differentiation (changes) in nLab

Showing changes from revision #30 to #31: Added | Removed | Changed

See also differentiable function.

Contents

Idea

Differentiation is the process that assigns to a function f:X→Yf : X \to Y its derivative, sometimes denoted dfd f. The derivative is a function that, roughly speaking, assigns to each point x∈Xx \in X the linear transformation df xd f_x that maps infinitesimal differences y−xy - x (for points yy infinitesimally close to xx) to infinitesimal differences f(y)−f(x)f(y) - f(x).

Definitions

Classical (analytic) definitions

Classically, if f:ℝ n→ℝf: \mathbb{R}^n \to \mathbb{R} is a function, the derivative df xd f_x at a point x∈ℝ nx \in \mathbb{R}^n is defined by a limiting process: it is the unique (when it exists) linear transformation L:ℝ n→ℝL: \mathbb{R}^n \to \mathbb{R} such that

lim h→0f(x+h)−f(x)−Lh|h|=0\lim_{h \to 0} \frac{f(x + h) - f(x) - L h}{{|h|}} = 0

or in other words such that

f(y)=f(x)+L(y−x)+g(y)f(y) = f(x) + L(y - x) + g(y)

where g(y)g(y) is o(|y−x|)o({|y - x|}) (see o-notation?). We say in that case that ff is differentiable at xx, with derivative df x=Ld f_x = L. When n=1n = 1, such a linear transformation L:ℝ→ℝL: \mathbb{R} \to \mathbb{R} is multiplication by a scalar λ∈ℝ\lambda \in \mathbb{R}. This is often denoted f′(x)f'(x), so that the derivative may be regarded as another function f′:ℝ→ℝf': \mathbb{R} \to \mathbb{R}.

We say f:ℝ n→ℝf: \mathbb{R}^n \to \mathbb{R} is differentiable if it is differentiable at every point xx of its domain. In that case, we get a global derivative function ℝ n×ℝ n→ℝ\mathbb{R}^n \times \mathbb{R}^n \to \mathbb{R} which takes a pair (x,h)(x, h) to df x(h)d f_x(h). We can then ask whether this function is itself differentiable, i.e., whether the differentiation process can be iterated. The nn-th derivative is the result of differentiating nn times (the 00-th derivative of ff is just ff). We say that f:ℝ n→ℝf: \mathbb{R}^n \to \mathbb{R} is infinitely differentiable (or more briefly, that ff is smooth) if it has an nn-th derivative for each natural number nn.

Notice that a differentiable function is necessarily continuous (indeed, it is locally a Lipschitz function). One denotes the class of functions having a continuous¹ n thn^{th} derivative by C n(ℝ n)C^n(\mathbb{R}^n). Thus continuous functions f:ℝ n→ℝf: \mathbb{R}^n \to \mathbb{R} form a filtration

…⊂C n(ℝ k)⊂…⊂C 1(ℝ k)⊂C 0(ℝ k)\ldots \subset C^n(\mathbb{R}^k) \subset \ldots \subset C^1(\mathbb{R}^k) \subset C^0(\mathbb{R}^k)

and the intersection C ∞(ℝ k)≔⋂ n≥0C n(ℝ k)C^\infty(\mathbb{R}^k) \coloneqq \bigcap_{n \geq 0} C^n(\mathbb{R}^k) coincides with the class of smooth functions.

Evidently differentiation gives a function D:C n+1(ℝ k)→C n(ℝ k×ℝ k)D: C^{n+1}(\mathbb{R}^k) \to C^n(\mathbb{R}^k \times \mathbb{R}^k). The convenient setting for developing the differential calculus is to go directly to the smooth case, where we have D:C ∞(ℝ k)→C ∞(ℝ k×ℝ k)D: C^\infty(\mathbb{R}^k) \to C^\infty(\mathbb{R}^k \times \mathbb{R}^k) without having to keep track of the degree of differentiability.

This development carries over more generally to smooth manifolds XX, which are locally (smoothly) isomorphic to Euclidean spaces ℝ k\mathbb{R}^k.

