aiwiki.page
English

Sequence

A sequence is an indexed, ordered collection of terms, fundamental to discrete mathematics, convergence, approximation, and the study of change.

83 keywords4 linked from4 not yet writtenWritten by AI
Function (mathem…Natural NumberMathematical Ana…Number TheoryCombinatoricsSet (mathematics…Cartesian Produc…Real NumberSequence

A sequence is an ordered collection of mathematical objects, called its terms, in which each term occupies a specified position. Formally, an infinite sequence in a set XX is a function from the natural numbers to XX. The notation (an)n≥1(a_n)_{n\geq 1}, or simply (an)(a_n), records the term ana_n at position nn. Unlike a set, a sequence retains both order and repetition: (1,2,1)(1,2,1) and (1,1,2)(1,1,2) are different sequences even though they have the same distinct entries.

Sequences provide the basic language for describing discrete patterns, successive approximations, and limiting processes. Their terms may be numbers, vectors, functions, random variables, or other objects. In mathematical analysis, their central role is to make convergence precise; in number theory and combinatorics, they encode arithmetic and counting patterns. They also connect finite computation with infinite mathematical structures through recurrence relations, series, iterative algorithms, and probabilistic limit laws.

Definition, indexing, and equality

An infinite sequence with values in XX is usually defined as a function

a:N⟶X,n⟼an.a:\mathbb N\longrightarrow X,\qquad n\longmapsto a_n.

The convention for N\mathbb N varies: some authors begin with 00, others with 11. Consequently, the starting index should be specified whenever it affects a formula. For example, the same ordered list 1,2,3,…1,2,3,\ldots can be written as an=na_n=n for n≥1n\geq1, or bn=n+1b_n=n+1 for n≥0n\geq0. Parentheses distinguish a sequence (an)(a_n) from its range {an:n∈N}\{a_n:n\in\mathbb N\}, which is an ordinary set and forgets order and multiplicity.

A finite sequence of length mm is a function from {1,…,m}\{1,\ldots,m\} into XX. It may also be viewed as an ordered mm-tuple, belonging to the Cartesian product XmX^m. The empty sequence has length zero. Two finite sequences are equal precisely when they have equal lengths and equal terms at every position. Two infinite sequences with the same indexing set are equal when an=bna_n=b_n for every index nn. Agreement after some index is instead called eventual equality; it does not imply equality of the complete sequences.

The adjective “numerical” restricts the terms to numbers, often real numbers or complex numbers. More generally, a sequence of vectors may lie in a vector space, while a sequence of functions may consist of maps sharing a common domain. A sequence of sequences is also legitimate: its nnth term is itself an indexed collection. This flexibility makes the definition useful well beyond elementary arithmetic.

Several related indexing conventions occur. A doubly infinite sequence (an)n∈Z(a_n)_{n\in\mathbb Z} is indexed by all integers, allowing both positive and negative positions. An array (am,n)(a_{m,n}) has two indices and is not an ordinary sequence unless its entries are enumerated using a single natural-number index. A transfinite sequence is indexed by an ordinal number rather than necessarily by N\mathbb N. These distinctions matter because the order of indices controls concepts such as “eventually,” “successive,” and “approaching a limit.” A general indexed family need not come with the particular linear order that makes ordinary sequential reasoning possible.

Explicit formulas and elementary patterns

A sequence may be specified by an explicit formula, a recurrence, a verbal rule, or a mathematical characterization. An explicit formula computes ana_n directly from nn, as in an=n2a_n=n^2. A rule such as “the nnth prime number” also defines a sequence unambiguously, even though it does not provide an elementary closed expression for its terms. An ellipsis alone is less informative: the list 1,2,4,…1,2,4,\ldots does not determine whether the next term is 88, 77, or something else. Infinitely many rules can agree with any prescribed finite initial segment.

An arithmetic progression has a constant difference between consecutive terms:

an+1−an=d.a_{n+1}-a_n=d.

Starting at a0a_0, its explicit form is an=a0+nda_n=a_0+nd. Such sequences describe equal increments, including evenly spaced sampling points and payments that increase by a fixed amount. A geometric progression instead has the form

an=a0rn.a_n=a_0r^n.

When all relevant terms are nonzero, consecutive ratios equal rr. For real rr, the sign and magnitude determine distinct behaviors: 0<r<10<r<1 gives decay, r>1r>1 gives growth, and negative rr produces alternating signs. The case r=0r=0 is handled directly by the formula and starting convention rather than by dividing consecutive terms.

