6. Differential calculus of multivariate functions

Definition 1:

Let XX and YY be normed spaces. A mapping f:EYf: E \rightarrow Y of a set EXE \subset X into YY is differentiable at an interior point xEx \in E if there exists a continuous linear transformation L(x):XYL(x): X \rightarrow Y such that

f(x+h)f(x)=L(x)h+α(x;h)f(x+h)-f(x)=L(x) h+\alpha(x ; h)

where α(x;h)=o(h)\alpha(x ; h)=o(h) as h0,x+hEh \rightarrow 0, x+h \in E.

We often denote that L(x)L(x)as df(x),Df(x)\mathrm{d} f(x), D f(x), or f(x)f^{\prime}(x).

remark: If the condition for differentiability of the mapping ff at some point aa is written as

f(x)f(a)=L(x)(xa)+α(a;x)f(x)-f(a)=L(x)(x-a)+\alpha(a ; x)where α(a;x)=o(xa)\alpha(a;x) = o(x − a) as xax \to a.

tangent space

We remark that strictly speaking f(x)L(TXx;TYf(x))f^{\prime}(x) \in \mathcal{L}\left(T X_{x} ; T Y_{f(x)}\right) here and hh is a vector of the tangent space TXxT X_{x} . But parallel translation of a vector to any point xXx \in X is defined in a vector space, and this allows us to identify the tangent space TXxT X_{x} with the vector space XX itself.

Consequently, after choosing a basis in XX, we can extend it to all the tangent spaces TXxT X_{x} . This means that if, for example, X=RmX=\mathbb{R}^{m}, Y=RnY=\mathbb{R}^{n},and the mapping fL(Rm;Rn)f \in \mathcal{L}\left(\mathbb{R}^{m} ; \mathbb{R}^{n}\right) is given by the matrix (aij),\left(a_{ij}\right), then at every point xRmx \in \mathbb{R}^{m} the tangent mapping f(x):TRxmTRf(x)nf^{\prime}(x): T \mathbb{R}_{x}^{m} \rightarrow T \mathbb{R}_{f(x)}^{n} will be given by the same matrix.

Chain Rule

If the mapping f:UVf: U \rightarrow V is differentiable at a point xUXx \in U \subset X, and the mapping g:VZg: V \rightarrow Z is differentiable at f(x)=yVYf(x)=y \in V \subset Y, then the composition gfg \circ f of these mappings is differentiable at xx, and

(gf)(x)=g(f(x))f(x)(g \circ f)^{\prime}(x)=g^{\prime}(f(x)) \circ f^{\prime}(x)

Differentiation of the Inverse (#)

Let f:UYf: U \rightarrow Y be a mapping that is continuous at xUXx \in U \subset X and has an inverse f1:VXf^{-1}: V \rightarrow X that is defined in a neighborhood of y=f(x)y=f(x) and continuous at that point. If the mapping ff is differentiable at xx and its tangent mapping f(x)L(X;Y)f^{\prime}(x) \in \mathcal{L}(X ; Y) has a continuous inverse [f(x)]1L(Y;X),\left[f^{\prime}(x)\right]^{-1} \in \mathcal{L}(Y ; X), then the mapping f1f^{-1} is differentiable at y=f(x)y=f(x) and

[f1](f(x))=[f(x)]1\left[f^{-1}\right]^{\prime}(f(x))=\left[f^{\prime}(x)\right]^{-1}

Partial derivative

Let U=U(a)U=U(a) be a neighborhood of the point aX=X1××Xma \in X=X_{1} \times \cdots \times X_{m} in the direct product of the normed spaces X1,,XmX_{1}, \cdots, X_{m} , and let f:UYf: U \rightarrow Y be a mapping of UU into the normed space VV . In this case

y=f(x)=f(x1,,xm)y=f(x)=f\left(x_{1}, \ldots, x_{m}\right)

and hence, if we fix all the variables but xix_{i} in the above equation by setting xk=akx_{k}=a_{k} for k{1,,m}\i,k \in \{1, \ldots, m\} \backslash i, we obtain a function

f(a1,,ai1,xi,ai+1,,am)=:φi(xi)f\left(a_{1}, \ldots, a_{i-1}, x_{i}, a_{i+1}, \ldots, a_{m}\right)=: \varphi_{i}\left(x_{i}\right)

defined in some neighborhood UiU_{i} of aia_{i} in XX.