Differentiation as a functor

Differentiation in more conceptual terms is a functor d:Diff→Diffd : Diff \to Diff on the category Diff of smooth manifolds that sends

• a manifold XX to its tangent bundle TXT X; recall that the points of TXT X are ordered pairs (x,v)(x, v) where x∈Xx \in X and vv is a tangent vector at xx, i.e., an (augmented) derivation v:C ∞(X)→ℝv: C^\infty(X) \to \mathbb{R} on the algebra of smooth functions, augmented by evaluation ev x:C ∞(X)→ℝev_x: C^\infty(X) \to \mathbb{R} at xx;

• a smooth function f::X→Y f : \colon X \to Y to its differential df:TX→TYd f : T X \to T Y : (cf.pushforward of vector fields):

  if γ:[−1,1]→X\gamma : [-1,1] \to X is a path in XX that represents a vector v∈T xXv \in T_x X, then (df)(v)∈T f(x)Y(d f)(v) \in T_{f(x)} Y is the vector represented by the path [−1,1]→γX→fY[-1,1] \stackrel{\gamma}{\to} X \stackrel{f}{\to} Y.

Equivalently, a smooth map f:X→Yf: X \to Y induces an algebra map f *=−∘f:C ∞(Y)→C ∞(X)f^\ast = - \circ f: C^\infty(Y) \to C^\infty(X), and hence by composition sends a derivation v:C ∞(X)→ℝv: C^\infty(X) \to \mathbb{R} to a derivation v∘f *:C ∞(Y)→ℝv \circ f^\ast: C^\infty(Y) \to \mathbb{R} augmented by ev f(x):C ∞(Y)→ℝev_{f(x)}: C^\infty(Y) \to \mathbb{R}). Concretely: v∘f *(ϕ)=v(ϕ∘f)v \circ f^\ast(\phi) = v(\phi \circ f) for ϕ∈C ∞(Y)\phi \in C^\infty(Y). Then the differential df:TX→TYd f: T X \to T Y is defined by df(x,v)=(f(x),v∘f *)d f(x, v) = (f(x), v \circ f^\ast).

According to either of these descriptions, differentiation is manifestly functorial. In the English literature, this fact is known as the chain rule. In more traditional form:

d(f∘g) x=df g(x)∘dg x.d(f \circ g)_x = d f_{g(x)} \circ d g_x.

In a context HH of synthetic differential geometry (SDG) the functor dd is simply the internal hom out of the infinitesimal interval DD

d=[D,−]:H→H. d = [D, -] : H \to H \,.

where D↪ℝD \hookrightarrow \mathbb{R} is the smooth locus defined by x 2=0x^2 = 0. (Classically, i.e., in the topos SetSet, this locus consists of just a single point², but in toposes that model SDG it should be regarded as an infinitesimal neighborhood about a point.) This description as internally representable functor underlies the following classical fact:

For example, the product of two smooth functions f,g:X→ℝf, g: X \to \mathbb{R} is a composition

X→ΔX×X→f×gℝ×ℝ→multℝX \stackrel{\Delta}{\to} X \times X \stackrel{f \times g}{\to} \mathbb{R} \times \mathbb{R} \stackrel{mult}{\to} \mathbb{R}

and the product rule for differentiation derives by applying the functor dd which results in the composite

TX→ΔTX×TX→df×dgTℝ×Tℝ→dmultTℝT X \stackrel{\Delta}{\to} T X \times T X \stackrel{d f \times d g}{\to} T \mathbb{R} \times T \mathbb{R} \stackrel{d mult}{\to} T \mathbb{R}

which reduces to computing dmultd mult.

Exposition of differentiation via infinitesimals

We attempt to give an exposition of differentiation not via taking limits in the sense of analysis, but rather, by working with infinitesimal spaces as in synthetic differential geometry.

