Skip to content

Continuous Dependence of Solution on Initial Value

Here we focus on

\[ \begin{equation} \dot{\mathbfit{y}} = \mathbfit{f}(x, \mathbfit{y},\pmb{\lambda}),\quad \mathbfit{y}(x_0) = \mathbfit{y}_0 \label{eq1} \end{equation} \]

where \(\mathbfit{x}\in \mathbb{R}^n\) and \(\mathbfit{f}: \mathbb{R}^{n+1} \mapsto \mathbb{R}^n \in C(D)\) is a vector function(or vector field).

Continuous Dependency

Continuous Dependency means that the solution of ODE with initial values would not differ very much from real solution when the errors of parameters are small enough.

Now we consider making a transformation. Let

\[ t=x-x_0,\quad \pmb{u}=\pmb{y}-\pmb{y}_0 \]

then

\[ \begin{align*} \frac{d \pmb{u}(t)}{dt} &= \frac{d\pmb{y}}{d x} \\ &=\pmb{f}(x,\pmb{y},\pmb{\lambda})\\ &=\pmb{f}(t+x_0,\pmb{u}+\pmb{y_0},\pmb{\lambda}) \end{align*} \]

and initial value becomes

\[ \pmb{u}(0)=\pmb{0} \]

So we only consider dependence on parameters \(\pmb{\lambda}\) of ODE

\[ \begin{equation} \frac{d\pmb{y}}{dx}=\pmb{f}(x,\pmb{y},\pmb{\lambda}),\quad \pmb{y}(0)=\pmb{0} \label{eq2} \end{equation} \]

Theorem for ODE with parameters

Consider a region

\[ G=\{(x,\pmb{y},\pmb{\lambda}):|x|\leq a,|\pmb{y}|\leq b, |\pmb{\lambda}|\leq c\} \]

and \(\pmb{f}(x,\pmb{y},\pmb{\lambda})\in C(G)\) and satisfies Lypschitz condition with respect to \(\pmb{y}\), i.e. \(\forall (x,\pmb{y}_1,\pmb{\lambda})\), \((x,\pmb{y}_2,\pmb{\lambda})\) we have

\[ |\pmb{f}(x,\pmb{y}_1,\pmb{\lambda})-\pmb{f}(x,\pmb{y}_2,\pmb{\lambda})|\leq L|\pmb{y}_1-\pmb{y}_2| \]

where \(L\) is Lypschitz constant. If we let

\[ M=\max_{(x,\pmb{y},\pmb{\lambda})\in G}\{\pmb{f}(x,\pmb{y},\pmb{\lambda})\}, \quad \alpha = \min\{a,b/M\} \]

then the solution \(\pmb{y}=\pmb{\phi}(x,\pmb{\lambda})\) of ODE \(\ref{eq2}\) is continuous on region

\[ D=\{|x|\leq \alpha, |\pmb{\lambda}|\leq c\} \]

Similar to Picard Theorem

Corollary

Assume region

\[ R=\{|x-x_0|\leq a, |\pmb{y}-\pmb{y}_0|\leq b\} \]

If \(\pmb{f}(x,\pmb{y})\in C(R)\) and satisfies Lypschitz condition with respect to \(\pmb{y}\). Then initial problem

\[ \frac{d \pmb{y}}{dx}=\pmb{f}(x,\pmb{y}),\quad \pmb{y}(x_0)=\pmb{\eta} \]

has a solution \(\pmb{y}=\pmb{\phi}(x,\pmb{\eta})\) is continuous on region

\[ Q=\left\{(x,\pmb{y}): |x-x_0|\leq \frac{\alpha}{2}, |\pmb{y}-\pmb{y}_0|\leq \frac{b}{2}\right\} \]

where

\[ M=\max_{(x,\pmb{y},\pmb{\lambda})\in G}\{\pmb{f}(x,\pmb{y},\pmb{\lambda})\}, \quad \alpha = \min\{a,b/M\}. \]

Actually, we do not need Lypschitz condition but the uniqueness of solution to ODE \(\ref{eq1}\). Here comes the following theorem.

