A differential equation becomes ordinary when the unknown function depends on one independent variable. This restriction turns questions about motion, growth, decay, oscillation, and flow into questions about curves. The central problem is not only to solve an equation, but to understand when the equation determines a curve, how long that curve exists, how it changes when the initial state changes, and what qualitative behaviour is forced by the vector field. The first warning is that integration alone does not explain ordinary differential equations. A formula such as $\dot{x}=f(t)$ can be solved by taking an antiderivative, but a law such as $\dot{x}=x^2$ asks the current state to determine its own future rate of change. The solution is a trajectory, and the equation can break down in finite time even when the velocity rule is smooth. [example: Finite-Time Blow-Up for $\dot{x}=x^2$] Let $x:(-\infty,1)\to \mathbb{R}$ be defined by \begin{align*} x(t)=\frac{1}{1-t}=(1-t)^{-1}. \end{align*} At the initial time, \begin{align*} x(0)=\frac{1}{1-0}=1. \end{align*} For every $t<1$, differentiating $(1-t)^{-1}$ by the chain rule gives \begin{align*} \dot{x}(t)=(-1)(1-t)^{-2}\cdot(-1)=(1-t)^{-2}. \end{align*} Since \begin{align*} x(t)^2=\left(\frac{1}{1-t}\right)^2=\frac{1}{(1-t)^2}=(1-t)^{-2}, \end{align*} we have $\dot{x}(t)=x(t)^2$ for every $t\in(-\infty,1)$. Thus $x$ solves the initial value problem $\dot{x}=x^2$, $x(0)=1$, on $(-\infty,1)$. The endpoint $t=1$ is not a removable endpoint for a finite real-valued extension: if $M>0$ and $t\in(1-1/M,1)$, then $0<1-t<1/M$, so \begin{align*} x(t)=\frac{1}{1-t}>M. \end{align*} Hence $x(t)\to\infty$ as $t\uparrow 1$, so the solution exists locally but blows up in finite forward time. [/example] This example separates three issues that are often conflated. There may be a local solution, it may be unique, and it may still fail to exist for all time. The theory of ordinary differential equations begins by making those questions precise. ## Definition The basic object is an equation prescribing a derivative. To avoid hiding analytic assumptions in notation, we specify the interval of time, the state space, and the regularity expected of the unknown curve. This is the minimal framework in which the phrase "the derivative is determined by the present state" has a precise meaning. [definition: Ordinary Differential Equation] Let $m,n,q\in \mathbb{N}$, let $I\subset \mathbb{R}$ be an interval, let $\Omega\subset(\mathbb{R}^n)^{m+1}$ be open, and let \begin{align*} \mathcal{F}:I\times \Omega\to \mathbb{R}^q \end{align*} be a function. An ordinary differential equation of order $m$ is an equation of the form \begin{align*} \mathcal{F}(t,x(t),\dot{x}(t),\ldots,x^{(m)}(t))=0, \end{align*} where $0$ denotes the zero vector in $\mathbb{R}^q$ and the unknown is an $m$-times differentiable curve $x:J\to \mathbb{R}^n$ on an interval $J\subset I$ depending on the single independent variable $t$, with $(x(t),\dot{x}(t),\ldots,x^{(m)}(t))\in\Omega$ wherever the equation is imposed. [/definition] This definition captures the parent idea: ordinary differential equations are equations in one independent variable, not necessarily first-order equations. It also includes implicit equations, where the highest derivative may not be isolated. Most of the theory below is developed in first-order normal form because that is the setting in which existence, uniqueness, flows, and stability have their cleanest statements; passing from an implicit equation to normal form can fail at singular points, can introduce branches, and can change the natural domain of the problem. ## Initial Value Problems and Integral Form The most important normal form is the one in which the highest derivative is already isolated. It is the form used for existence, uniqueness, flows, and stability. The word ordinary records that the curve has one independent variable, usually time. The value $x(t)$ may have many components, so a single ordinary differential equation can encode a coupled system in $\mathbb{R}^n$. To turn the equation into a prediction rather than a family of