Setup

For that purpose we regard objects such as the real line ℝ\mathbb{R} and more generally any smooth manifold as being objects of a well adapted smooth topos of smooth spaces. But in fact much less is required for most of the discussion. For the basic theory, it is sufficient to work in the canonical site of such a topos, namely, the category SmoothLocSmoothLoc of smooth loci. This site is defined via a categorial equivalence with the opposite category of the category of smooth algebras:

C ∞:SmoothLoc⟶∼SmoothAlg op C^\infty: SmoothLoc \stackrel{\sim}{\longrightarrow} SmoothAlg^{op}

Apart from the real line ℝ\mathbb{R}, the other object involved in our discussion of the basics of differentiation is the first order infinitesimal line segment DD (an infinitesimally thickened point), a subobject of ℝ\mathbb{R}. This subobject collects those points of ℝ\mathbb{R} that square to zero under the canonical multiplicative structure of ℝ\mathbb{R}

D={ϵ∈ℝ|ϵ 2=0}. D = \{\epsilon \in \mathbb{R} | \epsilon^2 = 0\} \,.

Although this is just {0}\{0\} when interpreted in the usual context of Set, here we are to interpret this in the ambient smooth topos, or equivalently just in our site SmoothLocSmoothLoc for it. That is to say, DD satisfies a pullback square in SmoothLocSmoothLoc:

D → * ↓ ↓ 0 ℝ →(−) 2 ℝ \array{ D &\to& * \\ \downarrow && \downarrow^{\mathrlap{0}} \\ \mathbb{R} &\stackrel{(-)^2}{\to}& \mathbb{R} }

Dually, the categorial equivalence C ∞C^\infty from SmoothLocSmoothLoc to the opposite category of smooth algebras carries the above pullback square to a pushout square of smooth algebras:

C ∞(D) ← C ∞(*)≃ℝ ↑ ↑ f↦f(0) C ∞(ℝ) ←x↦x 2 C ∞(ℝ) \array{ C^\infty(D) &\leftarrow& C^\infty(*) \simeq \mathbb{R} \\ \uparrow && \uparrow^{\mathrlap{f \mapsto f(0)}} \\ C^\infty(\mathbb{R}) &\stackrel{x \mapsto x^2}{\leftarrow}& C^\infty(\mathbb{R}) }

Here along the bottom and on the right are the ordinary smooth algebras of smooth functions on the smooth manifolds ℝ 1=ℝ\mathbb{R}^1 = \mathbb{R} and ℝ 0=*\mathbb{R}^0 = *, whereas C ∞(D)C^\infty(D) is new notation, suggestive and justified, for whatever the algebra of functions on the object DD is, where DD is defined by the fact that it has this algebra of functions. The bottom morphism sends a smooth function f:ℝ→ℝf : \mathbb{R} \to \mathbb{R} to the function f((−) 2):ℝ→ℝf((-)^2) : \mathbb{R} \to \mathbb{R}. If we write x=Id:ℝ→ℝx = Id : \mathbb{R} \to \mathbb{R} for the canonical coordinate function, then this sends xx to the function x 2x^2.

Notice that the commutativity of this pushout square means that the image in C ∞(D)C^\infty(D) of the squared coordinate function x 2∈C ∞(ℝ)x^2 \in C^\infty(\mathbb{R}) is zero. This is because the former is the image under the bottom morphism of the coordinate function xx and that maps under the right vertical morphism to 0∈C ∞(*)≃ℝ0 \in C^\infty(*) \simeq \mathbb{R}.

In fact one can show that the pushout C ∞(D)C^\infty(D) is generated by the coordinate function x∈C ∞(ℝ))x \in C^\infty(\mathbb{R})) subject to the relation x 2=0x^2 = 0: it is the ring of dual numbers

C ∞(D) ≃C ∞(ℝ)/(x 2) ≃ℝ[x]/(x 2). \begin{aligned} C^\infty(D) & \simeq C^\infty(\mathbb{R})/(x^2) \\ & \simeq \mathbb{R}[x]/(x^2) \end{aligned} \,.

(Here in the second step we have used Hadamard's lemma.)

This computation of the smooth algebra C ∞(D)C^\infty(D) of smooth functions on DD expresses the intuitive idea that DD is such a small neighbourhood of 00 in ℝ\mathbb{R} that while the canonical coordinate function of ℝ\mathbb{R} restricted to DD is non-vanishing (it would vanish only when restricted entirely to the point 0) it is “so small that its square is always 0”. Readers who like this intuitive statement should keep it in mind, as it is accurate and useful, whereas readers who do not like it could stick with the above precise definition, which unwinds this slogan in the internal logic of the ambient smooth topos that we keep alluding to.

