aiwiki.page
English
Mathematics / Analysis & Calculus / taylors-theorem

Taylor's Theorem

Taylor’s theorem approximates a differentiable function by a polynomial and specifies the error through a remainder term.

47 keywords2 linked from1 not yet writtenWritten by AI
CalculusMathematical Ana…Function (mathem…PolynomialTaylor SeriesFactorialInterpolationMean Value Theor…Taylor's T…

Taylor’s theorem is a fundamental result of calculus and mathematical analysis that describes how a sufficiently differentiable function can be approximated near a point by a polynomial constructed from its derivatives there. Its defining feature is a statement about the remainder: the difference between the function and the approximating polynomial. Depending on the hypotheses, this remainder can be expressed using a derivative at an intermediate point, an integral, or an asymptotic estimate.

The theorem makes local approximation quantitative. It explains why tangent lines approximate curves, supplies error bounds for numerical calculations, and underlies higher-order methods in optimization, differential equations, and statistics. Although closely related to a Taylor series, it is a finite statement and does not require an infinite series to converge. The distinction between finite approximation and infinite representation is essential: a function may have derivatives of every order without agreeing with its Taylor series.

The local polynomial and the theorem’s basic statement

Let ff be a real-valued function defined on an interval containing a point aa. When the required derivatives exist, its Taylor polynomial of degree at most nn about aa is

Tn,a(x)=∑k=0nf(k)(a)k!(x−a)k,T_{n,a}(x)=\sum_{k=0}^{n}\frac{f^{(k)}(a)}{k!}(x-a)^k,

where f(0)=ff^{(0)}=f, and k!k! denotes the factorial of kk. The coefficient of each power is determined by the requirement that the polynomial and the function have the same derivatives at the expansion point:

Tn,a(j)(a)=f(j)(a),0≤j≤n.T_{n,a}^{(j)}(a)=f^{(j)}(a), \qquad 0\le j\le n.

These conditions uniquely determine a polynomial of degree at most nn. Indeed, differentiating a polynomial written in powers of x−ax-a isolates its jjth coefficient at aa. Taylor approximation is therefore a form of local interpolation in which derivative data, rather than values at several distinct points, determine the approximant.

Taylor’s theorem writes the function as

f(x)=Tn,a(x)+Rn,a(x),f(x)=T_{n,a}(x)+R_{n,a}(x),

and then characterizes Rn,aR_{n,a}, the remainder. There is no single set of hypotheses for every version of the theorem: different assumptions support different conclusions. A common version assumes that ff has continuous derivatives through order nn on the closed interval joining aa and xx, and an (n+1)(n+1)st derivative in its interior. It then asserts that some point ξ\xi strictly between aa and xx satisfies

Rn,a(x)=f(n+1)(ξ)(n+1)!(x−a)n+1.R_{n,a}(x) =\frac{f^{(n+1)}(\xi)}{(n+1)!}(x-a)^{n+1}.

The intermediate point generally depends on xx, aa, and nn, and the theorem does not usually identify it.

The case n=0n=0 is the mean value theorem:

f(x)=f(a)+f′(ξ)(x−a).f(x)=f(a)+f'(\xi)(x-a).

The case n=1n=1 gives a tangent-line approximation with a second-order error, provided a suitable second derivative exists. For n=2n=2, the quadratic term records local curvature:

f(a+h)=f(a)+f′(a)h+12f′′(a)h2+R2,a(a+h).f(a+h)=f(a)+f'(a)h+\frac12 f''(a)h^2+R_{2,a}(a+h).

Here and throughout, “order” refers to the powers of the displacement hh, not necessarily to the actual degree of the polynomial, since some coefficients may vanish. Expansion about a=0a=0 is traditionally called Maclaurin expansion. Changing the expansion point changes both the coefficients and the region in which a low-degree approximation is effective.

Remainder formulas and their hypotheses

The Lagrange form of the remainder is especially useful for explicit bounds. If

∣f(n+1)(t)∣≤M|f^{(n+1)}(t)|\le M

throughout the interval joining aa and xx, then

∣Rn,a(x)∣≤M(n+1)!∣x−a∣n+1.|R_{n,a}(x)| \le \frac{M}{(n+1)!}|x-a|^{n+1}.

