Normal Number

A real number whose base-b expansion gives every finite length-k digit block its uniform limiting frequency b^-k, for every k, with the base kept explicit.

Version

v3 · 2026-09-06 · History

Domain-specific #

2393

Origin domain

number theory

Subdomain

normal numbers and uniform distribution

Aliases

B Normal Number, Base B Normal Number

Core Idea

A normal number in base \(b\) is a real number whose base-\(b\) digit expansion gives every finite block the limiting frequency expected under the uniform distribution on digits. The base is part of the claim. Fix an integer \(b\ge 2\) and write the fractional part of \(x\) in its canonical base-\(b\) expansion

\[ \{x\}=\sum_{n=1}^{\infty} d_n b^{-n}=0.d_1d_2d_3\ldots{}_b, \qquad d_n\in\{0,1,\ldots,b-1\}, \]

using the representation that does not end in an infinite tail of \(b-1\) digits when two positional representations are possible.^[1] For a word \(w=(w_1,\ldots,w_k)\) of length \(k\), let

\[ N_w(N;x,b)=\#\{1\le j\le N:(d_j,\ldots,d_{j+k-1})=w\}. \]

The occurrences are allowed to overlap. Then \(x\) is normal in base \(b\) exactly when, for every \(k\ge1\) and every \(w\in\{0,\ldots,b-1\}^k\),

\[ \lim_{N\to\infty}\frac{N_w(N;x,b)}{N}=b^{-k}. \]

Using \(N-k+1\) rather than \(N\) in the denominator gives the same limit. The quantifiers are load-bearing: one fixed base; every finite block length; every block of that length; then a limit over successively longer prefixes. Equal single-digit frequencies alone are insufficient, and a long finite prefix that looks uniform is evidence rather than proof.

Becher and Carton present this non-aligned block-frequency formulation and its equivalent aligned, finite-automaton, and uniform-distribution formulations as the modern theory of Borel normality.^[1] The identity is simultaneously concrete and asymptotic. It examines a deterministic digit sequence, but the target \(b^{-k}\) is the frequency that independent uniform base-\(b\) digits would have. It says nothing by itself about how quickly the frequencies converge, whether the digits can be computed, or whether the same real number is normal in a different base.

The entry deliberately retains this base-relative identity. Some authors reserve unqualified “normal number” for a number normal in every integer base \(b\ge2\), while others say “normal in base \(b\)” and call the all-bases property absolute normality. Here normal_number means the former technical predicate with the base declared; absolutely_normal_number is a stronger subtype requiring the predicate simultaneously for every integer base.

This is an autonomous domain-specific abstraction. Its core is a family of convergence statements for empirical digit-block frequencies, but generic Convergence, Measure, and Randomness do not select positional expansions, block alphabets, the \(b^{-k}\) targets, or the all-length quantifier. Those number-theoretic obligations survive catalog decomposition.

Structural Signature

A normal-number claim has these mandatory roles:

• The real number \(x\) — more precisely, its fractional part, since changing the finite integer part does not affect limiting block frequencies.
• The declared integer base \(b\ge2\) — the alphabet size and positional representation. Without it, ordinary normality is underspecified.
• The canonical infinite digit expansion \((d_n)\) — digits in \(\{0,\ldots,b-1\}\), with the dual terminating/repeating representation convention fixed.
• The finite word \(w\) and its length \(k\) — an arbitrary member of the \(b^k\) possible blocks, not merely a selected diagnostic pattern.
• The overlapping occurrence count \(N_w(N;x,b)\) — the number of starting positions for \(w\) in a growing prefix.
• The empirical block frequency — \(N_w(N;x,b)/N\), a prefix-dependent quantity rather than a probability assigned in advance.
• The uniform target \(b^{-k}\) — every word of length \(k\) receives the same limiting share, and the \(b^k\) targets sum to one.
• The universal limiting verdict — convergence must hold for every finite \(k\) and every length-\(k\) word.
• The scope label — base-relative normality, simple normality, or absolute normality must be stated rather than allowed to drift.