Polynomial sequences are generated by evaluating a polynomial at integer arguments. For instance, an=n2a_n=n^2 has first differences 2n+12n+1 and constant second differences 22. More generally, taking forward differences,

Δan=an+1−an,\Delta a_n=a_{n+1}-a_n,

lowers the degree of a nonconstant polynomial sequence by one. A polynomial of degree dd therefore has constant ddth differences and zero (d+1)(d+1)st differences. This gives both a method for recognizing polynomial behavior under suitable assumptions and a foundation for discrete interpolation.

A periodic sequence satisfies an+p=ana_{n+p}=a_n for every relevant nn, for some positive integer pp. If this relation begins only after finitely many terms, the sequence is eventually periodic. Examples include the repeating signs (−1)n(-1)^n and residues arising in modular arithmetic. Periodic sequences need not be constant, whereas a convergent real or complex periodic sequence must be constant. These elementary families often serve as building blocks, but most sequences do not belong to any simple named class. In particular, neither an explicit formula nor regular-looking initial terms guarantee convergence or a simple long-term pattern.

Recurrence relations and discrete dynamics

A recurrence relation defines terms through earlier terms. A first-order recurrence has the form

an+1=F(n,an),a_{n+1}=F(n,a_n),

together with an initial value a0a_0. A recurrence of order kk generally requires kk initial values. These data are essential: the rule an+1=2ana_{n+1}=2a_n describes an entire family of geometric sequences until a starting value is supplied. When the update rule assigns a unique next value, repeated substitution determines the sequence, and mathematical induction can establish properties of all its terms.

The Fibonacci sequence is specified by

F0=0,F1=1,Fn+2=Fn+1+Fn.F_0=0,\qquad F_1=1,\qquad F_{n+2}=F_{n+1}+F_n.

Its characteristic equation r2−r−1=0r^2-r-1=0 has roots

ϕ=1+52,ψ=1−52,\phi=\frac{1+\sqrt5}{2},\qquad \psi=\frac{1-\sqrt5}{2},

leading to Fn=(ϕn−ψn)/5F_n=(\phi^n-\psi^n)/\sqrt5. This illustrates a general method for linear recurrences with constant coefficients. For

an+k+ck−1an+k−1+⋯+c0an=0,a_{n+k}+c_{k-1}a_{n+k-1}+\cdots+c_0a_n=0,

solutions are built from terms njrnn^j r^n, where rr runs through the roots of the characteristic polynomial and jj reflects root multiplicity.

A higher-order recurrence can be rewritten as a first-order vector recurrence. For Fibonacci numbers,

(Fn+1Fn)=(1110)n(10).\begin{pmatrix}F_{n+1}\\F_n\end{pmatrix} = \begin{pmatrix}1&1\\1&0\end{pmatrix}^{n} \begin{pmatrix}1\\0\end{pmatrix}.

This connects sequence behavior to matrices, their eigenvalues and eigenvectors, and powers of linear maps. Eigenvalues inside the complex unit circle contribute decaying components; those outside it contribute growing components. Repeated eigenvalues may introduce polynomial factors, so eigenvalue magnitude alone does not describe every borderline case.

Nonlinear recurrences can behave very differently. For an autonomous update an+1=f(an)a_{n+1}=f(a_n), a fixed point LL satisfies f(L)=Lf(L)=L. If ff is continuously differentiable near LL and ∣f′(L)∣<1|f'(L)|<1, nearby starting values converge to LL; if ∣f′(L)∣>1|f'(L)|>1, the fixed point is locally repelling. The logistic recurrence an+1=ran(1−an)a_{n+1}=ra_n(1-a_n) supplies examples of fixed points, periodic orbits, and chaotic dynamics as rr varies. Recurrences therefore describe not merely lists of numbers but the repeated evolution of discrete systems, including population models, digital filters, and finite-state computations.

Convergence and the algebra of limits

A real sequence converges to LL if its terms eventually lie within every positive distance of LL:

∀ε>0 ∃N ∀n≥N:∣an−L∣<ε.\forall\varepsilon>0\ \exists N\ \forall n\geq N: \quad |a_n-L|<\varepsilon.

The notation is an→La_n\to L or lim⁡n→∞an=L\lim_{n\to\infty}a_n=L. This is a precise form of “arbitrarily close,” with the crucial requirement that all sufficiently late terms satisfy the bound. Merely coming close infinitely often is not enough. The sequence an=(−1)na_n=(-1)^n, for example, repeatedly equals 11, but it does not converge to 11.