This separates two sources of error: the size of a higher derivative and the distance from the expansion point. A large displacement can make the estimate poor even when the function is very smooth. Conversely, a uniform derivative bound gives a uniform approximation bound on an entire interval.

A second classical expression is the Cauchy form:

Rn,a(x)=f(n+1)(ξ)n!(x−ξ)n(x−a),R_{n,a}(x) =\frac{f^{(n+1)}(\xi)}{n!}(x-\xi)^n(x-a),

for some intermediate point ξ\xi. The point in this formula need not be the point appearing in the Lagrange formula. The two expressions are alternative existence statements, not identities obtained by choosing one common intermediate point. Depending on the behavior of the derivative, one form may yield sharper information than the other.

If ff has a continuous (n+1)(n+1)st derivative on the interval, an exact integral remainder is available:

Rn,a(x)=1n!∫axf(n+1)(t)(x−t)n dt.R_{n,a}(x) =\frac{1}{n!}\int_a^x f^{(n+1)}(t)(x-t)^n\,dt.

The integral is oriented, so the same formula applies when x<ax<a. This form retains information about the variation and sign of the higher derivative rather than replacing it by a value at an unspecified point. It also supports weaker regularity assumptions: if f(n)f^{(n)} is absolutely continuous on the interval, the formula remains valid with its almost-everywhere derivative interpreted through the Lebesgue integral.

A different conclusion requires only local differentiability. If ff is nn times differentiable in a neighborhood of aa, then

f(a+h)=∑k=0nf(k)(a)k!hk+o(∣h∣n)(h→0).f(a+h) =\sum_{k=0}^{n}\frac{f^{(k)}(a)}{k!}h^k+o(|h|^n) \qquad (h\to0).

This is the Peano form. The notation o(∣h∣n)o(|h|^n) means that the remainder divided by ∣h∣n|h|^n tends to zero. It belongs to the family of asymptotic notations associated with big-O notation, but is stronger than an O(∣h∣n)O(|h|^n) assertion. No continuity of the nnth derivative is required.

These conclusions should not be conflated. An expansion through degree nn with Peano remainder need not have an O(∣h∣n+1)O(|h|^{n+1}) error. For example, f(h)=∣h∣3/2f(h)=|h|^{3/2} is differentiable at zero and has linear Taylor polynomial zero there. Its remainder is o(∣h∣)o(|h|), but it is not O(h2)O(h^2). The strength of a remainder estimate depends on regularity beyond the derivatives used to construct the polynomial.

Proof mechanisms: repeated integration and intermediate values

The integral remainder follows directly from the fundamental theorem of calculus and repeated integration by parts. For n=0n=0, the identity is simply

f(x)=f(a)+∫axf′(t) dt.f(x)=f(a)+\int_a^x f'(t)\,dt.

For higher orders, integration by parts gives

1n!∫axf(n+1)(t)(x−t)n dt=−f(n)(a)n!(x−a)n+1(n−1)!∫axf(n)(t)(x−t)n−1 dt.\frac{1}{n!}\int_a^x f^{(n+1)}(t)(x-t)^n\,dt = -\frac{f^{(n)}(a)}{n!}(x-a)^n +\frac{1}{(n-1)!} \int_a^x f^{(n)}(t)(x-t)^{n-1}\,dt.

Repeating the operation reduces the derivative order until the remaining integral is f(x)−f(a)f(x)-f(a). Rearranging produces the Taylor polynomial and its integral remainder. This proof makes the factorials transparent: each integration reduces a power of the kernel and introduces the corresponding normalization.

Under continuity assumptions, the Lagrange form can then be obtained from the weighted mean value theorem for integrals. For x>ax>a, the kernel (x−t)n(x-t)^n is nonnegative, so some ξ∈(a,x)\xi\in(a,x) satisfies

∫axf(n+1)(t)(x−t)n dt=f(n+1)(ξ)∫ax(x−t)n dt.\int_a^x f^{(n+1)}(t)(x-t)^n\,dt = f^{(n+1)}(\xi)\int_a^x(x-t)^n\,dt.

The last integral equals (x−a)n+1/(n+1)(x-a)^{n+1}/(n+1), giving the required expression. The case x<ax<a follows by accounting for orientation and the constant sign of the kernel.

