Differential (mathematics)

• In calculus, the differential represents a change in the linearization of a function.
  • The total differential is its generalization for functions of multiple variables.
• In traditional approaches to calculus, the differentials (e.g. dx, dy, dt, etc.) are interpreted as infinitesimals. There are several methods of defining infinitesimals rigorously, but it is sufficient to say that an infinitesimal number is smaller in absolute value than any positive real number, just as an infinitely large number is larger than any real number.
• The differential is another name for the Jacobian matrix of partial derivatives of a function from Rⁿ to R^m (especially when this matrix is viewed as a linear map).
• More generally, the differential or pushforward refers to the derivative of a map between smooth manifolds and the pushforward operations it defines. The differential is also used to define the dual concept of pullback.
• Stochastic calculus provides a notion of stochastic differential and an associated calculus for stochastic processes.
• The integrator in a Stieltjes integral is represented as the differential of a function. Formally, the differential appearing under the integral behaves exactly as a differential: thus, the integration by substitution and integration by parts formulae for Stieltjes integral correspond, respectively, to the chain rule and product rule for the differential.

There is a simple way to make precise sense of differentials, first used on the Real line by regarding them as linear maps. It can be used on ${\displaystyle \mathbb {R} }$, ${\displaystyle \mathbb {R} ^{n}}$, a Hilbert space, a Banach space, or more generally, a topological vector space. The case of the Real line is the easiest to explain. This type of differential is also known as a covariant vector or cotangent vector, depending on context.

Suppose ${\displaystyle f(x)}$ is a real-valued function on ${\displaystyle \mathbb {R} }$. We can reinterpret the variable ${\displaystyle x}$ in ${\displaystyle f(x)}$ as being a function rather than a number, namely the identity map on the real line, which takes a real number ${\displaystyle p}$ to itself: ${\displaystyle x(p)=p}$. Then ${\displaystyle f(x)}$ is the composite of ${\displaystyle f}$ with ${\displaystyle x}$, whose value at ${\displaystyle p}$ is ${\displaystyle f(x(p))=f(p)}$. The differential ${\displaystyle \operatorname {d} f}$ (which of course depends on ${\displaystyle f}$) is then a function whose value at ${\displaystyle p}$ (usually denoted ${\displaystyle df_{p}}$) is not a number, but a linear map from ${\displaystyle \mathbb {R} }$ to ${\displaystyle \mathbb {R} }$. Since a linear map from ${\displaystyle \mathbb {R} }$ to ${\displaystyle \mathbb {R} }$ is given by a ${\displaystyle 1\times 1}$ matrix, it is essentially the same thing as a number, but the change in the point of view allows us to think of ${\displaystyle df_{p}}$ as an infinitesimal and compare it with the standard infinitesimal ${\displaystyle dx_{p}}$, which is again just the identity map from ${\displaystyle \mathbb {R} }$ to ${\displaystyle \mathbb {R} }$ (a ${\displaystyle 1\times 1}$ matrix with entry ${\displaystyle 1}$). The identity map has the property that if ${\displaystyle \varepsilon }$ is very small, then ${\displaystyle dx_{p}(\varepsilon )}$ is very small, which enables us to regard it as infinitesimal. The differential ${\displaystyle df_{p}}$ has the same property, because it is just a multiple of ${\displaystyle dx_{p}}$, and this multiple is the derivative ${\displaystyle f'(p)}$ by definition. We therefore obtain that ${\displaystyle df_{p}=f'(p)\,dx_{p}}$, and hence ${\displaystyle df=f'\,dx}$. Thus we recover the idea that ${\displaystyle f'}$ is the ratio of the differentials ${\displaystyle df}$ and ${\displaystyle dx}$.

This would just be a trick were it not for the fact that:

1. it captures the idea of the derivative of ${\displaystyle f}$ at ${\displaystyle p}$ as the best linear approximation to ${\displaystyle f}$ at ${\displaystyle p}$;
2. it has many generalizations.