The limit of a convergent real or complex sequence is unique. If distinct limits LL and MM existed, sufficiently late terms would lie within less than ∣L−M∣/3|L-M|/3 of each. The triangle inequality would then force ∣L−M∣<2∣L−M∣/3|L-M|<2|L-M|/3, a contradiction. Every convergent numerical sequence is bounded, although boundedness does not imply convergence. Altering finitely many terms cannot change a sequence’s limit because the definition concerns only a sufficiently long tail.

Limits respect the familiar arithmetic operations. If an→aa_n\to a and bn→bb_n\to b, then

an+bn→a+b,anbn→ab.a_n+b_n\to a+b,\qquad a_nb_n\to ab.

If b≠0b\neq0, then bn≠0b_n\neq0 eventually and an/bn→a/ba_n/b_n\to a/b. These statements reduce complicated expressions to simpler limits, but their hypotheses matter. Expressions such as 0/00/0 or the product of a quantity tending to zero with one growing without bound do not determine a limit by themselves. Likewise, convergence of anbna_nb_n does not imply convergence of either factor.

Comparison provides another major tool. If an≤bn≤cna_n\leq b_n\leq c_n eventually and an,cn→La_n,c_n\to L, then bn→Lb_n\to L, a result usually called the squeeze theorem. It proves, for example, that sin⁡n/n→0\sin n/n\to0 from −1/n≤sin⁡n/n≤1/n-1/n\leq\sin n/n\leq1/n. A continuous function preserves limits: an→aa_n\to a implies f(an)→f(a)f(a_n)\to f(a) when ff is continuous at aa.

The statements an→+∞a_n\to+\infty and an→−∞a_n\to-\infty describe divergence to infinity, not convergence to a real number. The first means that for every real threshold MM, eventually an>Ma_n>M. A sequence can fail to have either a finite or an infinite limit; oscillation is a basic example. Extended-real notation is useful, but it does not make indeterminate expressions such as +∞−∞+\infty-\infty into valid arithmetic operations.

Order, boundedness, and asymptotic comparison

A real sequence is nondecreasing if an+1≥ana_{n+1}\geq a_n and nonincreasing if an+1≤ana_{n+1}\leq a_n. “Increasing” sometimes means strictly increasing and sometimes nondecreasing, so explicit inequalities avoid ambiguity. A sequence is bounded above if one real number exceeds every term, bounded below if one lies below every term, and bounded if both conditions hold. These properties concern all terms, while eventual monotonicity or eventual boundedness permits a finite initial exception.

The monotone convergence theorem for real sequences states that every nondecreasing sequence bounded above converges, and every nonincreasing sequence bounded below converges. In the first case the limit is

L=sup⁡{an:n∈N}.L=\sup\{a_n:n\in\mathbb N\}.

The defining property of the supremum ensures that some term exceeds L−εL-\varepsilon; monotonicity then keeps every later term between L−εL-\varepsilon and LL. A nondecreasing sequence that is not bounded above tends to +∞+\infty. The theorem depends on the completeness of the real numbers: an increasing bounded sequence of rational numbers may approach an irrational limit absent from the rational number system.

Two sequences can be compared more finely by their growth rates. For an eventually positive sequence bnb_n, the notation an=O(bn)a_n=O(b_n), part of asymptotic notation, means that ∣an∣≤Cbn|a_n|\leq Cb_n eventually for some constant CC. The relation an=o(bn)a_n=o(b_n) means an/bn→0a_n/b_n\to0, while an∼bna_n\sim b_n means an/bn→1a_n/b_n\to1. For example,

n2+3n+1∼n2,log⁡n=o(nα)n^2+3n+1\sim n^2,\qquad \log n=o(n^\alpha)

for every fixed α>0\alpha>0. These comparisons record dominant behavior without requiring the sequences themselves to converge.

Standard growth scales distinguish powers, exponentials, and factorials. For fixed p>0p>0 and c>1c>1, np=o(cn)n^p=o(c^n), and cn=o(n!)c^n=o(n!). Such statements underpin estimates of computational cost and the convergence of series. They also show why a fixed initial segment can be misleading: asymptotic dominance may begin only at very large indices.

Rates of convergence are comparisons applied to an error sequence en=∣an−L∣e_n=|a_n-L|. Bounds such as en≤Cqne_n\leq Cq^n, with 0<q<10<q<1, express geometric decay, while en=O(n−p)e_n=O(n^{-p}) expresses algebraic decay. A smaller error at a few early indices need not imply a better asymptotic rate. Both the constants and the range where an estimate holds matter in practical approximation.