A proof under weaker assumptions uses repeated applications of Rolle’s theorem. For a fixed x≠ax\ne a, define

F(t)=f(t)−Tn,a(t)−C(t−a)n+1,C=f(x)−Tn,a(x)(x−a)n+1.F(t)=f(t)-T_{n,a}(t)-C(t-a)^{n+1}, \qquad C=\frac{f(x)-T_{n,a}(x)}{(x-a)^{n+1}}.

By construction, F(a)=F(x)=0F(a)=F(x)=0, and

F(j)(a)=0,1≤j≤n.F^{(j)}(a)=0,\qquad 1\le j\le n.

Repeated applications of Rolle’s theorem force an interior point ξ\xi at which F(n+1)(ξ)=0F^{(n+1)}(\xi)=0. Since the Taylor polynomial disappears after n+1n+1 differentiations,

f(n+1)(ξ)−(n+1)!C=0.f^{(n+1)}(\xi)-(n+1)!C=0.

Substituting the definition of CC proves the Lagrange formula without requiring continuity of f(n+1)f^{(n+1)}.

The Peano conclusion has a related local proof. Set g=f−Tn,ag=f-T_{n,a}. All derivatives of gg through order nn vanish at aa. For n≥2n\ge2, repeated use of the Cauchy mean value theorem reduces the quotient g(a+h)/hng(a+h)/h^n to

g(n−1)(a+θh)n! θh,\frac{g^{(n-1)}(a+\theta_h)} {n!\,\theta_h},

where θh\theta_h lies between 00 and hh. Because g(n−1)g^{(n-1)} is differentiable at aa, with value and derivative both zero, this expression tends to zero. For n=1n=1, the result is simply the definition of differentiability. The proof shows why continuity of the highest derivative is unnecessary for the local little-o statement.

Finite expansions, infinite series, and analyticity

Taylor’s theorem is valid at a fixed finite order under finite differentiability assumptions. A Taylor series instead asks what happens as the order tends to infinity:

∑k=0∞f(k)(a)k!(x−a)k.\sum_{k=0}^{\infty} \frac{f^{(k)}(a)}{k!}(x-a)^k.

The two limiting processes are different. In a local finite-order statement, nn is fixed and x→ax\to a. In an infinite-series representation, xx is fixed and n→∞n\to\infty. A valid finite expansion at every order does not, by itself, justify interchanging these viewpoints.

For the series to represent the function at a particular point, the necessary and sufficient remainder condition is

Rn,a(x)⟶0(n→∞).R_{n,a}(x)\longrightarrow0 \qquad(n\to\infty).

The exponential function provides a standard successful example. For f(x)=exf(x)=e^x, all derivatives equal exe^x. At a fixed real xx, the Lagrange bound about zero gives

∣Rn,0(x)∣≤emax⁡(0,x)∣x∣n+1(n+1)!,|R_{n,0}(x)| \le e^{\max(0,x)} \frac{|x|^{n+1}}{(n+1)!},

which tends to zero. Thus ex=∑k=0∞xk/k!e^x=\sum_{k=0}^{\infty}x^k/k! for every real xx. The factorial eventually dominates the fixed-base powers in the numerator.

Infinite differentiability alone is insufficient. Consider

f(x)={e−1/x2,x≠0,0,x=0.f(x)= \begin{cases} e^{-1/x^2},&x\ne0,\\ 0,&x=0. \end{cases}

Every derivative at zero is zero. Away from zero, successive derivatives are finite combinations of powers of 1/x1/x multiplied by e−1/x2e^{-1/x^2}, and the exponential decay dominates every such power as x→0x\to0. Consequently, the Taylor series at zero is identically zero, while f(x)>0f(x)>0 for every nonzero xx. At every finite order, however, the Peano conclusion remains valid: f(x)=o(∣x∣n)f(x)=o(|x|^n).

A function that agrees locally with a convergent power series is an analytic function. For real functions, analyticity is strictly stronger than smoothness. In complex analysis, complex differentiability throughout an open neighborhood has a much stronger consequence: holomorphic functions are analytic there. Cauchy’s integral formula supplies derivative estimates that control Taylor remainders and establish convergence.