If ${\displaystyle f}$ is a function from ${\displaystyle \mathbb {R} ^{n}}$ to ${\displaystyle \mathbb {R} }$, then we say that ${\displaystyle f}$ is differentiable[8] at ${\displaystyle p\in \mathbb {R} ^{n}}$ if there is a linear map ${\displaystyle df_{p}}$ from ${\displaystyle \mathbb {R} ^{n}}$ to ${\displaystyle \mathbb {R} }$ such that for any ${\displaystyle \varepsilon >0}$, there is a neighbourhood ${\displaystyle N}$ of ${\displaystyle p}$ such that for ${\displaystyle x\in N}$,

$${\displaystyle \left|f(x)-f(p)-df_{p}(x-p)\right|<\varepsilon \left|x-p\right|.}$$

We can now use the same trick as in the one-dimensional case and think of the expression ${\displaystyle f(x_{1},x_{2},\ldots ,x_{n})}$ as the composite of ${\displaystyle f}$ with the standard coordinates ${\displaystyle x_{1},x_{2},\ldots ,x_{n}}$ on ${\displaystyle \mathbb {R} ^{n}}$ (so that ${\displaystyle x_{j}(p)}$ is the ${\displaystyle j}$-th component of ${\displaystyle p\in \mathbb {R} ^{n}}$). Then the differentials ${\displaystyle \left(dx_{1}\right)_{p},\left(dx_{2}\right)_{p},\ldots ,\left(dx_{n}\right)_{p}}$ at a point ${\displaystyle p}$ form a basis for the vector space of linear maps from ${\displaystyle \mathbb {R} ^{n}}$ to ${\displaystyle \mathbb {R} }$ and therefore, if ${\displaystyle f}$ is differentiable at ${\displaystyle p}$, we can write ${\displaystyle \operatorname {d} f_{p}}$ as a linear combination of these basis elements:

$${\displaystyle df_{p}=\sum _{j=1}^{n}D_{j}f(p)\,(dx_{j})_{p}.}$$

The coefficients ${\displaystyle D_{j}f(p)}$ are (by definition) the partial derivatives of ${\displaystyle f}$ at ${\displaystyle p}$ with respect to ${\displaystyle x_{1},x_{2},\ldots ,x_{n}}$. Hence, if ${\displaystyle f}$ is differentiable on all of ${\displaystyle \mathbb {R} ^{n}}$, we can write, more concisely:

$${\displaystyle \operatorname {d} f={\frac {\partial f}{\partial x_{1}}}\,dx_{1}+{\frac {\partial f}{\partial x_{2}}}\,dx_{2}+\cdots +{\frac {\partial f}{\partial x_{n}}}\,dx_{n}.}$$

In the one-dimensional case this becomes

$${\displaystyle df={\frac {df}{dx}}dx}$$

as before.

This idea generalizes straightforwardly to functions from ${\displaystyle \mathbb {R} ^{n}}$ to ${\displaystyle \mathbb {R} ^{m}}$. Furthermore, it has the decisive advantage over other definitions of the derivative that it is invariant under changes of coordinates. This means that the same idea can be used to define the differential of smooth maps between smooth manifolds.

Aside: Note that the existence of all the partial derivatives of ${\displaystyle f(x)}$ at ${\displaystyle x}$ is a necessary condition for the existence of a differential at ${\displaystyle x}$. However it is not a sufficient condition. For counterexamples, see Gateaux derivative.

The same procedure works on a vector space with a enough additional structure to reasonably talk about continuity. The most concrete case is a Hilbert space, also known as a complete inner product space, where the inner product and its associated norm define a suitable concept of distance. The same procedure works for a Banach space, also known as a complete Normed vector space. However, for a more general topological vector space, some of the details are more abstract because there is no concept of distance.

For the important case of a finite dimension, any inner product space is a Hilbert space, any normed vector space is a Banach space and any topological vector space is complete. As a result, you can define a coordinate system from an arbitrary basis and use the same technique as for ${\displaystyle \mathbb {R} ^{n}}$.

This approach works on any differentiable manifold. If

1. U and V are open sets containing p
2. ${\displaystyle f\colon U\to \mathbb {R} }$ is continuous
3. ${\displaystyle g\colon V\to \mathbb {R} }$ is continuous

then f is equivalent to g at p, denoted ${\displaystyle f\sim _{p}g}$, if and only if there is an open ${\displaystyle W\subseteq U\cap V}$ containing p such that ${\displaystyle f(x)=g(x)}$ for every x in W. The germ of f at p, denoted ${\displaystyle [f]_{p}}$, is the set of all real continuous functions equivalent to f at p; if f is smooth at p then ${\displaystyle [f]_{p}}$ is a smooth germ. If