In order to remind us about the infinitesimal nature of the generator xx in C ∞(D)C^\infty(D) we shall usually call this “dxdx” and write

C ∞(D)=ℝ[dx]/((dx) 2)≃ℝ⊕dxℝ. C^\infty(D) = \mathbb{R}[dx]/((dx)^2) \simeq \mathbb{R} \oplus dx \mathbb{R} \,.

Functions and differential 1-forms

We discussion of differential 1-forms are functions on the space of infinitesimal paths. For more comprehensive such discussion see (Stel 13).

Define the smooth locus ℝ (Δ inf 1)\mathbb{R}^{(\Delta^1_{inf})} as the product ℝ×D\mathbb{R} \times D of the real line with the first order infinitesimal line segment. We think of this as the space of infinitesimal paths in ℝ\mathbb{R} (see Spaces of infinitesimal k-simplicies). We do so by thinking of a generalized element of ℝ (Δ inf 1)\mathbb{R}^{(\Delta^1_{inf})} in the category SmoothLocSmoothLoc, namely a pair (x,ϵ)∈ℝ×D(x, \epsilon) \in \mathbb{R} \times D, as the linear path in ℝ\mathbb{R} stretching from xx to x+ϵx + \epsilon (both regarded as generalized elements of ℝ\mathbb{R}).

We say a smooth differential 1-form on ℝ\mathbb{R} is a function (meaning: a morphism in SmoothLocSmoothLoc)

ω:X (Δ inf 1)→ℝ \omega : X^{(\Delta^1_{inf})} \to \mathbb{R}

with the property that its restriction to ℝ\mathbb{R} vanishes, hence so that the composite morphism

0:ℝ↪X (Δ inf 1)→ωℝ 0 : \mathbb{R} \hookrightarrow X^{(\Delta^1_{inf})} \stackrel{\omega}{\to} \mathbb{R}

vanishes, where the first morphism is the inclusion of ℝ\mathbb{R} as the constant infinitesimal paths in ℝ\mathbb{R}, hence (Id,0):ℝ→ℝ×D(Id, 0) : \mathbb{R} \to \mathbb{R} \times D.

We call the coproduct of smooth algebras their tensor product (classically it is the smoothly completed algebraic tensor product). For each smooth loci XX and YY, we have

C ∞(X×Y)≃C ∞(X)⊗C ∞(Y). C^\infty(X \times Y) \simeq C^\infty(X) \otimes C^\infty(Y) \,.

Using this property, we can dualize the above definition of a diferential 1-form. Namely, in the opposite category of smooth algebras, a 1-form on the line is a morphism

C ∞(ℝ)⊗ℝ[dx]/((dx) 2)←C ∞(ℝ):ω *, C^\infty(\mathbb{R}) \otimes \mathbb{R}[dx]/((dx)^2) \leftarrow C^\infty(\mathbb{R}) : \omega^* \,,

hence an assignment x↦g(x)+h(x)dx x \mapsto g(x) + h(x) dx such that g(x)=0g(x) = 0. In other words, it is any element in C ∞(R)⊗ℝ[dx]/((dx) 2)C^\infty(\mathrm{R}) \otimes \mathbb{R}[dx]/((dx)^2) that is proportional to the generator dxdx. We write ω=hdx \omega = h dx with h∈C ∞(ℝ)h \in C^\infty(\mathbb{R}).

Derivatives as infinitesimal differences

We can now “synthetically” describe differentiation as the process of taking infinitesimal differences of functions.

Consider the two morphisms of smooth loci

+:ℝ (Δ inf 1)=ℝ×D→ℝ + : \mathbb{R}^{(\Delta^1_{inf})} = \mathbb{R} \times D \to \mathbb{R}

and

p 1:ℝ (Δ inf 1)=ℝ×D→ℝ, p_1 : \mathbb{R}^{(\Delta^1_{inf})} = \mathbb{R} \times D \to \mathbb{R} \,,

where the first is the restriction of the canonical addition operation +:ℝ×ℝ→ℝ+ : \mathbb{R} \times \mathbb{R} \to \mathbb{R} of real numbers, and where the second is the projection on the first factor out of the product.

