a). For \(p<q<1\), \(I\) is continuous around \(q\), so the half-line LDP yields \(\lim_{n\to\infty}n^{-1}\log P(\overline X_n\ge q)=-I(q)\). This is an exponential rate, not an exact binomial tail formula.[1][2]
Mapped back: carrier = i.i.d. binary trials and their success fraction; light-tail hypothesis = bounded variables with finite MGF; rate = log-MGF convex conjugate \(D(q\Vert p)\); event geometry = the eligible half-line \([q,\infty)\); conclusion = logarithmic tail rate \(-D(q\Vert p)\), without a prefactor claim.
Gaussian sample averages¶
Let \(X_i\) be independent standard normal observations and ask for \(P(\overline X_n\ge b)\) with fixed \(b>0\). Here \(\Lambda(t)=t^2/2\), \(I(x)=x^2/2\), and the rate at the threshold is \(b^2/2\). Goodman computes the Gaussian tail directly and displays an \(e^{-nb^2/2}\) factor accompanied by an order-\(n^{-1/2}\) prefactor. That direct calculation corroborates the logarithmic conclusion while demonstrating what Cramér's theorem alone leaves unresolved.[3][2]
Mapped back: carrier = i.i.d. continuous Gaussian measurements and their average; light-tail hypothesis = Gaussian MGF finite everywhere; rate = quadratic conjugate \(x^2/2\); event geometry = \([b,\infty)\) with continuous \(I\); conclusion = \(n^{-1}\log P(\overline X_n\ge b)\to-b^2/2\), not equality to \(e^{-nb^2/2}\).
Boundary: ultimate ruin¶
An all-time crossing event such as “there exists a time when cumulative net losses exceed capital” is not an event of one fixed empirical mean at growing \(n\). It may involve related exponential-change-of-measure ideas, but equating its threshold asymptotic with the theorem above would silently exchange the event and the limit being taken. The seed's insurance setting therefore remains a possible downstream application requiring a separate result, not an instance proven here.[4]
Structural Tensions¶
T1 — Sharp exponential rate versus finite-sample probability. The LDP gains generality by retaining only \(n^{-1}\log P\), while a practical finite-\(n\) calculation can be decided by a prefactor or error term that the theorem discards. A precise Gaussian tail restores that detail but no longer offers a distribution-agnostic result. Diagnostic: Is the question about the asymptotic logarithmic rate, or would an order-\(n^{-1/2}\) prefactor change the finite-sample conclusion?[1][3]
T2 — General event validity versus simple scalar equality. Interior/closure bounds tolerate irregular event boundaries but require two rate infima; a one-number limit is simpler yet valid only when those infima meet. Keeping the generality costs an immediately usable equality, while imposing equality without the check risks a false theorem. Diagnostic: Do the interior and closure of the proposed event have the same infimum of \(I\)?[1][2]
Structural–Framed Character¶
This is a structural mathematical theorem with a domain-specific probabilistic carrier. Evaluative weight: it quantifies likelihood rather than judging whether a rare event is good or bad. Human-practice dependence: the theorem's validity depends on mathematical assumptions about a model, not on an institution declaring a category. Institutional origin: the eponym and disciplinary presentation are historical conventions; the probability inequalities are not. Vocabulary travel: “rate,” “tilt,” and “deviation” can migrate into applications, but their mathematical roles require an i.i.d. measure, an exponential moment and a specified event. Import versus recognition: a new field literally instantiates the theorem only after those roles and bounds are verified, not because an observed frequency declines quickly.[1][2]
The abstract operation of converting local distributional information into a global tail rate has portable appeal, but the exact open/closed-set theorem is not a generic principle of all aggregation. Its character: a formal probability result whose structural logic is stable across qualifying laws, while its hypotheses and logarithmic meaning resist metaphorical transfer.