Theorem for dependence on initial value

Consider ODE

\[ \frac{d\pmb{y}}{dx}=\pmb{f}(x,\pmb{y},\pmb{\lambda}) \]

where \(\pmb{f}\) is bounded and continuous on region \(G\subset \mathbb{R}\times \mathbb{R}^n \times \mathbb{R}^m\). Assume for all initial points \((x_0,\pmb{y}_0)\), the solution \(\pmb{y}=\pmb{\phi}(x;x_0,\pmb{y}_0,\pmb{\lambda})\) to the ODE exists uniquely on interval \(I_0\), where \(I_0\subset \mathbb{R}\) is finite. Then \(\forall \varepsilon>0\), \(\exists \delta>0\), \(\forall (\xi, \pmb{\eta}, \pmb{\lambda}')\in G\) s.t.

\[ |(\xi, \pmb{\eta}, \pmb{\lambda}')- (x_0, \pmb{y}_0, \pmb{\lambda})|<\delta \]

we have

\[ |\pmb{\phi}(x;\xi,\pmb{\eta},\pmb{\lambda}')-\pmb{\phi}(x;x_0,\pmb{y}_0,\pmb{\lambda})|<\varepsilon,\quad \forall x\in I_0 \]

By contradiction.

We only prove for ODE \(\ref{eq1}\).

  • Transfer the problem into another form

If we introduce \(\pmb{z}=\pmb{\lambda}\), then

\[ \frac{d\pmb{z}}{dx}=\pmb{0} \]

cause \(\pmb{\lambda}\) is a comstant. Then ODE \(\ref{eq1}\) becomes

\[ \frac{d\pmb{y}}{dx}=\pmb{f}(x,\pmb{y},\pmb{z}), \quad \frac{d\pmb{z}}{dx}=\pmb{0} \]

with initial value

\[ \pmb{y}(x_0)=\pmb{y}_0, \quad \pmb{z}(x_0)=\pmb{0} \]

We transfer the dependence on parameters into the dependence on initial values. In the following part, we only consider dependence on initial value.

  • By contradiction

Assume the above theorem does not hold, then \(\exists \varepsilon_0>0\), \(\forall \delta_n=1/n\), \(\exists(\xi_n,\pmb{\eta}_n, \pmb{\lambda}'_n)\in G\) with

\[ |(\xi_n,\pmb{\eta}_n, \pmb{\lambda}'_n)-(x_0,\pmb{y}_0,\pmb{\lambda})|<\delta_i \]

which satisfies that, there exsits \(x_n\in I_0\)

\[ \begin{equation} |\pmb{\phi}(x_n;\xi_n,\pmb{\eta}_n,\pmb{\lambda}'_n)-\pmb{\phi}(x_n;x_0,\pmb{y}_0,\pmb{\lambda})|\geq\varepsilon_0\label{ieq1} \end{equation}\]

Because \(I_0\) is bouned closed interval, so \(\{x_n\}\) has a convergent subsquence \(\{x_{n_j}\}\) which converges to \(\overline{x}\in I_0\).

Notice that

\[ \pmb{\phi}(x;\xi_n,\pmb{\eta}_n)=\pmb{\eta}_n+\int_{{\xi}_n}^{x}\pmb{f}(s,\pmb{\phi}(s;\xi_n,\pmb{\eta}_n))ds \]

it is easy to show that sequence \(\{\pmb{\phi}(x;\xi_n,\pmb{\eta}_n)\}\) is uniforly bounded and equicontinuous. Actually, if we assume \(M=\max_{(x,\pmb{y},\pmb{\lambda})\in G}|\pmb{f}|\) and denote \(\pmb{\psi}_n(x)=\pmb{\phi}(x;\xi_n,\pmb{\eta}_n)\), then

\[ |\pmb{\psi}_n(x)-\pmb{\eta}|\leq M|I_0|,\quad \forall n \]
\[ |\pmb{\psi}_n(x_1)-\pmb{\psi}_n(x_2)|\leq M|x_1-x_2|,\quad \forall n \]