The structural sequence is

\[ \text{real }x + \text{ base }b \longrightarrow \text{ infinite digit word} \longrightarrow \text{ all finite-block counts} \longrightarrow \text{ convergence to the uniform word law}. \]

Two equivalent readings help verify the identity. First, normality in base \(b\) is equivalent to simple normality in each power base \(b^k\): grouping \(k\) base-\(b\) digits into one base-\(b^k\) symbol converts block balance into digit balance.^[1] Second, Wall's criterion says that \(x\) is normal in base \(b\) exactly when the fractional parts \(\{b^n x\}\) are uniformly distributed in \([0,1)\).^[1][2] The block and orbit formulations are two views of one property, not two separate candidates.

What It Is Not

• Not simple normality. A number is simply normal in base \(b\) when each single digit has limiting frequency \(1/b\). Normality demands the correct limit for every longer block as well. The periodic binary expansion \(0.010101\ldots{}_2\) is simply normal but fails normality because 00 and 11 never occur.
• Not absolute normality. Absolute normality requires normality in every integer base \(b\ge2\). A proof in base 10 does not silently extend to base 2, base 3, or all bases.^[1]
• Not disjunctivity or richness. A disjunctive expansion contains every finite word, often infinitely many times. Normality implies that occurrence property because \(b^{-k}>0\), but disjunctivity imposes no uniform limiting frequencies and is strictly weaker.^[3]
• Not a normal distribution. The adjective does not refer to a Gaussian law, mean, variance, or bell curve. The target distribution on fixed-length digit words is finite and uniform.
• Not Normal Order of an Arithmetic Function. That analytic-number-theory construct describes the typical relative size of \(f(n)\) on a natural-density-one set. It does not examine positional digits or block frequencies.
• Not algorithmic randomness. A normal expansion has the correct finite-block limiting statistics, but it can be produced by a short deterministic algorithm. Champernowne's computable concatenation is the canonical witness.^[4] Normality is much weaker than unpredictability to every algorithm.
• Not an independence assertion about digit positions. The limiting frequencies agree with those of independent uniform symbols, but deterministic dependencies can persist. Normality is a frequency property of one infinite word.
• Not a finite statistical test. No finite number of examined digits establishes all infinitely many limits. A prefix can pass extensive frequency and correlation tests while the unexamined tail destroys normality.
• Not mere irrationality or transcendence. Rational numbers are excluded because their expansions are eventually periodic, but irrationality alone does not yield normality. Constructed normal numbers can be computable, and other irrational numbers can be non-normal.
• Not the claim that every finite word merely appears once. Presence is qualitative. Normality additionally fixes the asymptotic density of every word.

Scope of Application

Normal numbers belong primarily to number theory, especially metric number theory and uniform distribution modulo one. Borel introduced the property in 1909 and proved that almost every real number, in the sense of Lebesgue measure, is absolutely normal.^[5][1] For a fixed base, the non-normal set therefore has Lebesgue measure zero; intersecting the full-measure normal sets over the countably many integer bases preserves full measure. This is a theorem about the size of a class, not a practical classifier for an individually named real.

The same identity appears in symbolic dynamics and combinatorics on words. A base-\(b\) expansion is an infinite word over a \(b\)-symbol alphabet, and normality makes that word generic for the uniform Bernoulli measure at the level of every finite cylinder block. The map \(T_b(x)=\{bx\}\) shifts the digit expansion, so Wall's orbit-equidistribution criterion connects number representation to an expanding dynamical system.^[1]

Theoretical computer science studies normality through finite automata, selection rules, finite-state gambling, and finite-state compression. These characterizations sharpen the useful but limited phrase “digits look random”: normality controls what finite-state observers can exploit, while stronger algorithmic-randomness notions quantify resistance to unrestricted effective tests. Computable normal and absolutely normal constructions show that statistical balance and algorithmic unpredictability are separable.^[1][6]