Given an element f∈C ∞(ℝ)f \in C^\infty(\mathbb{R}), hence a smooth function f:ℝ→ℝf : \mathbb{R} \to \mathbb{R}, we can then form two functions on infinitesimal paths by precomposing with these two morphisms:

first we get

f:X (Δ inf 1)→p 1ℝ→fℝ. f : X^{(\Delta^1_{inf})} \stackrel{p_1}{\to} \mathbb{R} \stackrel{f}{\to} \mathbb{R} \,.

This is the function on infinitesimal paths in the line that only depends on the starting point of a path, not on the length of the path, and which sends that starting point to the value of ff on that point. Therefore by slight abuse of notation we just keep writing “ff” for this function.

The second combination is

f˜:X (Δ inf 1)→+ℝ→fℝ. \tilde f : X^{(\Delta^1_{inf})} \stackrel{+}{\to} \mathbb{R} \stackrel{f}{\to} \mathbb{R} \,.

This is the function on infinitesimal paths that sends any path (x)→(x+ϵ)(x) \to (x + \epsilon) to the value of the function on the endpoint (x+ϵ)(x + \epsilon) of the paths: in terms of these generalized elements f˜\tilde f is the assignment

f˜:(x,ϵ)↦f(x+ϵ). \tilde f : (x , \epsilon) \mapsto f(x + \epsilon) \,.

For instance consider the function f:x↦x 2f : x \mapsto x^2. Then

f˜:(x,ϵ) ↦(x+ϵ) 2 =x 2+2xϵ+ϵ 2 =x 2+2xϵ. \begin{aligned} \tilde f : (x, \epsilon) &\mapsto (x + \epsilon)^2 \\ & = x^2 + 2 x \epsilon + \epsilon^2 \\ & = x^2 + 2 x \epsilon \end{aligned} \,.

This is evident, if maybe still somewhat mysterious, in the internal language of the ambient smooth topos that we keep alluding to, but to make it entirely explicit and concrete notice that the functions here are dually morphisms of smooth algebras

C ∞(ℝ)⊗(ℝ⊕dxℝ)←C ∞(ℝ):(f˜) * C^\infty(\mathbb{R}) \otimes (\mathbb{R} \oplus dx \mathbb{R}) \leftarrow C^\infty(\mathbb{R}) : (\tilde f)^*

and the equation

(x+dx) 2=x 2+2xdx (x + dx)^2 = x^2 + 2 x dx

holds in C ∞(ℝ)⊗(ℝ⊕dxℝ)C^\infty(\mathbb{R}) \otimes (\mathbb{R} \oplus dx \mathbb{R}) by the above discussion, where now x∈C ∞(ℝ)x \in C^\infty(\mathbb{R}) and dx∈C ∞(D)=ℝ⊕dxℝdx \in C^\infty(D) = \mathbb{R} \oplus dx \mathbb{R} are the two canonical generators.

But notice that the expression x 2+2xdxx^2 + 2 x dx is not a dfifferential 1-form by the above definition, equivalently the function f˜\tilde f does not vanish when restricted to constant infinitesimal paths.

To get such, we form the derivative or infinitesimal difference, the function

f˜−f:X (Δ inf 1)→ℝ \tilde f - f : X^{(\Delta^1_{inf})} \to \mathbb{R}

that acts on generalized elements of the space of infinitesimal paths in the line as

(f˜−f):(x,ϵ)↦f(x+ϵ)−f(x). (\tilde f - f) : (x, \epsilon) \mapsto f(x + \epsilon) - f(x) \,.

By construction on the right this now is something linear in the infinitesimal length ϵ\epsilon of the path. Write f′∈C ∞(ℝ)f' \in C^\infty(\mathbb{R}) for the coefficient:

f(x+ϵ)−f(x)=ϵf′(x). f(x + \epsilon) - f(x) = \epsilon f'(x) \,.

This function f′f' is the differential of ff. It is the function that gives the change of the function ff along a “unit infinitesimal path”.

