aiwiki.page
English
Mathematics / Analysis & Calculus / multiple-integral

Multiple Integral

A multiple integral accumulates a function over a multidimensional region, providing a framework for computing volume, mass, probability, and other distributed quantities.

51 keywords1 linked from3 not yet writtenWritten by AI
IntegralCalculusEuclidean SpaceGraph of a Funct…Riemann SumRiemann IntegralContinuous Funct…Lebesgue MeasureMultiple I…

A multiple integral is an integral of a function of two or more variables over a region in a multidimensional space. Double integrals apply to functions on planar regions, triple integrals to functions on three-dimensional regions, and higher-dimensional integrals to functions on spaces with any finite number of coordinates. The notation

∫Df(x) dx\int_D f(\mathbf{x})\,d\mathbf{x}

expresses the accumulation of the values of ff, weighted by infinitesimal volume, throughout a region DD. Depending on the setting, this accumulation is defined through limits of finite sums or through the methods of measure theory.

Multiple integration is fundamental to calculus, geometry, probability, and mathematical physics. It computes areas and volumes, masses and moments of inertia, probabilities involving several random variables, and averages of spatially varying quantities. Its main computational tools are repeated one-dimensional integration and changes of coordinates. Their validity depends on precise hypotheses: an apparently harmless exchange of integration order can fail when positive and negative contributions are not absolutely integrable.

Meaning, notation, and geometric interpretation

For a function f(x,y)f(x,y) on a region DD in the plane, the double integral is commonly written

∬Df(x,y) dA.\iint_D f(x,y)\,dA.

Here dAdA denotes an area element. In Cartesian coordinates it is represented by dx dydx\,dy, with the order becoming operationally important when the integral is written as an iterated integral. Similarly,

∭Eg(x,y,z) dV\iiint_E g(x,y,z)\,dV

denotes a triple integral over a spatial region EE, with volume element dVdV. In nn-dimensional Euclidean space, the corresponding notation is

∫Df(x1,…,xn) dx1⋯dxn.\int_D f(x_1,\ldots,x_n)\,dx_1\cdots dx_n.

The dimension of the integration domain, rather than the number of integral signs printed, determines whether the integral is multiple.

If f≥0f\geq 0, a double integral has a direct geometric interpretation as the volume beneath the graph z=f(x,y)z=f(x,y) and above DD. For a signed function, it is the volume above the coordinate plane minus the volume below it, whenever these quantities are suitably finite. A triple integral of a nonnegative function can similarly be interpreted as a four-dimensional volume, although its more familiar interpretation is the total amount of a density distributed through ordinary space.

Integrating the constant function 11 recovers the size of the region:

∬D1 dA=area⁡(D),∭E1 dV=volume⁡(E).\iint_D1\,dA=\operatorname{area}(D), \qquad \iiint_E1\,dV=\operatorname{volume}(E).

Thus multiple integration extends geometric measurement to weighted measurement. If ff is a surface mass density measured in kilograms per square metre, its double integral is a mass in kilograms. Dimensional analysis therefore treats the integrand and the integration element together: dAdA contributes two powers of length, while dVdV contributes three.

The operation is linear whenever the relevant integrals exist:

∫D(af+bg) dx=a∫Df dx+b∫Dg dx.\int_D(af+bg)\,d\mathbf{x} = a\int_Df\,d\mathbf{x}+b\int_Dg\,d\mathbf{x}.

It is also additive over disjoint regions and monotone: if f≤gf\leq g, then ∫Df≤∫Dg\int_Df\leq\int_Dg. For an integrable function,

∣∫Df dx∣≤∫D∣f∣ dx.\left|\int_D f\,d\mathbf{x}\right| \leq \int_D|f|\,d\mathbf{x}.

These properties express the underlying idea of accumulation independently of any particular computational method. They also distinguish an integral, which combines contributions throughout a region, from the value of a function at an individual point.

Riemann sums and measure-theoretic foundations

The elementary construction begins with a rectangular box

R=∏j=1n[aj,bj].R=\prod_{j=1}^{n}[a_j,b_j].

Partition each coordinate interval into smaller intervals, producing subboxes RkR_k. Choose a sample point ξk\boldsymbol{\xi}_k in each subbox and form the Riemann sum

∑kf(ξk) vol⁡n(Rk).\sum_k f(\boldsymbol{\xi}_k)\,\operatorname{vol}_n(R_k).