Then by Ascoli-Arzelà Theorem, we have a uniforly convergent subsequence \(\{\pmb{\phi}(x;\xi_{n_j},\pmb{\eta}_{n_j})\}\) which converges to \(\pmb{\psi}(x)\) and they all satisfy

\[ \pmb{\phi}(x;\xi_{n_j},\pmb{\eta}_{n_j})=\pmb{\eta}_{n_j}+\int_{\xi_{n_j}}^{x}\pmb{f}(s,\pmb{\phi}(s;\xi_{n_j},\pmb{\eta}_{n_j}))ds \]

let \(j\rightarrow \infty\) and we get

\[ \pmb{\psi}(x)=\pmb{y}_0+\int_{x_0}^x\pmb{f}(x_0,\pmb{\psi}(x))ds \]

which means \(\pmb{\psi}(x)\equiv \pmb{\phi}(x;x_0,\pmb{y}_0)\).

  • Use the subsequence of \(\{x_{n_j}\}\)

from inequation \(\ref{ieq1}\), we have

\[ |\pmb{\psi}_{n_j}(x_{n_j})-\pmb{\psi}(x_{n_j})|\geq \varepsilon_0 \]

let \(j\rightarrow \infty\), we deduce

\[ |\pmb{\psi}(\overline{x})-\pmb{\psi}(\overline{x})|\geq \varepsilon_0 \]

which contradicts the uniqueness!

Here interval \(I_0\) is finite is necessary because in the proof we make use of the finite length of the interval. If a solution can satisfy the property on an infinite interval like \((\beta,\infty)\), then we call it Lyapunov stable.

Continuous Differentiability

Similar to the above part, if \(\pmb{f}(x,\pmb{y},\pmb{\lambda})\in C^1(G)\) in ODE \(\ref{eq1}\), then we can say the solution is continuous differential on a region. To be more specific, we have the following theorem.

Theorem for Continuous Differentiability

Assume \(\pmb{f}(x,\pmb{y},\pmb{\lambda})\in C(G)\) where

\[ G=\{(x,\pmb{y},\pmb{\lambda}):|x|\leq a,|\pmb{y}|\leq b, |\pmb{\lambda}|\leq c\} \]

and has continuous partial derivatives with respect to \(\pmb{y}\) and \(\pmb{\lambda}\). Then the solution \(\pmb{y}=\pmb{\phi}(x,\pmb{\lambda})\) to ODE \(\ref{eq1}\) is continuous differential on region \(D\), where

\[ D=\{|x|\leq \alpha, |\pmb{\lambda}|\leq c\} \]

Show that the partial derivatives can be the limit of divided difference.

Corollary 1

Assume \(\pmb{f}(x,\pmb{y})\in C(R)\) where

\[ R=\{(x,y): |x-x_0|\leq a, |\pmb{y}-\pmb{y}_0|\leq b\} \]

and has partial derivative with respect to \(\pmb{y}\), i.e. \(\pmb{f}'_{\pmb{y}}(x,\pmb{y})\). Then for all \(\pmb{\eta}\), s.t. \(|\pmb{\eta}-\pmb{y}_0|<\frac{b}{2}\), initial problem

\[ \frac{d\pmb{y}}{dx}=\pmb{f}(x,\pmb{y}), \quad \pmb{y}(x_0)=\pmb{\eta} \]

has solution \(\pmb{y}=\pmb{\phi}(x,\pmb{\eta})\) is conntinuous on region

\[ D=\left\{(x,y): |x-x_0|\leq \frac{h}{2}, |\pmb{\eta}-\pmb{y}_0|\leq \frac{b}{2}\right\} \]