possible motions, we must say where the trajectory starts. Initial data are the part of the model that connect the differential law to a particular experiment. Two falling particles may obey the same differential equation while starting from different positions and velocities. The initial value problem packages the law and the starting state together. [definition: Initial Value Problem] Let $I\subset \mathbb{R}$ be an interval, let $U\subset \mathbb{R}^n$ be open, let $F:I\times U\to \mathbb{R}^n$, let $t_0\in I$, and let $x_0\in U$. The initial value problem for $F$ with initial data $(t_0,x_0)$ is the system consisting of \begin{align*} \dot{x}(t)=F(t,x(t)) \end{align*} and \begin{align*} x(t_0)=x_0, \end{align*} for an unknown curve $x:J\to U$ on an interval $J\subset I$ with $t_0\in J$, differentiable at interior points of $J$ and one-sided differentiable at endpoints of $J$ when endpoint derivatives are used. [/definition] The interval is part of the data of a solution, not an afterthought. A curve that satisfies the equation before a singular time and a curve that can be continued through that time have different mathematical content. The next definition records what it means to solve the problem on a specified interval while remaining inside the state space. [definition: Solution of an Initial Value Problem] Let $I\subset \mathbb{R}$ be an interval, let $U\subset \mathbb{R}^n$ be open, let $F:I\times U\to \mathbb{R}^n$, and let $(t_0,x_0)\in I\times U$. A solution of the initial value problem is a curve $x:J\to U$, where $J\subset I$ is an interval with $t_0\in J$, such that $x$ is differentiable at every interior point of $J$ and one-sided differentiable at endpoints of $J$ when those endpoints belong to $J$, \begin{align*} x(t_0)=x_0 \end{align*} and \begin{align*} \dot{x}(t)=F(t,x(t)) \quad \text{for every } t\in J \text{ where the derivative is defined as above.} \end{align*} [/definition] Throughout this page, $B(x_0,r)$ denotes the open Euclidean ball $\{x\in\mathbb{R}^n:|x-x_0|<r\}$, $\overline{B}(x_0,r)$ denotes the closed Euclidean ball $\{x\in\mathbb{R}^n:|x-x_0|\le r\}$, and $\mathcal{L}^1$ denotes [Lebesgue measure](/page/Lebesgue%20Measure) on the time line. In matrix formulas, $I_n$ denotes the $n\times n$ identity matrix, and $I$ denotes the same matrix when the dimension is fixed by context. These conventions let the integral and neighbourhood statements below keep their formulas short. The differential definition is the right conceptual object, but it is not always the right analytic object. To prove existence, compare two possible solutions, or apply fixed point methods, we need an equivalent formulation involving only integration of the vector field along the unknown curve. [quotetheorem:8311] The integral form explains why continuity of $F$ is enough to ask for existence: integrating a continuous expression produces a differentiable curve. Uniqueness, however, needs more than continuity, and that distinction becomes visible once equations are treated as systems. ## First-Order Systems and Reduction of Order ### Systems as State-Space Equations Many equations first appear with higher derivatives, such as Newton's law $\ddot{q}=G(t,q,\dot{q})$. The first-order formalism is still general because higher derivatives can be treated as new state variables. This is why first-order systems are the natural language of the subject. [definition: First-Order System] Let $I\subset \mathbb{R}$ be an interval and let $U\subset \mathbb{R}^n$ be open. A first-order system on $I\times U$ is an ordinary differential equation \begin{align*} \dot{x}(t)=F(t,x(t)), \end{align*} where $F:I\times U\to \mathbb{R}^n$ and $x:J\to U$ is an unknown curve. [/definition] A scalar equation of order $m$ stores position, velocity, and successive derivatives as one vector. This conversion is not cosmetic: existence and uniqueness theorems are usually stated for first-order systems. To use those theorems for Newtonian equations, we need a formal definition of the higher-order equation being converted. [definition: Scalar Ordinary Differential Equation of Order $m$] Let $m\in \mathbb{N}$, let $I\subset \mathbb{R}$ be an interval, let $V\subset \mathbb{R}^m$ be open, and let $G:I\times V\to \mathbb{R}$. A scalar ordinary differential equation of order $m$ is an equation \begin{align*} y^{(m)}(t)=G(t,y(t),\dot{y}(t),\ldots,y^{(m-1)}(t)) \end{align*} for an unknown $m$-times differentiable function $y:J\to \mathbb{R}$ whose jet $(y,\dot{y},\ldots,y^{(m-1)})$ lies in $V$. [/definition] After this definition, the state variables are forced rather than chosen by convenience: the first coordinate is position, the next is velocity, and so on. The reduction theorem says that no information is lost in this packaging, so first-order theory is broad enough for higher-order scalar equations. [quotetheorem:8312] ### Phase Variables The harmonic oscillator shows the meaning of this conversion. A second-order equation becomes a planar vector field, and questions about oscillation become questions about curves in the phase plane. [example: The Harmonic Oscillator as a First-Order System] Let $\omega>0$ and consider \begin{align*} \ddot{q}(t)+\omega^2q(t)=0. \end{align*} Set $x_1=q$ and $x_2=\dot{q}$. Then $\dot{x}_1=\dot{q}=x_2$, and the equation $\ddot{q}+\omega^2q=0$ gives $\ddot{q}=-\omega^2q$, so \begin{align*} \dot{x}_2=\ddot{q}=-\omega^2q=-\omega^2x_1. \end{align*} Thus the second-order equation becomes the planar system \begin{align*} \dot{x}_1=x_2 \end{align*} and \begin{align*} \dot{x}_2=-\omega^2x_1. \end{align*} Define $E:\mathbb{R}^2\to\mathbb{R}$ by \begin{align*} E(x_1,x_2)=\frac{1}{2}x_2^2+\frac{1}{2}\omega^2x_1^2. \end{align*} Along any solution, the chain rule gives \begin{align*} \frac{d}{dt}E(x_1(t),x_2(t))=x_2(t)\dot{x}_2(t)+\omega^2x_1(t)\dot{x}_1(t). \end{align*} Substituting $\dot{x}_1=x_2$ and $\dot{x}_2=-\omega^2x_1$ gives \begin{align*} \frac{d}{dt}E(x_1(t),x_2(t))=x_2(t)(-\omega^2x_1(t))+\omega^2x_1(t)x_2(t). \end{align*} The two terms cancel, so \begin{align*} \frac{d}{dt}E(x_1(t),x_2(t))=0. \end{align*} Hence $E$ is constant along solutions, and every phase curve lies in a level set \begin{align*} E(x_1,x_2)=c. \end{align*} For $c>0$, this level set is \begin{align*} \omega^2x_1^2+x_2^2=2c, \end{align*} an ellipse in the $(x_1,x_2)$ plane. To see the rotation explicitly, introduce scaled coordinates \begin{align*} y_1=\omega x_1 \end{align*} and \begin{align*} y_2=x_2. \end{align*} Then \begin{align*} \dot{y}_1=\omega\dot{x}_1=\omega x_2=\omega y_2 \end{align*} and \begin{align*} \dot{y}_2=\dot{x}_2=-\omega^2x_1=-\omega y_1. \end{align*} If $y_1(0)=a$ and $y_2(0)=b$, the functions \begin{align*} y_1(t)=a\cos(\omega t)+b\sin(\omega t) \end{align*} and \begin{align*} y_2(t)=b\cos(\omega t)-a\sin(\omega t) \end{align*} satisfy \begin{align*} \dot{y}_1(t)=-a\omega\sin(\omega t)+b\omega\cos(\omega t)=\omega y_2(t) \end{align*} and \begin{align*} \dot{y}_2(t)=-b\omega\sin(\omega t)-a\omega\cos(\omega t)=-\omega y_1(t). \end{align*} Since $\cos(\omega(t+2\pi/\omega))=\cos(\omega t+2\pi)=\cos(\omega t)$ and $\sin(\omega(t+2\pi/\omega))=\sin(\omega t+2\pi)=\sin(\omega t)$, both $y_1$ and $y_2$ return to their initial values after time $2\pi/\omega$. On a level set with $c>0$, the vector field $(x_2,-\omega^2x_1)$ cannot vanish, because vanishing would force $x_2=0$ and $x_1=0$, which gives $E(0,0)=0$ rather than $c>0$. Thus the nonzero energy ellipses are periodic phase curves: they record recurring oscillation, not merely bounded motion. [/example] This phase-plane viewpoint replaces a search for closed-form solutions by geometric information about trajectories. It also prepares the key analytic question: when does a vector field determine exactly one trajectory through a given point? ## Existence and Uniqueness ### Existence Without Uniqueness The central local question is whether the velocity rule actually determines a trajectory through the prescribed point. Continuity of $F$ is enough to create at least one local curve, because the integral formulation can be approached by compactness. It does not prevent two curves from leaving the same initial point in different ways. In Peano's theorem, the Cauchy problem is the paired system