Experimental mathematics tests very long prefixes of constants such as \(\pi\). Such studies can quantify discrepancies and reject some non-normal models, but they do not promote empirical agreement to an infinite proof. Bailey, Borwein, Calude, and Dinneen report strong base-16 empirical evidence for \(\pi\) while explicitly framing the conclusion as a model-based assessment rather than a proof of asymptotic normality.^[7]

Clarity

The fastest diagnostic is to expand the quantifiers. “Normal in base 10” means that each digit has frequency \(1/10\), each pair—including overlapping pairs—has frequency \(1/100\), each triple has frequency \(1/1000\), and the analogous statement holds for every finite length. If an argument checks only digits, it proves at most simple normality. If it checks blocks only through a maximum length \(K\), it proves a finite-prefix or order-\(K\) property, not normality.

A second diagnostic separates base, block, and limit. The base chooses the alphabet and representation. The block chooses the finite event being counted. The limit turns prefix counts into an asymptotic property. Omitting any of those roles causes a recognizable error: base ambiguity, selective pattern checking, or extrapolation from finite data.

A third diagnostic asks what “almost every” is doing. Borel's theorem says that the set of exceptions has Lebesgue measure zero. It does not choose an explicit number from the full-measure set, and it does not assign probability one to a mathematical constant whose value is fixed. A named constant needs its own proof.

Manages Complexity

An infinite digit expansion presents infinitely many possible tests: counts for every word, every length, and every prefix cutoff. Normality compresses that family into one reusable verdict. Once established, it guarantees that no fixed finite word has a persistent frequency advantage or deficit. The result lets number theorists, dynamical-systems researchers, and automata theorists move between block counts, base-power simple normality, orbit equidistribution, and finite-state characterizations without re-proving the identity from scratch.

The abstraction also stratifies questions that would otherwise be conflated. Simple normality asks about letters; order-\(k\) normality asks about blocks up to a fixed resolution; base-\(b\) normality quantifies over every block length; absolute normality quantifies again over every integer base; discrepancy asks how fast finite prefixes approach the target. Each layer adds a distinct quantifier or rate obligation.

Finally, it localizes uncertainty about famous constants. A computation can report prefix discrepancy at specified lengths, bases, and block sizes. The unresolved mathematical claim remains visibly separate: whether every one of the required limiting frequencies exists and equals its target. This prevents “billions of apparently balanced digits” from being mistaken for a theorem.

Abstract Reasoning

Normal-number reasoning follows directly from the limit definition.

1. Positive-frequency implication. If \(x\) is normal in base \(b\), every finite base-\(b\) word appears infinitely often because its limiting frequency \(b^{-|w|}\) is positive. Thus normality implies disjunctivity.
2. Converse failure. A sequence can contain every finite word yet place increasingly long zero runs between them. It remains disjunctive while the zero frequency tends toward one, so occurrence completeness does not imply normality.
3. Simple-normality gap. Balancing individual digits does not balance pairs or longer correlations. The alternating binary expansion gives a complete counterexample with one line of counting.
4. Finite-change invariance. Inserting, deleting, or changing finitely many digits changes each fixed block count by at most a bounded amount depending on the edit count and \(|w|\). Dividing by \(N\) makes that difference vanish, so normality is a tail-stable limiting property.
5. Rational exclusion. An eventually periodic base-\(b\) expansion has only finitely many distinct blocks beyond its period structure and cannot realize all \(b^k\) words with frequency \(b^{-k}\) for every \(k\). Therefore no rational is normal in any integer base.^[6]
6. Base-power transfer. Reblocking \(k\) base-\(b\) digits as one base-\(b^k\) symbol turns aligned word counts into digit counts. The equivalence requires every \(k\), not one convenient power.
7. Orbit transfer. Multiplication by \(b\) modulo one shifts the expansion. Uniform distribution of \(\{b^n x\}\) is therefore equivalent to balanced visits to all base-\(b\) cylinder intervals.^[1]
8. No convergence-rate inference. The definition supplies a limit but no universal rate. A normal number may have prefixes with substantial local imbalance; discrepancy is an additional quantitative property.