Corollary 2

Assume \(\pmb{f}(x,\pmb{y})\in C(G)\) were

\[ G=\{(x,y): |x-x_0|\leq a, |\pmb{y}-\pmb{y}_0|\leq b,|\pmb{\lambda}|\leq c\} \]

and has partial derivative with respect to \(\pmb{y}\) and \(\pmb{\lambda}\). Then initial problem

\[ \begin{align} \frac{d\pmb{y}}{dx}=\pmb{f}(x,\pmb{y},\pmb{\lambda}), \quad \pmb{y}(x_0)=\pmb{y}_0\label{eq3} \end{align} \]

has unique solution \(\pmb{y}=\pmb{\phi}(x;x_0;\pmb{y}_0,\pmb{\lambda})\) which is differentiable for \((x_0,\pmb{y}_0,\pmb{\lambda})\) on region

\[ R=\{|x-x_0|\leq, |\pmb{\lambda}|\leq c\} \]

Solve for Derivatives

Assume \(\pmb{y}=\pmb{\phi}(x;x_0,\pmb{y}_0,\pmb{\lambda})\) is the solution to ODE \(\ref{eq3}\), then it satisfies the integral equation

\[ \begin{equation} \pmb{\phi}(x;x_0,\pmb{y}_0,\pmb{\lambda})=\pmb{y}_0+\int_{x_0}^x \pmb{f}(s,\pmb{\phi}(x;x_0,\pmb{y}_0,\pmb{\lambda}),\pmb{\lambda})ds \label{eq4} \end{equation} \]

Take partial derivative on both sides with respect to \(x_0\), \(\pmb{y}_0\), \(\pmb{\lambda}\), we get

\[ \begin{align*} \frac{\partial \pmb{\phi}}{\partial x_0}&= -\pmb{f}(x_0,\pmb{y}_0,\pmb{\lambda}) + \int_{x_0}^x \frac{\partial }{\partial \pmb{y}}\pmb{f}(s,\pmb{\phi}(x;x_0,\pmb{y}_0,\pmb{\lambda}),\pmb{\lambda}) \frac{\partial \pmb{\phi}}{\partial x_0}ds\\ \frac{\partial \pmb{\phi}}{\partial \pmb{y}_0}&=\pmb{E}_n+ \int_{x_0}^x \frac{\partial }{\partial \pmb{y}}\pmb{f}(s,\pmb{\phi}(x;x_0,\pmb{y}_0,\pmb{\lambda}),\pmb{\lambda}) \frac{\partial \pmb{\phi}}{\partial \pmb{y}_0}ds\\ \frac{\partial \pmb{\phi}}{\partial \pmb{\lambda}}&=\int_{x_0}^x \left(\frac{\partial }{\partial \pmb{y}}\pmb{f}(s,\pmb{\phi}(x;x_0,\pmb{y}_0,\pmb{\lambda}),\pmb{\lambda}) \frac{\partial \pmb{\phi}}{\partial \pmb{\lambda}}+\frac{\partial }{\partial \pmb{\lambda}}\pmb{f}(s,\pmb{\phi}(x;x_0,\pmb{y}_0,\pmb{\lambda}),\pmb{\lambda}) \right)ds \end{align*} \]

Or to be more concise:

\[ \begin{align*} \frac{\partial \pmb{\phi}}{\partial x_0}&= -\pmb{f}(x_0,\pmb{y}_0,\pmb{\lambda}) + \int_{x_0}^x \frac{\partial \pmb{f}}{\partial \pmb{y}}\frac{\partial \pmb{\phi}}{\partial x_0}ds\\ \frac{\partial \pmb{\phi}}{\partial \pmb{y}_0}&=\pmb{E}_n+ \int_{x_0}^x \frac{\partial \pmb{f}}{\partial \pmb{y}} \frac{\partial \pmb{\phi}}{\partial \pmb{y}_0}ds\\ \frac{\partial \pmb{\phi}}{\partial \pmb{\lambda}}&=\int_{x_0}^x \left(\frac{\partial \pmb{f}}{\partial \pmb{y}} \frac{\partial \pmb{\phi}}{\partial \pmb{\lambda}}+\frac{\partial }{\partial \pmb{\lambda}}\pmb{f} \right)ds \end{align*} \]