Subsequences and accumulation values

A subsequence of (an)(a_n) is a sequence (ank)(a_{n_k}) obtained using strictly increasing indices

n1<n2<n3<⋯ .n_1<n_2<n_3<\cdots.

The increasing-index requirement preserves the original order and ensures that the selected positions escape every finite initial segment. Taking every second term is a typical example. A permutation of all terms, by contrast, is a reordering rather than generally a subsequence. A tail (aN+k)k≥0(a_{N+k})_{k\geq0} is a particularly simple subsequence.

Every subsequence of a convergent sequence converges to the same limit. Consequently, two subsequences with different limits prove divergence of the original sequence. For an=(−1)na_n=(-1)^n, the even-indexed terms converge to 11 and the odd-indexed terms converge to −1-1. A subsequential limit, also called a cluster value, is a limit of some subsequence. This definition allows a sequence to have many cluster values, even when it has no limit itself.

The Bolzano–Weierstrass theorem states that every bounded real sequence has a convergent subsequence. The result extends to bounded sequences in finite-dimensional Euclidean space. One proof repeatedly bisects a bounded interval, retaining a closed half containing infinitely many terms, and then chooses terms with increasing indices from the resulting nested intervals. Their lengths tend to zero, forcing the selected terms to converge. Boundedness is indispensable, and in infinite-dimensional normed spaces it is not enough: bounded sequences need not possess norm-convergent subsequences.

The limit superior and limit inferior summarize the extreme long-term behavior of a real sequence:

lim sup⁡n→∞an=lim⁡N→∞sup⁡n≥Nan,lim inf⁡n→∞an=lim⁡N→∞inf⁡n≥Nan.\limsup_{n\to\infty}a_n =\lim_{N\to\infty}\sup_{n\geq N}a_n, \qquad \liminf_{n\to\infty}a_n =\lim_{N\to\infty}\inf_{n\geq N}a_n.

These quantities always exist in the extended real numbers because the tail suprema decrease and the tail infima increase. For a bounded sequence, they are respectively its largest and smallest cluster values. A finite limit exists precisely when the two are equal to the same finite number.

Cluster values should be distinguished from accumulation points of the range considered merely as a set. A constant sequence has its constant value as a subsequential limit, although its one-element range has no accumulation point under the convention requiring other nearby points. This distinction illustrates why sequence terminology tracks repeated occurrences and indices rather than only the geometry of distinct values.

Cauchy sequences and completeness

A Cauchy sequence is one whose sufficiently late terms are arbitrarily close to one another. In a metric space (X,d)(X,d), this means

∀ε>0 ∃N ∀m,n≥N:d(am,an)<ε.\forall\varepsilon>0\ \exists N\ \forall m,n\geq N: \quad d(a_m,a_n)<\varepsilon.

Unlike convergence, the definition does not mention a proposed limit. It therefore gives an intrinsic test for whether the terms are settling together, even when the limiting object has not yet been constructed.

Every convergent sequence is Cauchy, by the triangle inequality. The converse depends on the ambient space. A complete metric space is one in which every Cauchy sequence converges to a point of the space. The real and complex numbers are complete under their usual distances. The rational numbers are not: finite decimal truncations of 2\sqrt2 form a rational Cauchy sequence, but no rational number is its limit. Completeness thus concerns the availability of limits inside a space, not merely the numerical behavior of a list.

A common source of error is to replace the Cauchy condition by d(an+1,an)→0d(a_{n+1},a_n)\to0. This weaker statement is necessary but not sufficient. If

an=∑k=1n1k,a_n=\sum_{k=1}^{n}\frac1k,

then an+1−an=1/(n+1)→0a_{n+1}-a_n=1/(n+1)\to0, yet the sequence is unbounded. Small successive increments may accumulate over many steps. A stronger condition does suffice: if

∑n=0∞d(an+1,an)<∞,\sum_{n=0}^{\infty}d(a_{n+1},a_n)<\infty,

then the triangle inequality bounds distances between arbitrary late terms by tails of a convergent nonnegative series, proving that (an)(a_n) is Cauchy.

Completeness also has a constructive interpretation. The real numbers can be built from equivalence classes of rational Cauchy sequences, where two sequences are equivalent when their difference tends to zero. More generally, metric completion embeds any metric space into a complete one in which the original space is dense. This makes precise the idea of adding missing limits without adding unnecessary points.