These deductions reveal exactly what a proof must deliver. It can work combinatorially by controlling all word counts, dynamically by proving orbit equidistribution, or through an equivalent finite-state criterion. A simulation alone delivers none of the universal limiting quantifiers.

Knowledge Transfer

The role mapping travels literally across several mathematical practices:

• real number in base \(b\) \(\leftrightarrow\) infinite word over a \(b\)-symbol alphabet;
• digit block \(\leftrightarrow\) finite word or cylinder event;
• prefix count \(\leftrightarrow\) empirical visit count;
• target \(b^{-k}\) \(\leftrightarrow\) uniform Bernoulli mass of a length-\(k\) cylinder;
• multiplication by \(b\) modulo one \(\leftrightarrow\) left shift of the digit word;
• block discrepancy \(\leftrightarrow\) deviation of an empirical measure from its uniform target;
• simple/base-relative/absolute normality \(\leftrightarrow\) one-letter/all-word/all-base quantifier layers.

This mapping supports transfer among number theory, symbolic dynamics, ergodic theory, and finite-state computation without changing the identity. In each case, the object is one infinite sequence whose empirical frequencies converge to a prescribed uniform word law.

Transfer to ordinary statistical data is more limited. A finite random sample can have approximately uniform category counts, but no infinite-prefix limit has been established. A pseudorandom generator can pass a selected battery without producing a normal infinite sequence, and a normal computable sequence can be unsuitable for cryptography because future digits remain predictable. The generic primes Convergence, Measure, and Randomness carry the broader lessons; the term Normal Number should remain attached to positional expansions and their exact quantifiers.

Examples

Simple but not normal. Consider \(x=0.010101\ldots{}_2=1/3\). Zeros and ones each have limiting frequency \(1/2\), so the expansion is simply normal in base 2. The pair 00 has frequency zero instead of \(1/4\), as does 11. The example demonstrates why checking \(k=1\) cannot settle \(k=2\), much less every \(k\).

Champernowne's base-10 construction. Champernowne formed

\[ C_{10}=0.123456789101112131415\ldots \]

by concatenating the decimal representations of the positive integers and proved that it is normal in base 10.^[4] The construction is explicit and computable, so it simultaneously supplies a positive normality example and a counterexample to the claim that normality entails algorithmic unpredictability. The theorem is base-specific; it does not establish absolute normality.

Disjunctive but non-normal construction. Enumerate all finite binary words as \(w_1,w_2,\ldots\) and form

\[ w_1\,0^{1!}\,w_2\,0^{2!}\,w_3\,0^{3!}\cdots. \]

Every finite word occurs because it appears among the \(w_i\), so the sequence is disjunctive. The factorially growing zero runs dominate the total length, preventing the uniform block frequencies required for normality. The construction isolates presence from density.^[3]

Almost every versus a named example. Borel's theorem gives absolute normality to a Lebesgue-full set of real numbers.^[5] Becher and Figueira give a computable construction of an absolutely normal number, building on Sierpiński's effective work.^[6] These are different achievements: one measures the size of the class; the other specifies an individual member.

The \(\pi\) nonexample. Very large computed prefixes of \(\pi\) have digit and block statistics compatible with normality in tested bases.^[7] No known proof establishes that \(\pi\) is normal in even one integer base, and the same caution applies to \(e\) and \(\sqrt2\).^[1][2] A passing finite experiment is not an instance certificate.