consisting of the initial condition $X(0)=X_0$ and the evolution rule $\frac{d}{dt}X(t)=f(X(t))$. The quoted result is needed because it answers the first question about this combined system: under continuity alone, at least one local trajectory starts from the prescribed point. It does not yet answer whether that trajectory is the only possible one. [quotetheorem:68] Peano's theorem is deliberately only an existence result: continuity gives a local solution, but it does not impose enough control on the vector field to identify a unique future. The missing ingredient is visible in examples where the vector field is continuous but too flat near the initial point, allowing a solution to wait before moving. Existence alone is too weak for deterministic modelling. If the same starting point can generate several futures, the equation does not define a flow. [example: Non-Uniqueness for a Continuous Vector Field] Consider the scalar initial value problem \begin{align*} \dot{x}(t)=2\sqrt{|x(t)|} \end{align*} with \begin{align*} x(0)=0. \end{align*} The constant function $x(t)=0$ satisfies $x(0)=0$ and has $\dot{x}(t)=0$, while \begin{align*} 2\sqrt{|x(t)|}=2\sqrt{0}=0, \end{align*} so it is a solution. For every $a\ge 0$, define $x_a:[0,\infty)\to\mathbb{R}$ by \begin{align*} x_a(t)=\max\{t-a,0\}^2. \end{align*} At the initial time, \begin{align*} x_a(0)=\max\{-a,0\}^2=0, \end{align*} because $a\ge 0$. If $0\le t<a$, then $x_a(t)=0$, so $\dot{x}_a(t)=0$ and \begin{align*} 2\sqrt{|x_a(t)|}=2\sqrt{0}=0. \end{align*} If $t>a$, then $x_a(t)=(t-a)^2$, so \begin{align*} \dot{x}_a(t)=2(t-a). \end{align*} For the right-hand side, since $t-a>0$, \begin{align*} 2\sqrt{|x_a(t)|}=2\sqrt{(t-a)^2}=2|t-a|=2(t-a). \end{align*} At $t=a$, one has $x_a(a)=0$. For $h<0$ with $a+h\ge 0$, \begin{align*} \frac{x_a(a+h)-x_a(a)}{h}=\frac{0-0}{h}=0. \end{align*} For $h>0$, \begin{align*} \frac{x_a(a+h)-x_a(a)}{h}=\frac{h^2}{h}=h. \end{align*} Both one-sided difference quotients tend to $0$, so $\dot{x}_a(a)=0$, and the equation also holds there because \begin{align*} 2\sqrt{|x_a(a)|}=2\sqrt{0}=0. \end{align*} Thus every $x_a$ solves the same initial value problem on $[0,\infty)$. The vector field $f(x)=2\sqrt{|x|}$ is continuous, since $x\mapsto |x|$, $u\mapsto \sqrt{u}$ on $[0,\infty)$, and scalar multiplication are continuous. It is not Lipschitz near $0$: if $|f(x)-f(0)|\le L|x-0|$ held for all sufficiently small $x>0$, then \begin{align*} 2\sqrt{x}\le Lx. \end{align*} Dividing by $\sqrt{x}>0$ gives \begin{align*} 2\le L\sqrt{x}, \end{align*} which fails for any $0<x<4/L^2$. The parameter $a$ is a waiting time: the solution remains at the initial point until $t=a$ and then follows $(t-a)^2$, so continuity alone does not force uniqueness. [/example] ### The Lipschitz Condition The missing condition is a quantitative control on how much the vector field changes when the state changes. A Lipschitz estimate turns the distance between two possible solutions into an integral inequality. That estimate is exactly what rules out the waiting-time behaviour in the previous example. [definition: Local Lipschitz Continuity in the State Variable] Let $I\subset \mathbb{R}$ be an interval, let $U\subset \mathbb{R}^n$ be open, and let $F:I\times U\to \mathbb{R}^n$. The function $F$ is locally Lipschitz in the state variable if for every compact interval $K\subset I$ and every compact set $C\subset U$, there exists $L>0$ such that \begin{align*} |F(t,x)-F(t,y)|\le L|x-y| \end{align*} for all $t\in K$ and all $x,y\in C$. [/definition] Once this condition is imposed, the initial value problem behaves like a deterministic local law. The theorem is local because even smooth vector fields can blow up in finite time, as $\dot{x}=x^2$ already showed. The Cauchy problem in the theorem consists of the evolution equation together with the initial condition, so existence and uniqueness are asserted for that combined system. The quoted version below is stated in a slightly more general language than the Euclidean notation used so far. For this page, specialize its [Banach space](/page/Banach%20Space) $E$ to $\mathbb{R}^n$ with the usual norm, so $\|\cdot\|_E$ is just the Euclidean norm. The ball $B_{\mathbb{R}\times E}(u_0,r)$ is a neighbourhood of the initial point $u_0=(t_0,x_0)$ in time-state space. A function $k\in L^1(t_0-r,t_0+r)$ is an integrable time-dependent Lipschitz bound, and $C^1(I_\tau;E)$ means continuously differentiable $E$-valued curves on a smaller interval $I_\tau$. With these translations, the theorem is exactly the local existence-and-uniqueness principle for the Euclidean initial value problems discussed here. [quotetheorem:69] Picard--Lindelof is the local determinism theorem for ordinary differential equations. The Lipschitz hypothesis is the point: it lets differences between two candidate solutions be controlled by their accumulated past differences, so the non-uniqueness mechanism from the continuous example is no longer available. The conclusion is still local, since uniqueness does not by itself prevent later blow-up or escape from the domain. The theorem licenses notation: we can speak of the trajectory starting at $x_0$ at time $t_0$ because there is only one such trajectory locally. The next question is how far that trajectory can be followed. ## Maximal Solutions and Global Behaviour ### Extension and Maximality Local solutions can often be extended, but not indefinitely. A solution may stop because time leaves the allowed interval, because the trajectory reaches the boundary of the state space, or because its norm becomes unbounded. To distinguish an artificial endpoint from a real obstruction, we first need a notion of extension. [definition: Extension of a Solution] Let $I\subset \mathbb{R}$ be an interval, let $U\subset \mathbb{R}^n$ be open, let $F:I\times U\to \mathbb{R}^n$, and let $(t_0,x_0)\in I\times U$. Let $x:J\to U$ and $y:K\to U$ be solutions of the initial value problem determined by $F$ and $(t_0,x_0)$, where $J\subset K\subset I$ are intervals. The solution $y$ is an extension of $x$ if \begin{align*} y(t)=x(t) \quad \text{for all } t\in J. \end{align*} [/definition] A maximal solution is one that cannot be enlarged without leaving the class of solutions. This definition is needed because local existence gives many restricted solutions, but only the maximal interval records the full lifespan determined by the equation and initial data. [definition: Maximal Solution] Let $I\subset \mathbb{R}$ be an interval, let $U\subset \mathbb{R}^n$ be open, and let $F:I\times U\to \mathbb{R}^n$. A solution $x:J\to U$ of an initial value problem is maximal if there is no solution $y:K\to U$ of the same initial value problem such that $J\subsetneq K$ and $y$ extends $x$. [/definition] Local existence gives a solution only on a small interval, and different applications may produce different small intervals around the same initial time. The next theorem is needed to remove that dependence on an arbitrary local construction: under local uniqueness, all these compatible pieces assemble into a single largest trajectory, whose interval of definition is an invariant of the equation and the initial state. [quotetheorem:8313] Finite-time blow-up is not the only obstruction. If the state space removes a point, a bounded trajectory can still cease to be a solution after hitting the missing boundary. [example: Failure by Leaving the State Space] Let $U=(0,\infty)$ and consider the initial value problem \begin{align*} \dot{x}(t)=-1 \end{align*} with \begin{align*} x(0)=1. \end{align*} If $x:J\to U$ is any solution on an interval $J$ containing $0$, then for every $t\in J$ the [fundamental theorem of calculus](/theorems/632) gives \begin{align*} x(t)-x(0)=\int_0^t \dot{x}(s)\,d\mathcal{L}^1(s). \end{align*} Substituting $\dot{x}(s)=-1$ and $x(0)=1$ gives \begin{align*} x(t)-1=\int_0^t (-1)\,d\mathcal{L}^1(s)=-t. \end{align*} Hence every such solution satisfies \begin{align*} x(t)=1-t. \end{align*} This formula takes values in $U=(0,\infty)$ exactly when \begin{align*} 1-t>0. \end{align*} Equivalently, it remains in the state space exactly for $t<1$. Therefore the curve \begin{align*} x:(-\infty,1)\to(0,\infty), \qquad x(t)=1-t \end{align*} solves the initial