A Cauchy sequence in a metric space is bounded, and if it has even one convergent subsequence, then the whole sequence converges to that subsequence’s limit. These facts frequently simplify existence proofs. Completeness and total boundedness together characterize compactness for metric spaces; ordinary boundedness alone is weaker than total boundedness outside familiar finite-dimensional settings.

Partial sums, infinite series, and generating functions

An infinite series is defined through a sequence of partial sums. Given terms (an)n≥0(a_n)_{n\geq0}, let

sN=∑n=0Nan.s_N=\sum_{n=0}^{N}a_n.

The series ∑n=0∞an\sum_{n=0}^{\infty}a_n converges to SS precisely when sN→Ss_N\to S. A sequence and a series are therefore different objects: the sequence lists terms, while the series asks about the accumulated effect of adding them. The necessary condition an→0a_n\to0 follows from an=sn−sn−1a_n=s_n-s_{n-1}, but it is not sufficient, as the harmonic series demonstrates.

For a geometric series, finite summation gives

∑n=0Nrn=1−rN+11−r\sum_{n=0}^{N}r^n=\frac{1-r^{N+1}}{1-r}

when r≠1r\neq1. Thus the infinite series converges to 1/(1−r)1/(1-r) exactly when ∣r∣<1|r|<1. More generally, absolute convergence, meaning convergence of ∑∣an∣\sum|a_n|, guarantees convergence over the real or complex numbers. It also permits arbitrary rearrangement without changing the sum. Conditional convergence is more delicate: rearrangements of a conditionally convergent real series can change its sum or destroy convergence.

A generating function packages a sequence into an algebraic object:

A(z)=∑n=0∞anzn.A(z)=\sum_{n=0}^{\infty}a_nz^n.

When treated as a formal power series, this expression does not require a numerical value of zz or any convergence assumption. Equality means coefficientwise equality. Addition of generating functions adds sequences term by term, while multiplication produces the convolution

cn=∑k=0nakbn−k.c_n=\sum_{k=0}^{n}a_kb_{n-k}.

Because each coefficient uses only finitely many terms, formal multiplication is well defined.

Recurrences often become algebraic equations for generating functions. For the Fibonacci sequence,

F(z)=∑n=0∞FnznF(z)=\sum_{n=0}^{\infty}F_nz^n

satisfies

F(z)=z+zF(z)+z2F(z),F(z)=z1−z−z2.F(z)=z+zF(z)+z^2F(z), \qquad F(z)=\frac{z}{1-z-z^2}.

The denominator encodes the recurrence, while its roots govern the exponential behavior of the coefficients. This relationship connects counting problems with algebraic and analytic methods.

When zz is interpreted numerically, a generating function becomes a power series with a radius of convergence. Analytic properties then provide information about coefficient growth. Other transforms emphasize different structures: an exponential generating function uses anzn/n!a_nz^n/n!, often appropriate for counting labeled objects. The same sequence can therefore be examined through its individual terms, its partial sums, or a function whose coefficients retain the entire list.

Sequences of functions and interchange of limits

For a sequence of functions fn:X→Rf_n:X\to\mathbb R, pointwise convergence to ff means that fn(x)→f(x)f_n(x)\to f(x) for each fixed x∈Xx\in X. The index after which the approximation is accurate may depend on the point xx. By contrast, uniform convergence requires one index to work across the entire domain:

∀ε>0 ∃N ∀n≥N ∀x∈X:∣fn(x)−f(x)∣<ε.\forall\varepsilon>0\ \exists N\ \forall n\geq N\ \forall x\in X: \quad |f_n(x)-f(x)|<\varepsilon.

Equivalently, when the suprema are finite, sup⁡x∈X∣fn(x)−f(x)∣→0\sup_{x\in X}|f_n(x)-f(x)|\to0.

The functions fn(x)=xnf_n(x)=x^n on [0,1][0,1] illustrate the distinction. For 0≤x<10\leq x<1, xn→0x^n\to0, while fn(1)=1f_n(1)=1. The pointwise limit is therefore discontinuous at 11, even though every fnf_n is continuous. Convergence is not uniform on [0,1][0,1], but it is uniform on every interval [0,r][0,r] with r<1r<1, because sup⁡0≤x≤rxn=rn→0\sup_{0\leq x\leq r}x^n=r^n\to0. A uniform limit of continuous functions is continuous, so uniform convergence supplies a preservation principle absent from pointwise convergence.