ODE for Derivatives

If we let

\[ \begin{align*} \pmb{u}(x;x_0,\pmb{y}_0,\pmb{\lambda}) &= \frac{\partial \pmb{\phi}}{\partial x_0}\\ \pmb{V}(x;x_0,\pmb{y}_0,\pmb{\lambda}) &= \frac{\partial \pmb{\phi}}{\partial \pmb{y}_0}\\ \pmb{W}(x;x_0,\pmb{y}_0,\pmb{\lambda}) &= \frac{\partial \pmb{\phi}}{\partial \pmb{\lambda}_0}\\ \pmb{A}(x;x_0,\pmb{y}_0,\pmb{\lambda}) &= \frac{\partial \pmb{f}}{\partial \pmb{y}}(x,\pmb{\phi},\pmb{\lambda})\\ \pmb{B}(x;x_0,\pmb{y}_0,\pmb{\lambda}) &= \frac{\partial \pmb{f}}{\partial \pmb{\lambda}}(x,\pmb{\phi},\pmb{\lambda})\\ \end{align*} \]

then \(\pmb{u}\), \(\pmb{V}\) and \(\pmb{W}\) satisfy ODE with initial value respectively

\[ \begin{align*} \frac{d\pmb{u}}{dx}&=\pmb{A}(x;x_0,\pmb{y}_0,\pmb{\lambda})\pmb{u},\quad \pmb{u}(x_0)=-\pmb{f}(x_0,\pmb{y}_0,\pmb{\lambda})\\ \frac{d\pmb{V}}{dx}&=\pmb{A}(x;x_0,\pmb{y}_0,\pmb{\lambda})\pmb{V},\quad\pmb{V}(x_0)=\pmb{E}_0\\ \frac{d\pmb{W}}{dx}&=\pmb{A}(x;x_0,\pmb{y}_0,\pmb{\lambda})\pmb{W}+\pmb{B}(x;x_0,\pmb{y}_0,\pmb{\lambda}),\quad \pmb{W}(x_0)=0 \end{align*} \]

If \(\pmb{y}\) and \(\pmb{\lambda}\) is one dimension, then the above ODE becomes much simpler. And we can use method from Elementary Integration Method.

Q1. Assume \(\phi(x;x_0,y_0,\mu)\) is the solution to ODE

\[ y'=y+\mu(x^2+y^2),\quad y|_{x=x_0}=y_0 \]

Solve \(\frac{\partial y}{\partial \mu}\) when \(x_0=0\), \(y_0=1\), \(\mu=0\).

Write the integral form

\[ \phi(x;x_0,y_0,\mu)=y_0+\int_{x_0}^x(\phi(s;x_0,y_0,\mu)+\mu(s+\phi(s;x_0,y_0,\mu)^2))ds \]

we only show the dependence of variables:

\[ \phi(\mu)=y_0+\int_{x_0}^x(\phi(\mu)+\mu(s+\phi(\mu)^2))ds \]

take derivative of \(\mu\), we get

\[ \frac{\partial \phi}{\partial \mu}=\int_{x_0}^x\left(\frac{\partial\phi}{\partial \mu}+s+\phi(\mu)^2+2\phi\mu\frac{\partial \phi}{\partial \mu}\right)ds \]

So function \(\frac{\partial\phi}{\partial \mu}\) satisfies ODE

\[ \begin{equation} u'=x+\phi^2+(1-2\mu\phi)u, \quad u(0)=0 \label{mu} \end{equation}\]

For initial value, we can solve the solution to ODE

\[ \phi=e^x \]

substitute initial value and solution into equation \(\ref{mu}\) and we get

\[ u'=x+e^{2x}+u,\quad u(0)=0 \]

so we get

\[ \begin{align*} u &= e^x\left[\int_{0}^x(s+e^{2s})e^{-s}ds\right]\\ &=e^{2x}-x-1 \end{align*} \]