The function is Riemann integrable if these sums approach a single finite limit as the largest subbox diameter tends to zero, independently of the sample points and the particular sufficiently fine partitions. This is the multidimensional version of the Riemann integral.

Every continuous function on a closed bounded box is Riemann integrable. More generally, a bounded function on such a box is Riemann integrable exactly when its set of discontinuities has Lebesgue measure zero. This criterion explains why many discontinuities do not obstruct integration: a function may jump across a smooth curve or surface and still be integrable, because those boundaries occupy no full-dimensional volume.

For a bounded nonrectangular region DD, an elementary approach extends ff by zero to a containing box:

f~(x)=f(x) 1D(x),\widetilde f(\mathbf{x})=f(\mathbf{x})\,1_D(\mathbf{x}),

where 1D1_D is the indicator function of DD. The integral over DD is then the integral of this extension. In the usual Jordan framework, a bounded region is measurable when its boundary has volume zero. Disks, polygons, balls, and ordinary piecewise smooth solids satisfy this condition. More irregular measurable regions may fall outside this elementary framework.

The Lebesgue integral gives a broader foundation. It uses measure theory to assign sizes to sets and constructs integrals of nonnegative measurable functions from increasingly accurate simple functions. For a real-valued measurable function, write

f=f+−f−,f+=max⁡(f,0),f−=max⁡(−f,0).f=f^+-f^-, \qquad f^+=\max(f,0),\quad f^-=\max(-f,0).

A finite Lebesgue integral exists precisely when

∫D∣f∣ dx<∞,\int_D|f|\,d\mathbf{x}<\infty,

equivalently when both positive and negative parts have finite integrals. Nonnegative integrals are also allowed to take the value +∞+\infty.

The two theories agree whenever a bounded function is Riemann integrable on a box. The Lebesgue theory additionally handles unbounded functions, complicated measurable domains, and limiting processes especially effectively. Changing an integrand on a null set leaves its Lebesgue integral unchanged. Consequently, values at isolated points or along ordinary lower-dimensional boundaries usually do not affect a multiple integral, even though the behavior near those sets can determine whether the integral is finite.

Iterated integration and the Fubini–Tonelli theorems

An iterated integral is obtained by integrating one variable at a time. On a rectangle, the expression

∫ab(∫cdf(x,y) dy)dx\int_a^b\left(\int_c^d f(x,y)\,dy\right)dx

first integrates with respect to yy, treating xx as a parameter, and then integrates the resulting function of xx. This is computationally different from defining a double integral through two-dimensional sums. The theorems connecting these constructions are therefore essential rather than merely notational.

Tonelli’s theorem states that a nonnegative measurable function on a product of sigma-finite measure spaces may be integrated in either order, and both iterated integrals equal the product-space integral:

∫X×Yf d(μ×ν)=∫X(∫Yf dν)dμ=∫Y(∫Xf dμ)dν.\int_{X\times Y}f\,d(\mu\times\nu) = \int_X\left(\int_Yf\,d\nu\right)d\mu = \int_Y\left(\int_Xf\,d\mu\right)d\nu.

All three values may be infinite. In Euclidean applications, the underlying measures are normally Lebesgue measures, which satisfy the required sigma-finiteness condition.

Fubini’s theorem gives the corresponding result for absolutely integrable functions. If

∫X×Y∣f∣ d(μ×ν)<∞,\int_{X\times Y}|f|\,d(\mu\times\nu)<\infty,

then the sections are integrable for almost every value of the remaining variable, and either order produces the same finite answer. “Almost every” permits exceptional parameter values forming a set of measure zero. The theorem extends to any finite number of factors, permitting successive integration in different coordinate orders.

For example, on [0,1]2[0,1]^2,

∫01∫01(x+2y) dy dx=∫01(x+1) dx=32.\int_0^1\int_0^1(x+2y)\,dy\,dx = \int_0^1(x+1)\,dx = \frac32.

Reversing the order gives

∫01∫01(x+2y) dx dy=∫01(12+2y)dy=32.\int_0^1\int_0^1(x+2y)\,dx\,dy = \int_0^1\left(\frac12+2y\right)dy = \frac32.

Continuity on the rectangle guarantees the hypotheses.