We usually denote this partial derivative by one of the symbols

if(a),Dif(a),fxi(a),fxi(a)\partial_{i} f(a), \quad D_{i} f(a), \quad \frac{\partial f}{\partial x_{i}}(a), \quad f_{x_{i}}^{\prime}(a)

In accordance with these definitions Dif(a)L(Xi;Y).D_{i} f(a) \in \mathcal{L}\left(X_{i} ; Y\right) . More precisely, Dif(a)L(TXi(ai);TY(f(a)))D_{i} f(a) \in \mathcal{L}\left(T X_{i}\left(a_{i}\right) ; T Y(f(a))\right).

Denote that h=(h1,,hm)TX1(a1)××TXm(am)=TX(a)h=\left(h_{1}, \ldots, h_{m}\right) \in T X_{1}\left(a_{1}\right) \times \cdots \times T X_{m}\left(a_{m}\right)=T X(a), we have the equation

df(a)h=1f(a)h1++mf(a)hm\mathrm{d} f(a) h=\partial_{1} f(a) h_{1}+\cdots+\partial_{m} f(a) h_{m}

partial derivative of higher order

Consider U=U(x)U=U(x) be a neighborhood of the point x=(x1,,xn)x=(x^1,\cdots,x^n), Then if ffhas partial derivative with respect to xix^i, then fxi\frac{\partial f}{\partial x^{i}}becomes a new function which is defined in UU, if:

xi1(xi2(fxik))(x)\frac{\partial}{\partial x^{i_{1}}}\left(\frac{\partial}{\partial x^{i_{2}}} \cdots\left(\frac{\partial f}{\partial x^{i_{k}}}\right) \cdots\right)(x)

exist, we call it higher-order partial derivative, often denoted as:

kfxi1xik(x)\frac{\partial^{k} f}{\partial x^{i_{1}} \cdots \partial x^{i_{k}}}(x)

Taylor's theorem

If a mapping f:UYf: U \rightarrow Y from a neighborhood U=U(x)U=U(x) of a point xx in a normed space XX into a normed space YY has derivatives up to order n1n-1 inclusive in UU and has an nth order derivative f(n)(x)f^{(n)}(x) at the point xx, then

f(x+h)=f(x)+f(x)h++1n!f(n)(x)hn+R(x;h)f(x+h)=f(x)+f^{\prime}(x) h+\cdots+\frac{1}{n !} f^{(n)}(x) h^{n}+R\left(x;h\right)

R(x;h)=o(hn)R\left(x;h\right)= o(|h|^n)when h0h\to 0.

We can expand the f(n)(x)f^{(n)}(x), such as Jacobian matrix:

Tf(x)=df(x)=f(x)=(fx1(x),fx2(x),,fxn(x))\mathbf {T}f(x)= \mathrm df(x)=\nabla f(x)=\left({\frac {\partial f}{\partial x_{1}}}(x),{\frac {\partial f}{\partial x_{2}}}(x),\cdots ,{\frac {\partial f}{\partial x_{n}}}(x)\right)

or Hessian matrix:

Hf(x)=(2fx122fx1x22fx1xn2fx2x12fx222fx2xn2fxnx12fxnx22fxn2)\mathbf{H} f(x)=\left(\begin{array}{cccc} \frac{\partial^{2} f}{\partial x_{1}^{2}} & \frac{\partial^{2} f}{\partial x_{1} \partial x_{2}} & \cdots & \frac{\partial^{2} f}{\partial x_{1} \partial x_{n}} \\&&\\ \frac{\partial^{2} f}{\partial x_{2} \partial x_{1}} & \frac{\partial^{2} f}{\partial x_{2}^{2}} & \cdots & \frac{\partial^{2} f}{\partial x_{2} \partial x_{n}} \\&&\\ \vdots & \vdots & \ddots & \vdots \\&&\\ \frac{\partial^{2} f}{\partial x_{n} \partial x_{1}} & \frac{\partial^{2} f}{\partial x_{n} \partial x_{2}} & \cdots & \frac{\partial^{2} f}{\partial x_{n}^{2}} \end{array}\right)

fn(x)f^n(x)means all the nn-th order partial derivatives of ff

f(n)(x)hn=α1++αn=nnfx1α1xnαn(x)h1α1hnαnf^{(n)}(x)h^n=\sum_{\alpha _{1}+\cdots+\alpha _{n}=n}{\frac {\partial ^{n}f}{\partial x_{1}^{\alpha _{1}}\cdots \partial x_{n}^{\alpha_ {n}}}}(x)\cdot h_1^{\alpha_1}\cdots h_n^{\alpha_n}

and

R(x;h)=1n!01(1t)nDn+1f(x+th))dtR\left(x;h\right)=\frac{1}{n !} \int_{0}^{1}(1-t)^{n} D^{n+1} f(x+t\cdot h)) \mathrm d t

Lagrange multiplier

... pass.

\blacksquare

Last updated

Was this helpful?