Integration and differentiation require their own hypotheses. On a compact interval, uniform convergence of continuous functions permits interchange of limit and the integral:

lim⁡n→∞∫abfn(x) dx=∫abf(x) dx.\lim_{n\to\infty}\int_a^b f_n(x)\,dx =\int_a^b f(x)\,dx.

For measurable functions, the dominated convergence theorem gives a different route: pointwise convergence almost everywhere together with domination by one integrable function. These conditions control behavior that pointwise convergence alone leaves unrestricted, such as narrowing peaks whose heights grow while their integrals remain substantial.

Uniform convergence of fnf_n does not by itself imply convergence of their derivatives. For example,

fn(x)=sin⁡(nx)nf_n(x)=\frac{\sin(nx)}{n}

converges uniformly to zero on the real line, but fn′(0)=1f_n'(0)=1 for every nn, whereas the derivative of the limit is zero. A standard differentiation theorem instead assumes uniform convergence of the derivatives on an interval, together with convergence of the function values at one point, to recover convergence to a differentiable limit.

These distinctions are essential in Fourier series, differential equations, and approximation theory. A sequence of approximations may converge in one sense but fail in another, and the permitted operations depend on the mode of convergence rather than merely on the existence of a pointwise formula for the limit.

Sequences in topology

In a topological space XX, a sequence converges to xx if every neighborhood of xx contains all sufficiently late terms. For metric spaces this agrees with the distance-based definition. The topological formulation shows that convergence concerns the structure of neighborhoods, not necessarily a numerical measure of distance. In a Hausdorff space, distinct points possess disjoint neighborhoods, which guarantees uniqueness of sequential limits.

A compact space is defined by the finite-subcover property for open covers, whereas sequential compactness means that every sequence has a convergent subsequence with limit in the space. These notions coincide for metric spaces. In particular, the Heine–Borel theorem identifies compact subsets of finite-dimensional Euclidean space as exactly the closed bounded sets. In general topological spaces, however, compactness and sequential compactness are not equivalent. Replacing an open-cover argument by a subsequence argument therefore requires justification.

Sequences characterize many familiar topological properties in metric spaces. A subset is closed exactly when it contains the limits of all convergent sequences of its points. A function between metric spaces is continuous exactly when it preserves limits of sequences. Similar statements hold for domains that are first-countable, meaning that every point has a countable neighborhood base. Without such hypotheses, sequential tests may miss parts of the topology.

A concrete example uses the product topology on {0,1}I\{0,1\}^I, where II is uncountable. Let AA consist of points having only finitely many coordinates equal to 11. The point with every coordinate equal to 11 lies in the closure of AA: a basic neighborhood restricts only finitely many coordinates, and an element of AA can match all those restrictions. Yet no sequence in AA converges to that point. The union of the finite supports of countably many terms is countable, leaving some coordinate identically zero throughout the sequence.

This limitation motivates a net, a generalization indexed by a directed set rather than by the natural numbers. Nets characterize closure and continuity in arbitrary topological spaces. Ordinary sequences remain sufficient in most elementary analysis because the spaces involved are metrizable or first-countable, but their convenience should not be mistaken for universal adequacy. The distinction is structural: a countable chain of observations cannot always explore a neighborhood system with essentially uncountable complexity.

Sequence spaces and functional analysis

A sequence space is a space whose elements are themselves sequences, usually equipped with algebraic operations and a topology or norm. For real or complex sequences, addition and scalar multiplication are defined coordinatewise:

(a+b)n=an+bn,(λa)n=λan.(a+b)_n=a_n+b_n,\qquad (\lambda a)_n=\lambda a_n.

The collection of all scalar sequences is a vector space, but analysis typically restricts attention to subspaces satisfying boundedness, convergence, or summability conditions. Different choices capture different notions of size and approximation.

For 1≤p<∞1\leq p<\infty, the space ℓp\ell^p consists of sequences satisfying

∑n=1∞∣an∣p<∞,∥a∥p=(∑n=1∞∣an∣p)1/p.\sum_{n=1}^{\infty}|a_n|^p<\infty, \qquad \|a\|_p=\left(\sum_{n=1}^{\infty}|a_n|^p\right)^{1/p}.

The space ℓ∞\ell^\infty consists of bounded sequences with norm ∥a∥∞=sup⁡n∣an∣\|a\|_\infty=\sup_n|a_n|. These are discrete instances of LpL^p spaces, using counting measure on the natural numbers, and all are Banach spaces. The space ℓ2\ell^2 is additionally a Hilbert space, with inner product

⟨a,b⟩=∑n=1∞anbn‾.\langle a,b\rangle=\sum_{n=1}^{\infty}a_n\overline{b_n}.