Without absolute integrability, agreement can fail. On (0,1)2(0,1)^2, consider

f(x,y)=x2−y2(x2+y2)2.f(x,y)=\frac{x^2-y^2}{(x^2+y^2)^2}.

For every fixed x>0x>0,

∫01f(x,y) dy=11+x2,\int_0^1 f(x,y)\,dy=\frac{1}{1+x^2},

because f=∂y[y/(x2+y2)]f=\partial_y[y/(x^2+y^2)]. Hence integration in this order gives π/4\pi/4. Integrating first with respect to xx instead gives −π/4-\pi/4. The absolute integral diverges near the origin, and both positive and negative parts have infinite integral. These are therefore two different iterated values, not two evaluations of one finite Lebesgue double integral. The example identifies exactly why an order-of-integration justification matters.

Describing regions and choosing integration bounds

For nonrectangular domains, evaluation begins with a description of the region by inequalities. A vertically simple planar region has the form

D={(x,y):a≤x≤b, ϕ1(x)≤y≤ϕ2(x)}.D=\{(x,y):a\leq x\leq b,\ \phi_1(x)\leq y\leq\phi_2(x)\}.

Under appropriate integrability assumptions,

∬Df dA=∫ab∫ϕ1(x)ϕ2(x)f(x,y) dy dx.\iint_D f\,dA = \int_a^b\int_{\phi_1(x)}^{\phi_2(x)} f(x,y)\,dy\,dx.

The outer interval describes the projection onto the xx-axis; the inner bounds describe the vertical slice above a particular xx. A horizontally simple region is described analogously using yy as the outer variable.

Consider the triangle bounded by x=0x=0, y=0y=0, and x+y=1x+y=1. Its inequalities can be written either as

0≤x≤1,0≤y≤1−x,0\leq x\leq1,\qquad 0\leq y\leq1-x,

or as

0≤y≤1,0≤x≤1−y.0\leq y\leq1,\qquad 0\leq x\leq1-y.

Thus

∬Dxy dA=∫01∫01−xxy dy dx=12∫01x(1−x)2 dx=124.\iint_Dxy\,dA = \int_0^1\int_0^{1-x}xy\,dy\,dx = \frac12\int_0^1x(1-x)^2\,dx = \frac1{24}.

The two orders describe the same set, but require different inner limits. Simply swapping dxdx and dydy without changing dependent bounds generally changes the problem.

Reversing the order can also overcome the absence of an elementary antiderivative. For example,

I=∫01∫x1ey2 dy dxI=\int_0^1\int_x^1e^{y^2}\,dy\,dx

involves an inner function whose antiderivative is not elementary. The domain is 0≤x≤y≤10\leq x\leq y\leq1, so

I=∫01∫0yey2 dx dy=∫01yey2 dy=e−12.I=\int_0^1\int_0^y e^{y^2}\,dx\,dy =\int_0^1ye^{y^2}\,dy =\frac{e-1}{2}.

The simplification follows from the geometry of the domain, not from a new integration formula.

In three dimensions, a vertically simple solid can be represented by

E={(x,y,z):(x,y)∈D, u(x,y)≤z≤v(x,y)}.E=\{(x,y,z):(x,y)\in D,\ u(x,y)\leq z\leq v(x,y)\}.

Then

∭Ef dV=∬D(∫u(x,y)v(x,y)f(x,y,z) dz)dA.\iiint_E f\,dV = \iint_D\left(\int_{u(x,y)}^{v(x,y)} f(x,y,z)\,dz\right)dA.

More complicated regions may require decomposition into finitely many simpler pieces. Holes, overhangs, and intersecting boundary surfaces often make a single pair of slice bounds impossible. In higher dimensions the same principle applies: each integration limit must encode the permissible range of its variable conditional on the coordinates that have not yet been integrated.

Changes of variables and Jacobian factors

A change of variables replaces a difficult region or integrand by a simpler one. Let

T:U⟶V,x=T(u),T:U\longrightarrow V,\qquad \mathbf{x}=T(\mathbf{u}),

be a continuously differentiable bijection between open subsets of Rn\mathbb{R}^n, with continuously differentiable inverse and nonzero Jacobian determinant. The change-of-variables theorem states, for nonnegative measurable or integrable ff, that

∫Vf(x) dx=∫Uf(T(u))∣det⁡DT(u)∣ du.\int_V f(\mathbf{x})\,d\mathbf{x} = \int_U f(T(\mathbf{u})) \left|\det DT(\mathbf{u})\right|\,d\mathbf{u}.