value problem on $(-\infty,1)$, since $\dot{x}(t)=-1$ for every $t<1$ and $x(0)=1$. As $t\uparrow 1$, the values satisfy \begin{align*} \lim_{t\uparrow 1}x(t)=\lim_{t\uparrow 1}(1-t)=0. \end{align*} The limit is finite, but it is not a point of $U$. If an extension to an interval containing some time $t>1$ existed, that interval would also contain $1$, and continuity of the extended solution would force its value at $1$ to be $0$. This contradicts the requirement that the solution take values in $(0,\infty)$. Thus the obstruction is not blow-up: the trajectory reaches the boundary of the state space in finite time. [/example] ### Continuation Criteria The obstruction principle says that a solution can fail to continue only by running out of time or by escaping every compact subset of the domain. This result is the rigorous version of the local nature of existence: if the solution remains in a region where the hypotheses are still available, the local theorem can be restarted. [quotetheorem:8314] The continuation criterion is powerful, but it still asks us to control a solution whose formula may be unavailable. A usable global existence theorem should replace that unknown trajectory information by a direct hypothesis on the vector field. Linear growth is exactly such a condition: it prevents the speed from increasing faster than proportionally to the distance from the origin, so Gronwall-type estimates can keep the trajectory inside bounded regions on every finite time interval. [quotetheorem:8315] The theorem says that local uniqueness plus a growth estimate gives a global dynamical system. Without the growth estimate, smoothness alone does not stop a solution from escaping to infinity. ## Autonomous Equations and Flows ### Vector Fields and Equilibria When the vector field does not depend explicitly on time, the equation has time-translation symmetry. The same rule applies today and tomorrow, so solutions can be restarted from their current state. This gives the subject its dynamical-systems form. [definition: Autonomous Ordinary Differential Equation] Let $U\subset \mathbb{R}^n$ be open and let $f:U\to \mathbb{R}^n$. An autonomous ordinary differential equation is an equation \begin{align*} \dot{x}(t)=f(x(t)) \end{align*} for an unknown differentiable curve $x:J\to U$. [/definition] To draw phase portraits and reason geometrically, we need to name the arrow assignment on the state space itself. This separates the vector field, which is the static geometric data, from its integral curves, which are the motions generated by that data. [definition: Vector Field on an Open Subset of $\mathbb{R}^n$] Let $U\subset \mathbb{R}^n$ be open. A vector field on $U$ is a function \begin{align*} f:U\to \mathbb{R}^n. \end{align*} [/definition] Once a vector field is named, the first geometric question is where its arrows vanish. Those points are candidates for rest states, and they organize nearby trajectories even when no explicit solution formula is available. [definition: Equilibrium Point] Let $U\subset \mathbb{R}^n$ be open and let $f:U\to \mathbb{R}^n$ be a vector field. A point $x^*\in U$ is an equilibrium point of $\dot{x}=f(x)$ if \begin{align*} f(x^*)=0. \end{align*} [/definition] The definition should agree with the dynamical picture: zero velocity should mean no motion, and no motion should mean zero velocity. This small theorem is often the first check when translating between vector fields and trajectories. [quotetheorem:8316] This equivalence is small but structural. It lets the geometric zeros of the vector field be treated as actual trajectories, so equilibria are not a separate kind of object from solutions. In practice, it changes the first qualitative task from solving the differential equation to solving the algebraic equation $f(x)=0$. For the logistic equation, for instance, the equilibria $0$ and $K$ are found before one writes the explicit solution formula, and they already predict the two special trajectories that remain fixed for all time. The theorem also marks a limitation. It identifies exactly the trajectories that never move, but it does not say