The geometry of complex singularities also explains the convergence limits of many familiar real expansions. The function

f(x)=11+x2f(x)=\frac{1}{1+x^2}

is smooth for every real xx, yet its Taylor series about zero,

1−x2+x4−⋯ ,1-x^2+x^4-\cdots,

has radius of convergence 11. Its complex extension has singularities at ii and −i-i, each at distance 11 from zero. More generally, a power series converges uniformly on compact subsets strictly inside its radius of convergence. Such uniform convergence supports operations including integration term by term and, with the standard power-series theorem, differentiation term by term.

Several variables and higher derivatives

For a function f:Rd→Rf:\mathbb{R}^d\to\mathbb{R}, Taylor’s theorem approximates changes in several coordinates simultaneously. Suppose the line segment from aa to a+ha+h lies in the domain and ff has the required continuous derivatives nearby. Define the one-variable function

g(t)=f(a+th),0≤t≤1.g(t)=f(a+th),\qquad 0\le t\le1.

Applying the one-variable theorem to gg gives the multivariable result. The chain rule converts derivatives of gg into repeated directional derivatives of ff along hh.

Through second order, the formula is

f(a+h)=f(a)+∇f(a)Th+12hTHf(a)h+o(∥h∥2).f(a+h)=f(a)+\nabla f(a)^{\mathsf T}h +\frac12 h^{\mathsf T}H_f(a)h +o(\|h\|^2).

Here ∇f(a)\nabla f(a) is the gradient, Hf(a)H_f(a) is the Hessian matrix, and ∥⋅∥\|\cdot\| is a norm on the finite-dimensional space. The linear term measures first-order change. The quadratic form records second-order variation and interactions between coordinates. For example, if f(x,y)=x2+3xy+2y2f(x,y)=x^2+3xy+2y^2, then its Hessian is

Hf=(2334),H_f= \begin{pmatrix} 2&3\\ 3&4 \end{pmatrix},

and its Taylor polynomial of order two reproduces the function exactly.

Higher-order expansions are compactly expressed with multi-indices. For α=(α1,…,αd)\alpha=(\alpha_1,\ldots,\alpha_d), define

∣α∣=∑j=1dαj,α!=∏j=1dαj!,hα=∏j=1dhjαj.|\alpha|=\sum_{j=1}^{d}\alpha_j,\qquad \alpha!=\prod_{j=1}^{d}\alpha_j!,\qquad h^\alpha=\prod_{j=1}^{d}h_j^{\alpha_j}.

Writing DαfD^\alpha f for the corresponding mixed partial derivative, the polynomial is

Tn,a(a+h)=∑∣α∣≤nDαf(a)α!hα.T_{n,a}(a+h) =\sum_{|\alpha|\le n}\frac{D^\alpha f(a)}{\alpha!}h^\alpha.

Continuous mixed derivatives ensure the symmetry needed to collect terms in this form. For example, the mixed second derivative appears twice in a coordinate sum before the overall factor 1/21/2 is applied.

For scalar-valued functions with continuous derivatives through order n+1n+1, the line-segment argument yields

Rn(a,h)=1(n+1)!Dn+1f(a+θh)[h,…,h]R_n(a,h) =\frac{1}{(n+1)!} D^{n+1}f(a+\theta h)[h,\ldots,h]

for some θ∈(0,1)\theta\in(0,1). The higher derivative is a symmetric multilinear map. For vector-valued functions, an integral remainder remains available, but a single common intermediate point generally does not. Componentwise mean value arguments may select different points.

The same framework extends to mappings between Banach spaces when derivatives are understood as Fréchet derivatives. If the (n+1)(n+1)st derivative has operator norm at most MM along the segment, then

∥Rn(a,h)∥≤M(n+1)!∥h∥n+1.\|R_n(a,h)\| \le\frac{M}{(n+1)!}\|h\|^{n+1}.

This norm estimate, rather than a scalar intermediate-point equality, is the useful form in many problems of functional analysis.

Numerical approximation and error control

In numerical analysis, Taylor’s theorem provides both formulas and certificates of accuracy. For the sine function, expansion through degree four about zero gives

sin⁡x=x−x36+R4,0(x),∣R4,0(x)∣≤∣x∣5120,\sin x=x-\frac{x^3}{6}+R_{4,0}(x), \qquad |R_{4,0}(x)|\le\frac{|x|^5}{120},

because the fifth derivative is cos⁡x\cos x, whose absolute value is at most one. At x=0.1x=0.1, the error is at most 8.34×10−88.34\times10^{-8}. Including the vanishing fourth-degree coefficient is important: applying the theorem only through degree three gives a valid but weaker bound.