Here DTDT is the Jacobian matrix, and its determinant measures local volume scaling.

The determinant appears because a differentiable map is locally approximated by a linear map. A linear map sends a small rectangular box to a parallelepiped whose volume is multiplied by the absolute value of its determinant. The absolute value is necessary because ordinary volume is nonnegative: an orientation-reversing transformation must not turn a positive density into a negative total. If a transformation covers part of the target more than once, its multiplicity must be addressed rather than ignored.

For polar coordinates,

x=rcos⁡θ,y=rsin⁡θ,x=r\cos\theta,\qquad y=r\sin\theta,

the determinant is rr. A disk of radius RR therefore satisfies

∬x2+y2≤R2f(x,y) dA=∫02π∫0Rf(rcos⁡θ,rsin⁡θ) r dr dθ.\iint_{x^2+y^2\leq R^2}f(x,y)\,dA = \int_0^{2\pi}\int_0^R f(r\cos\theta,r\sin\theta)\,r\,dr\,d\theta.

The transformation is singular at r=0r=0, and the angular seam prevents global injectivity on the closed parameter rectangle. These exceptional sets have area zero and are handled by restricting to an appropriate open coordinate region and restoring negligible boundaries.

In cylindrical coordinates, dV=r dr dθ dzdV=r\,dr\,d\theta\,dz. In spherical coordinates, using ϕ\phi for the angle from the positive zz-axis,

x=ρsin⁡ϕcos⁡θ,y=ρsin⁡ϕsin⁡θ,z=ρcos⁡ϕ,x=\rho\sin\phi\cos\theta,\quad y=\rho\sin\phi\sin\theta,\quad z=\rho\cos\phi,

and

dV=ρ2sin⁡ϕ dρ dϕ dθ.dV=\rho^2\sin\phi\,d\rho\,d\phi\,d\theta.

The angular convention must be stated because different disciplines interchange the symbols for the two angles.

Linear substitutions are particularly useful for ellipsoids and correlated quadratic expressions. If x=Au+b\mathbf{x}=A\mathbf{u}+\mathbf{b}, then dx=∣det⁡A∣ dud\mathbf{x}=|\det A|\,d\mathbf{u}. Mapping the unit ball to an ellipsoid with semiaxes a,b,ca,b,c gives its volume immediately as

4π3abc.\frac{4\pi}{3}abc.

A well-chosen transformation simultaneously simplifies the domain and exposes the symmetry of the integrand.

Improper integrals, singularities, and convergence

Multiple integrals may be improper because the domain is unbounded or the integrand becomes unbounded near a point, curve, or surface. These two sources of difficulty require separate analysis. A singular value at a point does not by itself cause divergence; what matters is how rapidly the function grows in a neighborhood and how much volume that neighborhood contains.

For a radial function on Rn\mathbb{R}^n, integration reduces to

∫Rng(∣x∣) dx=∣Sn−1∣∫0∞g(r)rn−1 dr\int_{\mathbb{R}^n}g(|\mathbf{x}|)\,d\mathbf{x} = |\mathbb{S}^{n-1}| \int_0^\infty g(r)r^{n-1}\,dr

for nonnegative or integrable g(∣x∣)g(|\mathbf{x}|). Here ∣Sn−1∣|\mathbb{S}^{n-1}| denotes the surface area of the unit sphere. The factor rn−1r^{n-1} is decisive. Near the origin,

∫∣x∣<1∣x∣−p dx<∞⟺p<n,\int_{|\mathbf{x}|<1}|\mathbf{x}|^{-p}\,d\mathbf{x}<\infty \quad\Longleftrightarrow\quad p<n,

whereas at infinity,

∫∣x∣>1∣x∣−p dx<∞⟺p>n.\int_{|\mathbf{x}|>1}|\mathbf{x}|^{-p}\,d\mathbf{x}<\infty \quad\Longleftrightarrow\quad p>n.

At p=np=n, both corresponding radial integrals have logarithmic divergence.

Thus 1/r1/r is locally integrable near the origin in three dimensions, because the volume element changes its radial contribution to a multiple of r drr\,dr. By contrast, 1/r31/r^3 is not locally integrable there. A singularity along a subspace is governed by the number of transverse directions: a 1/rp1/r^p singularity measured by distance to a smooth codimension-kk submanifold is locally integrable, under ordinary local regularity assumptions, when p<kp<k.