Structural Tensions

Genericity versus explicitness. Absolute normality holds almost everywhere in Lebesgue measure, yet proving the property for familiar constants remains extraordinarily difficult. Full measure makes normality typical as a class without making any particular named member easy to certify.

Uniform statistics versus computable structure. Normality reproduces the limiting frequencies of uniform random digits, while Champernowne-type constructions generate every digit deterministically. The failure mode is to interpret one statistical projection as total randomness or cryptographic unpredictability.

Infinite verdict versus finite evidence. Every computation examines a finite prefix and finitely many block lengths. Normality quantifies over unbounded prefixes and all finite lengths. More data can strengthen empirical support and shrink observed discrepancy, but it cannot close the logical gap without a proof controlling the tail.

Base-relative identity versus unqualified language. A number can be studied as normal to one base, while absolute normality quantifies over all integer bases. Publications do not always use the unqualified phrase consistently. Every catalog claim and example must preserve its base scope.

Aligned convenience versus overlapping definition. Grouping digits into nonoverlapping \(k\)-blocks makes base-\(b^k\) reasoning convenient. Modern normality counts occurrences at every starting position. The equivalence is a theorem; silently switching counting conventions without the theorem can weaken a proof.

Qualitative limit versus convergence speed. Two numbers can both be normal while their finite-prefix discrepancies decay at very different rates. The label preserves the limiting destination and discards the path. Applications needing finite-sample quality must add discrepancy bounds.

Structural–Framed Character

Normal Number is structural-leaning but domain-bound. Its verdict is formal and non-evaluative: a countable family of empirical frequency sequences either converges to declared uniform targets or does not. No institution or human preference determines the result, and the block-frequency skeleton transfers cleanly to symbolic dynamics.

The biological, legal, or organizational framing tests are irrelevant; the binding frame is mathematical representation. The object must be a real number or infinite word, the alphabet comes from an integer positional base, and the target law is exactly \(b^{-k}\). Removing that apparatus leaves generic equidistribution or convergence, not the normal-number abstraction. This supports domain-specific rather than prime classification despite the strong structural character.

Structural Core vs. Domain Accent

The structural core is simultaneous empirical convergence: for every member of a countable family of finite patterns, observed prefix frequency approaches a prescribed reference weight. That residue maps to Convergence and uses Measure to express target and empirical distributions. Randomness supplies a motivating comparison but not an equivalence.

The domain accent fixes the object to a real number's positional expansion, the alphabet to \(b\) digits, the patterns to contiguous finite blocks, the target to \(b^{-|w|}\), and the quantifiers to all words and lengths. It also introduces base conversion, dual expansion conventions, simple and absolute variants, and classical explicit-construction problems.

A corpus whose word frequencies approach a model distribution may share the structural core, but it is not a normal number. A finite pseudorandom sequence may have low discrepancy without possessing an infinite normality verdict. Those cases should route to the generic primes or their own field abstractions rather than broadening this entry.

Instantiates / Related Primes

Convergence is the load-bearing DAG relation. For every word \(w\), the empirical frequency sequence \(N_w(N;x,b)/N\) must converge to \(b^{-|w|}\). Removing convergence turns normality into finite coincidence. Because a normal number is an object defined through a family of limit processes rather than literally a kind of convergence, the proposed relation is strict composition with part_of flavor.

Measure supplies the uniform Bernoulli weights on digit cylinders and the Lebesgue-measure meaning of Borel's “almost every” theorem. It is related and presupposed by some formulations, but it is not the candidate's minimal identity-bearing parent.

Randomness explains the intended resemblance to uniform random digits. Normality is a weak statistical projection and cannot inherit the prime's full unpredictability obligations; computable normal constructions block subsumption.

Asymptotic Behavior captures the deliberate disregard of finite prefixes and transient imbalance. Discreteness supplies the finite digit alphabet and word counts. Both are supporting structures rather than additional minimal parents.

One proposal-only edge is therefore recommended: domain_specific:normal_number contains prime:convergence as a strict identity-bearing component. No live DAG mutation is authorized.