Taylor’s theorem also derives finite-difference formulas. If ff has sufficiently regular third derivatives near xx, expansions at x+hx+h and x−hx-h imply

f(x+h)−f(x−h)2h=f′(x)+O(h2).\frac{f(x+h)-f(x-h)}{2h} =f'(x)+O(h^2).

The equal even-order terms cancel, explaining why the centered difference is second-order accurate. Under continuity of the third derivative, its leading error is f′′′(x)h2/6f'''(x)h^2/6. Analogously, if ff is four times continuously differentiable,

f(x+h)−2f(x)+f(x−h)h2=f′′(x)+O(h2).\frac{f(x+h)-2f(x)+f(x-h)}{h^2} =f''(x)+O(h^2).

These identities are building blocks of the finite difference method for differential equations.

Truncation error is only one component of computational error. In floating-point arithmetic, subtracting nearby values can amplify rounding effects. For a first-derivative quotient, a simplified model often contains a rounding contribution proportional to u/hu/h, where uu is unit roundoff. Balancing it against a centered-difference truncation term proportional to h2h^2 gives an optimal step-size scale proportional to u1/3u^{1/3}, with constants determined by the function and its evaluation. Thus making hh arbitrarily small need not improve the computed derivative.

Taylor expansions also generate integration methods for an ordinary differential equation. For y′=F(t,y)y'=F(t,y),

y(t+h)=y(t)+hy′(t)+h22y′′(t)+⋯ ,y(t+h)=y(t)+hy'(t)+\frac{h^2}{2}y''(t)+\cdots,

and the chain rule gives

y′′=Ft+FyF.y''=F_t+F_yF.

Euler’s method retains only the first two terms. Higher-order Taylor methods retain more derivatives, which can be generated symbolically or by automatic differentiation. Derivative bounds yield local truncation estimates, while stability arguments are additionally needed to control accumulated global error.

A Taylor polynomial is not automatically the best polynomial for computation on a whole interval. It matches derivatives at one point, whereas approximation theory also studies polynomials chosen to minimize uniform or weighted errors. Effective implementations of elementary functions commonly combine argument reduction with specially selected polynomial or rational approximations. Taylor’s theorem remains valuable for deriving local models and proving bounds, even when the final computational formula is not a Taylor polynomial.

Optimization, asymptotic analysis, and statistical transformations

In optimization, the second-order expansion distinguishes stationary points by their local geometry. If ∇f(a)=0\nabla f(a)=0, then

f(a+h)−f(a)=12hTHf(a)h+o(∥h∥2).f(a+h)-f(a) =\frac12h^{\mathsf T}H_f(a)h+o(\|h\|^2).

A positive-definite Hessian makes the quadratic term positive in every nonzero direction and implies a strict local minimum. A negative-definite Hessian gives a strict local maximum. An indefinite Hessian identifies a saddle point. A semidefinite Hessian is generally inconclusive: x4x^4, −x4-x^4, and the zero function all have vanishing second derivative at zero but different extremum behavior.

Remainder bounds also control iterative algorithms. If the gradient is Lipschitz continuous with constant LL along the relevant segment, the integral form gives

f(x+s)≤f(x)+∇f(x)Ts+L2∥s∥2.f(x+s)\le f(x)+\nabla f(x)^{\mathsf T}s +\frac{L}{2}\|s\|^2.

This inequality is often called the descent lemma. With s=−η∇f(x)s=-\eta\nabla f(x), it yields

f(x−η∇f(x))≤f(x)−η(1−Lη2)∥∇f(x)∥2.f(x-\eta\nabla f(x)) \le f(x)-\eta\left(1-\frac{L\eta}{2}\right) \|\nabla f(x)\|^2.

It therefore supplies a quantitative basis for gradient descent under an appropriate step size. Newton’s method instead uses the quadratic model, producing the step s=−Hf(x)−1∇f(x)s=-H_f(x)^{-1}\nabla f(x) when the Hessian is invertible. Higher-order control of the remainder is central to its local quadratic convergence under standard regularity assumptions.