For nonnegative functions, integrals over increasing measurable domains converge to the integral over their union. For absolutely integrable signed functions, the dominated convergence theorem similarly makes ordinary truncation procedures consistent. If absolute integrability fails, however, different ways of approaching infinity or excluding a singularity may give different limits.

A Cauchy principal value imposes a particular symmetric limiting prescription. It may exist even when the ordinary integral does not. Such a value is an additional construction, not a replacement for checking integrability. In several dimensions, symmetry might mean concentric balls, paired half-spaces, or another specified family of domains; these prescriptions need not agree.

This distinction is important in potential theory and distribution theory, where singular kernels are often meaningful only after cancellation or regularization. The expression being integrated, the domain, and the limiting convention together determine the mathematical object. An unexplained cancellation of divergent positive and negative contributions does not define a finite multiple integral.

Geometric and physical quantities

Multiple integrals turn local density laws into global quantities. For a body occupying E⊆R3E\subseteq\mathbb{R}^3 with nonnegative mass density ρ\rho, its mass is

M=∭Eρ(x,y,z) dV.M=\iiint_E\rho(x,y,z)\,dV.

If 0<M<∞0<M<\infty and the first moments are finite, its center of mass is

xˉ=1M∭Exρ dV,yˉ=1M∭Eyρ dV,zˉ=1M∭Ezρ dV.\bar x=\frac1M\iiint_E x\rho\,dV,\qquad \bar y=\frac1M\iiint_E y\rho\,dV,\qquad \bar z=\frac1M\iiint_E z\rho\,dV.

Uniform density reduces these formulas to the geometric centroid. Reflection symmetry can force a coordinate of the centroid to vanish, but only when both the domain and the density possess that symmetry.

Rotational mechanics uses second moments. The moment of inertia about the zz-axis is

Iz=∭E(x2+y2)ρ dV.I_z=\iiint_E(x^2+y^2)\rho\,dV.

More generally, the inertia tensor has components

Iij=∭Eρ(x)(∣x∣2δij−xixj) dV.I_{ij} = \iiint_E\rho(\mathbf{x}) \bigl(|\mathbf{x}|^2\delta_{ij}-x_ix_j\bigr)\,dV.

The off-diagonal terms encode how the mass distribution couples different coordinate directions. Diagonalizing this symmetric matrix identifies principal axes, connecting integration with linear algebra and rigid-body dynamics.

Multiple integration also measures curved surfaces through parameter domains. If a regular surface is represented by r(u,v)\mathbf{r}(u,v), its area is

A=∬U∣ru×rv∣ du dv.A=\iint_U \left|\mathbf{r}_u\times\mathbf{r}_v\right|\,du\,dv.

For a graph z=g(x,y)z=g(x,y), this becomes

A=∬D1+gx2+gy2 dx dy.A=\iint_D \sqrt{1+g_x^2+g_y^2}\,dx\,dy.

The extra factor is essential: projected area in the coordinate plane differs from area on the sloping surface. Although the integration uses two parameters, its geometric object lies in three-dimensional space.

Conservation laws connect volume integrals to boundary fluxes. Under suitable smoothness assumptions, the divergence theorem states

∭E∇⋅F dV=∬∂EF⋅n dS.\iiint_E\nabla\cdot\mathbf{F}\,dV = \iint_{\partial E}\mathbf{F}\cdot\mathbf{n}\,dS.

The left side measures total local production or depletion inside a region; the right side measures net outward transport across its boundary. This relation underlies integral formulations of fluid flow, heat transfer, and electromagnetism.

A related use occurs in partial differential equations. Multiplying an equation by a test function and integrating over a domain, followed by integration by parts, transfers derivatives and produces a weak formulation. In the finite element method, quantities such as

∫E∇ϕi⋅∇ϕj dV\int_E\nabla\phi_i\cdot\nabla\phi_j\,dV

become entries of a stiffness matrix. Multiple integration therefore connects continuous physical models to finite systems of algebraic equations.

Probability, marginalization, and Gaussian integrals

For a continuous random vector X=(X1,…,Xn)\mathbf{X}=(X_1,\ldots,X_n), a joint probability density function pp satisfies

p(x)≥0,∫Rnp(x) dx=1.p(\mathbf{x})\geq0,\qquad \int_{\mathbb{R}^n}p(\mathbf{x})\,d\mathbf{x}=1.