The Cauchy–Schwarz inequality ensures that this sum is well defined.

The space c0c_0 consists of sequences converging to zero, while cc consists of all convergent scalar sequences. Both are closed subspaces of ℓ∞\ell^\infty under the supremum norm. For finite pp, membership in ℓp\ell^p implies membership in c0c_0, but the reverse implication fails. The sequence an=1/na_n=1/n, for example, tends to zero and belongs to ℓp\ell^p for p>1p>1, but not to ℓ1\ell^1. Summability therefore imposes a stronger condition than termwise decay.

Convergence of a sequence of points in a sequence space must be distinguished from convergence along the coordinates of one point. Let e(k)e^{(k)} have value 11 in coordinate kk and 00 elsewhere. For each fixed coordinate nn, en(k)→0e^{(k)}_n\to0 as k→∞k\to\infty. Nevertheless, ∥e(k)∥p=1\|e^{(k)}\|_p=1, so these vectors do not converge to zero in norm. In ℓ2\ell^2, they do converge weakly to zero: their inner products with every fixed vector tend to zero.

Sequence spaces provide concrete models for functional analysis. Coordinates in an orthonormal basis turn vectors into square-summable coefficient sequences, and linear operators can often be studied through their action on these coefficients. This connects abstract infinite-dimensional geometry with Fourier coefficients, signal representations, and infinite systems of equations.

Random sequences and probabilistic convergence

A sequence of random variables (Xn)(X_n) assigns a random value to each index. Defined on a common probability space, it is a discrete-time stochastic process. For a fixed outcome ω\omega, the values (Xn(ω))(X_n(\omega)) form an ordinary sequence, called a sample path. Probabilistic convergence can concern these paths, the probability of large errors, average error size, or only the distributions of the terms. These are distinct questions.

Almost-sure convergence means that Xn(ω)→X(ω)X_n(\omega)\to X(\omega) outside a set of probability zero. Convergence in probability means

Pr⁡(∣Xn−X∣>ε)⟶0\Pr(|X_n-X|>\varepsilon)\longrightarrow0

for every ε>0\varepsilon>0. Almost-sure convergence implies convergence in probability, but the converse does not generally hold. Nevertheless, every sequence converging in probability has a subsequence that converges almost surely to the same limit. This subsequence principle is useful when pathwise arguments are easier than direct probabilistic estimates.

Convergence in LpL^p, for p≥1p\geq1, means

E∣Xn−X∣p⟶0.\mathbb E|X_n-X|^p\longrightarrow0.

It implies convergence in probability by Markov’s inequality. It need not follow from almost-sure convergence: on (0,1)(0,1) with uniform probability, Xn(ω)=n1(0,1/n)(ω)X_n(\omega)=n\mathbf1_{(0,1/n)}(\omega) tends to zero at every fixed ω>0\omega>0, but E∣Xn∣=1\mathbb E|X_n|=1. Increasingly rare events can carry enough magnitude to prevent mean convergence. Integrability conditions are therefore necessary when passing from pathwise limits to limits of expectations.

Convergence in distribution is weaker still. For real random variables it means that their distribution functions converge at every continuity point of the limiting distribution function. The variables need not even be defined on the same probability space. Convergence in probability implies convergence in distribution, but distributional convergence alone does not specify pathwise proximity to a particular random variable.

The law of large numbers and central limit theorem illustrate different uses of random sequences. For independent identically distributed variables with finite absolute mean μ\mu, the sample averages converge almost surely to μ\mu. Under finite positive variance σ2\sigma^2, the normalized deviations

n(X‾n−μ)σ\frac{\sqrt n(\overline X_n-\mu)}{\sigma}

converge in distribution to the standard normal law. One theorem describes stabilization of averages; the other describes the distribution of their fluctuations at a shrinking scale. Dependence, tail behavior, and normalization determine which analogues remain valid for more general random sequences.

Numerical iteration and computability

Many numerical methods produce a sequence intended to approach a solution. For solving f(x)=0f(x)=0, Newton’s method uses

xn+1=xn−f(xn)f′(xn),x_{n+1}=x_n-\frac{f(x_n)}{f'(x_n)},

when the denominator is nonzero. Near a simple root rr, suitable smoothness and a sufficiently close starting point yield quadratic convergence:

∣xn+1−r∣≤C∣xn−r∣2|x_{n+1}-r|\leq C|x_n-r|^2

for all sufficiently large nn. This is a local statement, not a guarantee for every initial value. Iterates can encounter zero derivatives, enter cycles, or move away from the desired root.