Relationships to Other Abstractions

Current abstraction Normal Number Domain-specific

Parents (1) — more general patterns this builds on

• Normal Number is part of Convergence Prime

  Convergence is the load-bearing DAG relation.

Hierarchy path (1) — routes to 1 parentless root

• Normal Number → Convergence

Neighborhood in Abstraction Space

Normal Number sits in a sparse region of the domain-specific corpus (83^rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

• Cyclic Number — 0.84
• Multiplicative Digital Root — 0.82
• Champernowne constant — 0.81
• Normal Order of an Arithmetic Function — 0.80
• Primefree Sequence — 0.80

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With

The frozen top semantic neighbor, Normal Order of an Arithmetic Function, is a homonym within number theory. Its object is an arithmetic function \(f(n)\), its comparator is a typical-size function \(g(n)\), and its limit ignores a natural-density-zero exceptional set. Normal Number instead fixes one real, one base, and every finite digit block. Neither aliases nor covers the other.

Natural Number is a basic kind of integer, not a statistical property of a real's positional expansion. Hermite Normal Form is a canonical integer-matrix form. Collectionwise Normal Space is a separation property in topology. Normalization of Deviance, Normal Fault, and Probability Distribution are further lexical neighbors without the block-frequency mechanism.

Measure and Probability can formulate the uniform reference law, but a probability space alone contains no infinite digit word whose empirical counts converge. Randomness adds predictor-relative unpredictability that normality does not guarantee. Mathematical Induction can appear in a proof but is not an identity neighbor.

Vocabulary must preserve the hierarchy: simply_normal_number is a weaker variant; normal_number_in_base_b, b_normal_number, and base_b_normal_number are base-explicit surfaces for this candidate; absolutely_normal_number is the all-integer-bases subtype; disjunctive_number or rich_number is weaker occurrence-only vocabulary; and Champernowne_number is a construction and instance, not an alias.

References

[1] Verónica Becher and Olivier Carton, “Normal Numbers and Computer Science,” survey chapter, June 15, 2018. https://www.irif.fr/~carton/Enseignement/MPRI/Dynam-symbol/Documents/chapter.pdf registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j

[2] Christoph Aistleitner, “Normal Numbers and the Normality Measure,” Combinatorics, Probability and Computing 22, no. 3 (2013): 342–345. https://doi.org/10.1017/S0963548313000084 registry ↩a ↩b

[3] Cristian S. Calude, Lutz Priese, and Ludwig Staiger, “Disjunctive Sequences: An Overview,” CDMTCS Research Report 063, University of Auckland, 1997. https://www.cs.auckland.ac.nz/research/groups/CDMTCS/researchreports/063cris.pdf registry ↩a ↩b

[4] D. G. Champernowne, “The Construction of Decimals Normal in the Scale of Ten,” Journal of the London Mathematical Society s1-8, no. 4 (1933): 254–260. https://doi.org/10.1112/jlms/s1-8.4.254 registry ↩a ↩b

[5] Émile Borel, “Les probabilités dénombrables et leurs applications arithmétiques,” Rendiconti del Circolo Matematico di Palermo 27 (1909): 247–271. https://doi.org/10.1007/BF03019651 registry ↩a ↩b

[6] Verónica Becher and Santiago Figueira, “An Example of a Computable Absolutely Normal Number,” Theoretical Computer Science 270, nos. 1–2 (2002): 947–958. https://doi.org/10.1016/S0304-3975(01)00170-0 registry ↩a ↩b ↩c

[7] David H. Bailey, Jonathan M. Borwein, Cristian S. Calude, and Michael J. Dinneen, “An Empirical Approach to the Normality of \(\pi\),” Experimental Mathematics 21, no. 4 (2012): 375–384. https://doi.org/10.1080/10586458.2012.665333 registry ↩a ↩b