1. ${\displaystyle U_{1}}$, ${\displaystyle U_{2}}$ ${\displaystyle V_{1}}$ and ${\displaystyle V_{2}}$ are open sets containing p
2. ${\displaystyle f_{1}\colon U_{1}\to \mathbb {R} }$, ${\displaystyle f_{2}\colon U_{2}\to \mathbb {R} }$, ${\displaystyle g_{1}\colon V_{1}\to \mathbb {R} }$ and ${\displaystyle g_{2}\colon V_{2}\to \mathbb {R} }$ are smooth functions
3. ${\displaystyle f_{1}\sim _{p}g_{1}}$
4. ${\displaystyle f_{2}\sim _{p}g_{2}}$
5. r is a real number

then

1. ${\displaystyle r*f_{1}\sim _{p}r*g_{1}}$
2. ${\displaystyle f_{1}+f_{2}\colon U_{1}\cap U_{2}\to \mathbb {R} \sim _{p}g_{1}+g_{2}\colon V_{1}\cap V_{2}\to \mathbb {R} }$
3. ${\displaystyle f_{1}*f_{2}\colon U_{1}\cap U_{2}\to \mathbb {R} \sim _{p}g_{1}*g_{2}\colon V_{1}\cap V_{2}\to \mathbb {R} }$

This shows that the germs at p form an algebra.

Define ${\displaystyle {\mathcal {I}}_{p}}$ to be the set of all smooth germs vanishing at p and ${\displaystyle {\mathcal {I}}_{p}^{2}}$ to be the product of ideals ${\displaystyle {\mathcal {I}}_{p}{\mathcal {I}}_{p}}$. Then a differential at p (cotangent vector at p) is an element of ${\displaystyle {\mathcal {I}}_{p}/{\mathcal {I}}_{p}^{2}}$. The differential of a smooth function f at p, denoted ${\displaystyle \mathrm {d} f_{p}}$, is ${\displaystyle [f-f(p)]_{p}/{\mathcal {I}}_{p}^{2}}$.

A similar approach is to define differential equivalence of first order in terms of derivatives in an arbitrary coordinate patch. Then the differential of f at p is the set of all functions differentially equivalent to ${\displaystyle f-f(p)}$ at p.

In algebraic geometry, differentials and other infinitesimal notions are handled in a very explicit way by accepting that the coordinate ring or structure sheaf of a space may contain nilpotent elements. The simplest example is the ring of dual numbers R[ε], where ε² = 0.

This can be motivated by the algebro-geometric point of view on the derivative of a function f from R to R at a point p. For this, note first that f − f(p) belongs to the ideal I_p of functions on R which vanish at p. If the derivative f vanishes at p, then f − f(p) belongs to the square I_p² of this ideal. Hence the derivative of f at p may be captured by the equivalence class [f − f(p)] in the quotient space I_p/I_p², and the 1-jet of f (which encodes its value and its first derivative) is the equivalence class of f in the space of all functions modulo I_p². Algebraic geometers regard this equivalence class as the restriction of f to a thickened version of the point p whose coordinate ring is not R (which is the quotient space of functions on R modulo I_p) but R[ε] which is the quotient space of functions on R modulo I_p². Such a thickened point is a simple example of a scheme.[5]

Differentials are also important in algebraic geometry, and there are several important notions.

• Abelian differentials usually mean differential one-forms on an algebraic curve or Riemann surface.
• Quadratic differentials (which behave like "squares" of abelian differentials) are also important in the theory of Riemann surfaces.
• Kähler differentials provide a general notion of differential in algebraic geometry.

A fifth approach to infinitesimals is the method of synthetic differential geometry[9] or smooth infinitesimal analysis.[10] This is closely related to the algebraic-geometric approach, except that the infinitesimals are more implicit and intuitive. The main idea of this approach is to replace the category of sets with another category of smoothly varying sets which is a topos. In this category, one can define the real numbers, smooth functions, and so on, but the real numbers automatically contain nilpotent infinitesimals, so these do not need to be introduced by hand as in the algebraic geometric approach. However the logic in this new category is not identical to the familiar logic of the category of sets: in particular, the law of the excluded middle does not hold. This means that set-theoretic mathematical arguments only extend to smooth infinitesimal analysis if they are constructive (e.g., do not use proof by contradiction). Some regard this disadvantage as a positive thing, since it forces one to find constructive arguments wherever they are available.

The final approach to infinitesimals again involves extending the real numbers, but in a less drastic way. In the nonstandard analysis approach there are no nilpotent infinitesimals, only invertible ones, which may be viewed as the reciprocals of infinitely large numbers.[7] Such extensions of the real numbers may be constructed explicitly using equivalence classes of sequences of real numbers, so that, for example, the sequence (1, 1/2, 1/3, ..., 1/n, ...) represents an infinitesimal. The first-order logic of this new set of hyperreal numbers is the same as the logic for the usual real numbers, but the completeness axiom (which involves second-order logic) does not hold. Nevertheless, this suffices to develop an elementary and quite intuitive approach to calculus using infinitesimals, see transfer principle.