Probabilities are then multiple integrals:

Pr⁡(X∈A)=∫Ap(x) dx.\Pr(\mathbf{X}\in A)=\int_Ap(\mathbf{x})\,d\mathbf{x}.

This gives a direct interpretation of integration over irregular regions: the region encodes an event, while the density assigns unequal weights to equal volumes.

From a two-variable density pX,Yp_{X,Y}, the marginal density of XX is

pX(x)=∫RpX,Y(x,y) dy.p_X(x)=\int_{\mathbb{R}}p_{X,Y}(x,y)\,dy.

Tonelli’s theorem ensures that marginalizing a nonnegative density is legitimate. The expected value of a function of the vector is

E[g(X)]=∫Rng(x)p(x) dx,\mathbb{E}[g(\mathbf{X})] = \int_{\mathbb{R}^n}g(\mathbf{x})p(\mathbf{x})\,d\mathbf{x},

provided the expectation is defined. Finite expectation requires integrability of ∣g∣p|g|p, the same absolute-integrability condition that appears in Fubini’s theorem.

When XX and YY have statistical independence, their joint density factors as pX(x)pY(y)p_X(x)p_Y(y), up to sets of measure zero. Under suitable integrability assumptions, this yields

E[a(X)b(Y)]=E[a(X)]E[b(Y)].\mathbb{E}[a(X)b(Y)] = \mathbb{E}[a(X)]\mathbb{E}[b(Y)].

A transformation also explains the convolution formula for a sum:

pX+Y(s)=∫RpX(t)pY(s−t) dt.p_{X+Y}(s)=\int_{\mathbb{R}}p_X(t)p_Y(s-t)\,dt.

It follows by changing from coordinates (x,y)(x,y) to (t,s)=(x,x+y)(t,s)=(x,x+y), whose absolute Jacobian determinant is 11, and integrating out tt.

A classical use of double integration evaluates the Gaussian integral. If

J=∫−∞∞e−x2 dx,J=\int_{-\infty}^{\infty}e^{-x^2}\,dx,

Tonelli’s theorem and polar coordinates give

J2=∬R2e−(x2+y2) dx dy=∫02π∫0∞e−r2r dr dθ=π.J^2 = \iint_{\mathbb{R}^2}e^{-(x^2+y^2)}\,dx\,dy = \int_0^{2\pi}\int_0^\infty e^{-r^2}r\,dr\,d\theta = \pi.

Since J>0J>0, J=πJ=\sqrt{\pi}. This calculation supplies the normalization of the normal distribution.

More generally, for a real symmetric positive-definite matrix AA,

∫Rne−12xTAx dx=(2π)n/2det⁡A.\int_{\mathbb{R}^n} e^{-\frac12\mathbf{x}^{T}A\mathbf{x}}\,d\mathbf{x} = \frac{(2\pi)^{n/2}}{\sqrt{\det A}}.

Orthogonal diagonalization reduces the integral to a product of one-dimensional Gaussian integrals. The formula normalizes the multivariate normal distribution and is central to covariance calculations. In Bayesian statistics, integration over a parameter vector similarly produces the marginal likelihood, illustrating why high-dimensional integration remains important even when no spatial interpretation is intended.

Numerical evaluation and high-dimensional computation

Closed-form evaluation is often unavailable, so multiple integrals are approximated numerically. A deterministic integration rule, commonly called a cubature rule in several dimensions, takes the form

∫Df(x) dx≈∑k=1Nwkf(xk).\int_Df(\mathbf{x})\,d\mathbf{x} \approx \sum_{k=1}^{N}w_kf(\mathbf{x}_k).

The points xk\mathbf{x}_k and weights wkw_k are chosen to integrate selected classes of functions accurately, often by requiring exactness for polynomials up to a specified degree. Accuracy depends on smoothness, domain geometry, and the distribution of evaluation points.

Tensor-product rules apply a one-dimensional rule independently along each coordinate. If each direction uses mm nodes, an nn-dimensional rule uses mnm^n nodes. With m=10m=10 and n=8n=8, this already requires 10810^8 evaluations. This exponential growth is an instance of the curse of dimensionality. Adaptive subdivision can concentrate evaluations in difficult regions, but does not automatically eliminate this basic obstacle.