Dually, we have a morphism

C ∞(ℝ)⊗(ℝ⊕dxℝ)←C ∞(ℝ):(f˜−f) * C^\infty(\mathbb{R})\otimes (\mathbb{R} \oplus dx \mathbb{R}) \leftarrow C^\infty(\mathbb{R}) : (\tilde f - f)^*

that sends

x↦f(x+dx)−f(x)=f′(x)dx. x \mapsto f(x + dx) - f(x) = f'(x) dx \,.

The 1-form on the right is the de Rham differential of ff, usually written

(df)(x)=f′(x)dx∈C ∞(ℝ×D). (d f)(x) = f'(x) dx \in C^\infty(\mathbb{R} \times D) \,.

As a map of tangent bundles

Equivalently we may rephrase this differential dfd f as a morphism of tangent bundles df:Tℝ⟶Tℝd f \colon T \mathbb{R} \longrightarrow T \mathbb{R} as follows.

First, by the Kock-Lawvere axiom of SDG we have the standard fact that Tℝ:=ℝ D≃ℝ×ℝT \mathbb{R} := \mathbb{R}^D \simeq \mathbb{R} \times \mathbb{R}. Therefore given a function

ℝ×D⟶ℝ \mathbb{R} \times D \longrightarrow \mathbb{R}

this is by the internal hom-adjunction equivalently a function of the form

ℝ⟶ℝ D=Tℝ. \mathbb{R} \longrightarrow \mathbb{R}^D = T \mathbb{R} \,.

The condition on the original function makes this adjunct be a section of the tangent bundle of ℝ\mathbb{R}. This section is x↦f′(x)x \mapsto f'(x), hence is the derivative of ff regarded as a tangent vector on ℝ\mathbb{R}. Moreover, since ℝ\mathbb{R} is a microlinear space, this induces by rescaling a function

Tℝ=ℝ×ℝ⟶Tℝ. T \mathbb{R} = \mathbb{R} \times \mathbb{R} \longrightarrow T \mathbb{R} \,.

This is the differential of ff regarded as a map of tangent bundles.

• product law, chain rule

• Jacobian matrix, Hessian matrix, extremum

• partial differentiation

• Hadamard lemma, Fermat theory

• derivative of distributions

• pullback of differential forms

• tangent space, tangent bundle

• jet space, jet bundle, metric jet

Examples of sequences of local structures

geometry	point	first order infinitesimal	⊂\subset	formal = arbitrary order infinitesimal	⊂\subset	local = stalkwise	⊂\subset	finite
			←\leftarrow differentiation		integration →\to
smooth functions		derivative		Taylor series		germ		smooth function
curve (path)		tangent vector		jet		germ of curve		curve
smooth space		infinitesimal neighbourhood		formal neighbourhood		germ of a space		open neighbourhood
function algebra		square-0 ring extension		nilpotent ring extension/formal completion				ring extension
arithmetic geometry	𝔽 p\mathbb{F}_p finite field			ℤ p\mathbb{Z}_p p-adic integers		ℤ (p)\mathbb{Z}_{(p)} localization at (p)		ℤ\mathbb{Z} integers
Lie theory		Lie algebra		formal group		local Lie group		Lie group
symplectic geometry		Poisson manifold		formal deformation quantization		local strict deformation quantization		strict deformation quantization

References

General

The Kock-Lawvere axiom for the axiomatization of differentiation in synthetic differential geometry was introduced in

• Anders Kock, A simple axiomatics for differentiation, Mathematica Scandinavica Vol. 40, No. 2 (October 24, 1977), pp. 183-193 (JSTOR)

Another synthetic formalization of the chain rule in differential cohesive homotopy type theory is given in

• Felix Wellen, section 3 of Cartan Geometry in Modal Homotopy Type Theory (arXiv:1806.05966)

Discussion of differential forms as functions on infinitesimal simplices is in

• Herman Stel, Cosimplicial C ∞C^\infty rings and the de Rham complex of Euclidean space (arXiv:1310.7407)

History

Discussion of the history of the concept of differentiation, with emphasis on its roots all the way back in Zeno's paradoxes of motion is in

• Carl Benjamin Boyer, The history of the Calculus and its conceptual development, Dover 1949