Structural Core vs. Domain Accent¶
The thin skeleton is “aggregate repeated units, transform one-unit information into a limiting cost, and bound exceptional outcomes.” That skeleton can be recognized outside probability, but it is too thin to license this theorem's rate formula. The live Probability prime provides a broad event-and-measure substrate; the live Convex Conjugate entry supplies the specific transform as a compositional parent. Neither alone asserts the two-sided large-deviation principle for i.i.d. empirical means.[1]
Whether the full portable skeleton “local law → convex-dual cost → sharp asymptotic deviation bounds” merits an additional prime across nonprobabilistic domains is a future-prime question, not an admitted parent here. The domain accent—independence, identical distribution, exponential moments, Borel event geometry, and logarithmic probability—does the theorem's identity-bearing work. Remove it and one has an analogy or a different result, not Cramér's theorem.[1][2]
Instantiates / Related Primes¶
This entry presupposes Convex conjugate.
- Probability (live prime): supplies events and measures, but does not by itself assign an asymptotic rate to i.i.d. sample means.
- Convex Conjugate (live domain-specific, proposed composition parent): \(I=\Lambda^*\) is a necessary rate construction in this formulation; the theorem is not a kind of convex conjugate.
- Large Deviations Theory (live domain-specific): a broader field-level neighbor. The present theorem is one result in its subject matter, not a strict subtype of a “theory” node on identity grounds.
The graph proposal is limited to the convex-conjugate presupposition.
Relationships to Other Abstractions¶
Current abstraction Cramér's Theorem (Large Deviations) Domain-specific
Parents (1) — more general patterns this builds on
-
Cramér's Theorem (Large Deviations) presupposes Convex conjugate Domain-specific
Cramér's rate is defined as the convex conjugate of the one-observation log MGF.The sourced formulation constructs the theorem's rate I(x)=sup_t(tx−Lambda(t)) using the convex-conjugate operation. The theorem therefore structurally presupposes that transform, while adding i.i.d. sample means, a light-tail hypothesis, and lower/upper logarithmic probability bounds. It is not a subtype of Convex Conjugate.
Hierarchy path (1) — routes to 1 parentless root
- Cramér's Theorem (Large Deviations) → Convex conjugate → Duality
Neighborhood in Abstraction Space¶
Cramér's Theorem (Large Deviations) sits in a moderately populated region (50th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.
Family — Foundations of Probability & Inference (29 abstractions)
Nearest neighbors
- Large deviations theory — 0.87
- Kolmogorov's Three-Series Theorem — 0.87
- Random Variable — 0.86
- Brownian Skorokhod Embedding — 0.85
- Yule–Simon Distribution — 0.85
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
Cramér's normal-decomposition theorem has a different probability statement despite the same mathematician's name. Cramér's rule solves linear systems, and the Cramér–Rao bound concerns estimator variance; neither is the sample-mean large-deviation principle. Chernoff's bound can share the exponent in a right-tail upper estimate but lacks the theorem's open-set lower assertion. A precise large-deviation asymptotic additionally obtains prefactors or error terms; Cramér's theorem as stated here supplies the logarithmic scale. Ultimate ruin and other first-passage probabilities require their own event-specific analysis rather than an unqualified substitution of \(\overline X_n\).[1][2][3]
References¶
[1] Timo Seppäläinen, Translation Invariant Exclusion Processes (2005), Appendix A.7, printed pp. 192–193, especially Eq. (A.16), Theorem A.9, Proposition A.1 and Exercise A.9(a). Author-text PDF. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y ↩z ↩27 ↩28 ↩29 ↩30
[2] MIT 6.265/15.070J, Lecture 3: Large Deviations Theory. Cramer's Theorem (2013), printed pp. 1–3, especially Theorem 1 and the post-theorem half-line continuity argument. Official course PDF. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u
[3] Jonathan Goodman, Variance Reduction (2005), §4.1, printed pp. 6–7, Gaussian sample-mean rare-tail calculation. Author-course PDF. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j
[4] Victor F. Araman, Theory and Applications of Stochastic Systems, Lecture 10 (2003), §§1–2, random-walk exponential martingale and change of measure. Author-course PDF. This supports the related method only, not an insurance ruin exponent. registry ↩a ↩b ↩c