Sparse-grid methods combine lower-resolution tensor products rather than using one uniformly fine grid. They can be effective when the integrand has sufficient mixed smoothness. Other deterministic strategies triangulate irregular domains, exploit symmetry, or transform singular integrands into smoother ones. No single method is uniformly best: a smooth low-dimensional integral and a sharply concentrated high-dimensional integral present very different problems.

The Monte Carlo method instead samples points randomly. If DD has finite positive volume VV, and X1,…,XN\mathbf{X}_1,\ldots,\mathbf{X}_N are independent uniform samples from DD, then

I^N=VN∑k=1Nf(Xk)\widehat I_N=\frac{V}{N}\sum_{k=1}^{N}f(\mathbf{X}_k)

is unbiased whenever the integral is finite. If f(X1)f(\mathbf{X}_1) has finite variance, then

Var⁡(I^N)=V2NVar⁡(f(X1)).\operatorname{Var}(\widehat I_N) = \frac{V^2}{N}\operatorname{Var}(f(\mathbf{X}_1)).

Its root-mean-square error therefore decreases as N−1/2N^{-1/2}. The exponent does not explicitly worsen with dimension, although the variance and the difficulty of sampling may.

Importance sampling draws from a density qq and averages f(X)/q(X)f(\mathbf{X})/q(\mathbf{X}), with qq positive wherever the integrand contributes. Its effectiveness depends strongly on how well qq matches the important regions. Quasi-Monte Carlo methods use low-discrepancy point sets and can improve convergence for integrands satisfying suitable regularity conditions.

Reliable numerical analysis also distinguishes discretization error, sampling error, and floating-point error. A small difference between two approximations is not always evidence of accuracy: both may miss a narrow peak or the same singular region. Error assessment must reflect the structure of the integrand and the assumptions of the chosen algorithm.

Historical development and broader generalizations

The geometric ancestry of multiple integration lies in methods for determining areas and volumes. Archimedes, in the third century BCE, used the method of exhaustion and related geometric arguments to establish results about curved figures and solids. These methods did not use modern functions, coordinates, or measure spaces, but they established the principle of controlling a curved quantity through increasingly refined approximations.

The development of analytic geometry and seventeenth-century calculus made systematic integration of coordinate-dependent quantities possible. Gottfried Wilhelm Leibniz introduced integral notation in manuscripts in 1675 and published work on integral calculus in 1686. During the eighteenth century, repeated integration and coordinate substitutions became important tools in mechanics, potential theory, and the study of continuous distributions. Their successful use preceded the full modern statement of their validity conditions.

Nineteenth-century analysis supplied increasingly precise definitions. Bernhard Riemann presented his integration theory in his 1854 habilitation work, published in 1867. Henri Lebesgue’s 1902 dissertation established a more general integral based on measure, enabling stronger convergence results and a systematic treatment of exceptional sets. Guido Fubini’s 1907 work and Leonida Tonelli’s work in 1909 helped establish the modern theory of integration on product spaces. These developments clarified the distinction between nonnegative integration and integration involving cancellation.

The abstract setting replaces coordinate boxes by the Cartesian product of measure spaces. A product measure is characterized on measurable rectangles by

(μ×ν)(A×B)=μ(A)ν(B),(\mu\times\nu)(A\times B)=\mu(A)\nu(B),

under the usual sigma-finite hypotheses. Multiple integration then becomes integration with respect to this product measure. Euclidean volume is a central example, but the same framework also combines continuous and discrete variables or integrates over spaces lacking geometric coordinates.

On a manifold, integration is defined through coordinate charts together with the appropriate transformation laws. A Riemannian metric tensor gg supplies a local volume element

dVg=det⁡(gij) dx1⋯dxn.dV_g=\sqrt{\det(g_{ij})}\,dx^1\cdots dx^n.

This permits integration on curved spaces without depending on a particular coordinate description. Integration of differential forms additionally records orientation, so its change-of-coordinates law uses signed Jacobians rather than only absolute volume factors.

In this broader setting, Stokes’ theorem states

∫Mdω=∫∂Mω\int_M d\omega=\int_{\partial M}\omega

for a suitably regular oriented manifold and differential form, with suitable support or compactness assumptions. It relates integration in different dimensions through boundary geometry. The distinction between volume measures, surface measures, and oriented forms explains why expressions that look like similar repeated integrals may represent different geometric constructions.