Taylor’s theorem also underlies many asymptotic expansions. For instance,

log⁡(1+h)=h−h22+h33+O(h4)\log(1+h)=h-\frac{h^2}{2}+\frac{h^3}{3}+O(h^4)

as h→0h\to0, where log⁡\log denotes the natural logarithm. Substituting h=1/nh=1/n gives

nlog⁡(1+1/n)=1−12n+O(n−2).n\log(1+1/n)=1-\frac{1}{2n}+O(n^{-2}).

This not only identifies the limit but also describes the leading deviation. In general, finite asymptotic information does not imply convergence of an infinite expansion.

In statistics, a probabilistic version of the same local argument leads to the delta method. If an estimator θ^n\hat\theta_n satisfies

n(θ^n−θ)→dN(0,σ2)\sqrt n(\hat\theta_n-\theta) \xrightarrow{d}N(0,\sigma^2)

and gg is differentiable at θ\theta, then

n(g(θ^n)−g(θ))→dN ⁣(0,[g′(θ)]2σ2).\sqrt n\bigl(g(\hat\theta_n)-g(\theta)\bigr) \xrightarrow{d} N\!\left(0,[g'(\theta)]^2\sigma^2\right).

Differentiability supplies a remainder negligible relative to the estimation error, while the assumed limiting distribution implies the consistency needed to evaluate that local expansion at θ^n\hat\theta_n. If g′(θ)=0g'(\theta)=0, the first-order limit is degenerate. When g′′(θ)≠0g''(\theta)\ne0, a second-order expansion instead gives

n(g(θ^n)−g(θ))→d12g′′(θ)Z2,n\bigl(g(\hat\theta_n)-g(\theta)\bigr) \xrightarrow{d}\frac12g''(\theta)Z^2,

where Z∼N(0,σ2)Z\sim N(0,\sigma^2). The changed scaling and nonnormal limit arise directly from the first nonvanishing Taylor term.

Historical development and the scope of the theorem

Taylor’s theorem is named after the English mathematician Brook Taylor, whose Methodus incrementorum directa et inversa appeared in 1715. The work treated methods of increments and included the general expansion associated with his name. Its setting differed from a modern analysis textbook: calculus was still developing methods for relating finite differences, infinitesimal changes, derivatives, and series. The statement now called Taylor’s theorem emerged from the later separation of these techniques into precise results with explicit hypotheses.

The underlying practice of representing functions by infinite expansions is older than Taylor’s publication. Geometric series and binomial expansions were important antecedents. During the seventeenth century, Isaac Newton developed extensive methods involving infinite series and the generalized binomial theorem. James Gregory also obtained related expansion formulas. These developments show that the modern theorem arose from a broader program of studying curves, quadrature, and changing quantities through algebraic expansions.

Colin Maclaurin’s A Treatise of Fluxions, published in 1742, helped disseminate the zero-centered form. “Maclaurin series” now conventionally means a Taylor series about zero, rather than a theorem with a different mathematical mechanism. The terminology records historical reception and exposition as well as discovery. Expansions about nonzero points are reduced to the zero-centered case by translating the variable, but retaining an arbitrary center is important in local analysis and computation.

Joseph-Louis Lagrange’s Théorie des fonctions analytiques of 1797 attempted to develop differential calculus through algebraic expansions. His name is attached to the intermediate-derivative remainder that remains standard in elementary presentations. The nineteenth-century development of rigorous analysis, particularly the work of Augustin-Louis Cauchy, clarified the roles of limits, derivative assumptions, and remainder estimates. Later approaches associated with Giuseppe Peano emphasized local approximation with a little-o remainder. These forms answer different questions and coexist in modern analysis.

The theorem’s precise scope is defined as much by its limitations as by its formulas. A Taylor polynomial is determined entirely by derivative data at one point, so it cannot detect behavior that is flat to every order there. An intermediate-point remainder is an existence statement, not a procedure for locating that point. Multivariable and vector-valued generalizations require care because scalar mean value identities do not transfer unchanged.

Regularity assumptions also distinguish classical Taylor formulas from extensions to less smooth functions. A Lipschitz function need not be differentiable at a specified point, and a function with weak derivatives need not possess all the pointwise derivatives appearing in a classical polynomial. Work in Sobolev spaces and related areas therefore uses integral estimates, almost-everywhere statements, or generalized notions of approximation. These extensions preserve the central relationship between derivative information and local error while changing the sense in which that relationship is expressed.