A broadly applicable convergence mechanism is the Banach fixed-point theorem. If a map TT on a nonempty complete metric space satisfies

d(Tx,Ty)≤q d(x,y),0≤q<1,d(Tx,Ty)\leq q\,d(x,y),\qquad 0\leq q<1,

then xn+1=T(xn)x_{n+1}=T(x_n) converges to a unique fixed point x∗x^\ast. It also gives the computable estimate

d(xn,x∗)≤qn1−q d(x1,x0).d(x_n,x^\ast)\leq \frac{q^n}{1-q}\,d(x_1,x_0).

The estimate converts qualitative convergence into an explicit bound on how many steps are sufficient for a desired accuracy.

Stopping criteria require comparable justification. A small difference ∣xn+1−xn∣|x_{n+1}-x_n| does not alone prove proximity to the answer. A small residual ∣f(xn)∣|f(x_n)| may also be misleading if the equation is poorly conditioned. Error bounds depend on the structure of the problem, not merely on the appearance of the latest terms. In numerical analysis, convergence of the ideal mathematical sequence is therefore distinguished from numerical stability of its implementation under floating-point arithmetic.

There is also a logical distinction between a well-defined sequence and a computable one. An integer sequence is computable when an algorithm can return its nnth term for every input nn. A sequence of rational approximations may be computable term by term yet lack a computable rate of convergence. A modulus of convergence is a function N(k)N(k) guaranteeing an error below, for example, 2−k2^{-k} after index N(k)N(k). Such a modulus provides the effective control needed to turn convergence into an approximation procedure.

Not every real number or infinite binary sequence is computable: algorithms have finite descriptions and form a countable collection, whereas infinite binary sequences form an uncountable set. Consequently, mathematical existence of a sequence does not imply the existence of a finite procedure for generating its terms. This boundary connects sequence theory with computation, constructive analysis, and the foundations of effective approximation.

Historical development

Sequence-like reasoning long predates the modern definition of a function on the natural numbers. Ancient work on figurate numbers studied arrangements corresponding to square, triangular, and other numerical patterns. In Greek mathematics, proportions and repeated constructions supplied tools for investigating indefinitely continuing processes without the modern symbolism of indexed terms. Euclid’s *Elements* includes results on numbers in continued proportion, while Archimedes used successive geometric approximations in determining areas and estimating the ratio of a circle’s circumference to its diameter.

Medieval and early modern arithmetic developed recurring numerical patterns in practical and theoretical settings. Leonardo of Pisa, known as Fibonacci, presented the rabbit problem associated with the Fibonacci numbers in the Liber abaci of 1202; a revised version appeared in 1228. Related recurrences had already arisen in Indian work on poetic meters. The significance of the example lies not in the biological realism of the rabbit model but in the transparent generation of each term from two preceding terms, a structure later connected with algebraic formulas and counting problems.

During the seventeenth and eighteenth centuries, sequences and infinite series became central to the development of calculus. Infinite expansions supported calculations with trigonometric, logarithmic, and exponential functions. Leonhard Euler used series and generating-function methods extensively in the eighteenth century, including in number theory and the study of partitions. Formal manipulation was extraordinarily productive, but questions about when an infinite operation preserved equality increasingly required systematic treatment.

Nineteenth-century analysis made convergence an explicit foundation rather than an informal expectation. In his Cours d’analyse of 1821, Augustin-Louis Cauchy developed a systematic account of limits, infinitesimals, continuity, and series, including the criterion now associated with Cauchy sequences. Karl Weierstrass and other analysts contributed to the precise inequality-based treatment of limiting processes. The construction of the real numbers and the study of completeness clarified why some Cauchy sequences have limits within a number system and others do not.

The late nineteenth and twentieth centuries broadened the objects carried by sequences. Georg Cantor used infinite listings and diagonal arguments in the study of countability and real numbers. The development of metric spaces, topology, and functional analysis separated convergence from any particular numerical formula and exposed the limits of purely sequential descriptions in general spaces. Probability added several inequivalent modes of convergence, while computing made iteration and effective error bounds central practical concerns. The resulting subject retains the elementary image of terms in order while supporting theories of infinite-dimensional spaces, random behavior, and algorithmic approximation.