whether nearby nonconstant trajectories approach, leave, oscillate around, or fail to exist for all forward time. Those questions require comparing whole solution curves rather than a single value of the vector field. The next construction packages those curves into flow maps, so an equilibrium becomes a fixed point of the time-evolution map and can be studied through the behaviour of nearby initial states. ### Flow Maps For autonomous equations with uniqueness, a trajectory can be stopped and restarted without changing its future. To express this restart property, we collect all solutions into a single map from time and initial state to current state. For a vector field $f:U\to\mathbb{R}^n$, the phrase locally Lipschitz means the one-variable version of the earlier condition: for every compact set $C\subset U$, there exists $L>0$ such that $|f(x)-f(y)|\le L|x-y|$ for all $x,y\in C$. This is the regularity assumption that gives a unique maximal trajectory through each initial state. [definition: Flow Domain] Let $U\subset \mathbb{R}^n$ be open and let $f:U\to \mathbb{R}^n$ be locally Lipschitz. For each $x_0\in U$, let $J_{x_0}\subset \mathbb{R}$ be the maximal interval of the solution of \begin{align*} \dot{x}(t)=f(x(t)) \end{align*} with \begin{align*} x(0)=x_0. \end{align*} The flow domain is \begin{align*} D=\{(t,x_0)\in \mathbb{R}\times U:t\in J_{x_0}\}. \end{align*} [/definition] Once the possible times have been collected into a single set, the flow itself can be defined as an honest map with a specified domain and codomain. This matters because finite-time blow-up means the natural domain need not be all of $\mathbb{R}\times U$. [definition: Flow Map] Let $U\subset \mathbb{R}^n$ be open, let $f:U\to \mathbb{R}^n$ be locally Lipschitz, and let $D\subset \mathbb{R}\times U$ be the flow domain. The flow map of $\dot{x}=f(x)$ is the function $\varphi:D\to U$ defined by \begin{align*} \varphi(t,x_0)=x_{x_0}(t), \end{align*} where $x_{x_0}:J_{x_0}\to U$ is the maximal solution with $x_{x_0}(0)=x_0$. [/definition] A flow map should behave like time evolution rather than an arbitrary two-variable function. The composition law is needed to express that evolving for time $s$ and then for time $t$ gives the same state as evolving for the combined time whenever all terms are defined. [quotetheorem:8317] The composition law turns local uniqueness into an algebraic structure: time evolution can be composed wherever the relevant pieces of trajectory exist. In models with global solutions this becomes an action of the additive group $\mathbb{R}$ on the state space; in models with finite escape time it remains a partial action on the flow domain. Logistic growth is a useful example because positive initial data give forward-global positive solutions, and the long-time behaviour can be read from both the sign of the vector field and the precise interval on which the separated formula is defined. [example: Logistic Growth and Equilibria] Let $r>0$, $K>0$, and define $f:(0,\infty)\to\mathbb{R}$ by \begin{align*} f(x)=rx\left(1-\frac{x}{K}\right). \end{align*} Since $r>0$ and $x>0$ on $U=(0,\infty)$, the equation $f(x)=0$ is equivalent to \begin{align*} 1-\frac{x}{K}=0. \end{align*} Thus the only equilibrium in $U$ is $x=K$. If the same formula is extended to $[0,\infty)$, then \begin{align*} f(0)=r\cdot 0\left(1-\frac{0}{K}\right)=0, \end{align*} so $0$ becomes a boundary equilibrium of the extended model. The sign of the vector field is also explicit. If $0<x<K$, then $x/K<1$, hence $1-x/K>0$, and therefore \begin{align*} rx\left(1-\frac{x}{K}\right)>0. \end{align*} If $x>K$, then $x/K>1$, hence $1-x/K<0$, and therefore \begin{align*} rx\left(1-\frac{x}{K}\right)<0. \end{align*} Now fix an initial value $x(0)=x_0>0$ and set \begin{align*} A=\frac{K-x_0}{x_0}. \end{align*} Consider \begin{align*} x(t)=\frac{K}{1+Ae^{-rt}} \end{align*} where the denominator is nonzero. At $t=0$, \begin{align*} 1+Ae^{0}=1+A=1+\frac{K-x_0}{x_0}=\frac{K}{x_0}, \end{align*} so \begin{align*} x(0)=\frac{K}{K/x_0}=x_0. \end{align*} Differentiating $K(1+Ae^{-rt})^{-1}$ gives \begin{align*} \dot{x}(t)=K(-1)(1+Ae^{-rt})^{-2}(-rAe^{-rt})=\frac{KrAe^{-rt}}{(1+Ae^{-rt})^2}. \end{align*} On the other hand, \begin{align*} 1-\frac{x(t)}{K}=1-\frac{1}{1+Ae^{-rt}}=\frac{Ae^{-rt}}{1+Ae^{-rt}}. \end{align*} Therefore \begin{align*} rx(t)\left(1-\frac{x(t)}{K}\right)=r\frac{K}{1+Ae^{-rt}}\frac{Ae^{-rt}}{1+Ae^{-rt}}=\frac{KrAe^{-rt}}{(1+Ae^{-rt})^2}. \end{align*} Hence $\dot{x}(t)=rx(t)(1-x(t)/K)$ wherever the formula is defined. For forward time, the formula is defined and positive for every $t\ge 0$. If $0<x_0<K$, then $A>0$, so \begin{align*} 1+Ae^{-rt}>1>0. \end{align*} If $x_0=K$, then $A=0$, so \begin{align*} x(t)=\frac{K}{1}=K. \end{align*} If $x_0>K$, then $-1<A<0$, because $0<x_0-K<x_0$, and for $t\ge 0$ one has $0<e^{-rt}\le 1$, so \begin{align*} 1+Ae^{-rt}\ge 1+A=\frac{K}{x_0}>0. \end{align*} Thus every positive initial value gives a positive forward solution. When $x_0>K$, the denominator vanishes at a negative time. Indeed $A<0$, and \begin{align*} 1+Ae^{-rt}=0 \end{align*} is equivalent to \begin{align*} e^{-rt}=-\frac{1}{A}. \end{align*} Taking logarithms gives \begin{align*} -rt=\log\left(-\frac{1}{A}\right)=-\log(-A), \end{align*} so \begin{align*} t_*=\frac{1}{r}\log(-A)=\frac{1}{r}\log\left(\frac{x_0-K}{x_0}\right)<0. \end{align*} Thus for $x_0>K$ the same maximal solution cannot be defined on all of $\mathbb{R}$ as a solution in $U=(0,\infty)$. Finally, since $r>0$, \begin{align*} \lim_{t\to\infty}e^{-rt}=0. \end{align*} Hence \begin{align*} \lim_{t\to\infty}x(t)=\lim_{t\to\infty}\frac{K}{1+Ae^{-rt}}=\frac{K}{1}=K. \end{align*} So positive logistic trajectories move toward the equilibrium $K$ in forward time, while the backward-time lifespan can still depend on the initial value. [/example] The logistic equation illustrates a qualitative lesson: the most important information may be the direction field, equilibria, and long-time behaviour rather than the closed formula. ## Linear Equations and Variation of Constants Nonlinear equations are the main source of difficulty, but linear equations are the reference model. They are the class where superposition, matrix methods, and explicit representation formulas survive. Even when the equation is nonlinear, its local behaviour is often studied by comparing it to a linear equation obtained by linearisation. [definition: Linear Ordinary Differential Equation] Let $I\subset \mathbb{R}$ be an interval, let $A:I\to \mathbb{R}^{n\times n}$ and $b:I\to \mathbb{R}^n$ be continuous. A linear ordinary differential equation is an equation of the form \begin{align*} \dot{x}(t)=A(t)x(t)+b(t) \end{align*} for an unknown differentiable curve $x:J\to \mathbb{R}^n$ with $J\subset I$. [/definition] When $b=0$, solving separately for each initial vector would hide the main structure: the homogeneous equation transports every initial state by the same linear operator. To build a variation-of-constants formula for the inhomogeneous equation, we first need a single matrix-valued solution that records this entire transport mechanism. That object is the fundamental matrix; its columns are independent homogeneous solutions, and its invertibility lets the initial state be recovered and propagated from any time in the interval. [definition: Fundamental Matrix] Let $I\subset \mathbb{R}$ be an interval and let $A:I\to \mathbb{R}^{n\times n}$ be continuous. A fundamental matrix for the homogeneous system $\dot{x}=A(t)x$ is a differentiable map $\Phi:I\to \mathbb{R}^{n\times n}$ such that \begin{align*} \dot{\Phi}(t)=A(t)\Phi(t) \end{align*} and $\Phi(t)$ is invertible for every $t\in I$. [/definition] The fundamental matrix solves the homogeneous problem, but models usually include an inhomogeneous forcing term $b(t)$. The next question is how to combine the homogeneous propagation of initial data with the accumulated effect of forcing; variation of constants gives exactly that representation by transporting each infinitesimal contribution through the homogeneous dynamics. [quotetheorem:8318] The formula is useful even when the integral cannot be evaluated in closed form, because it separates propagation by the homogeneous equation from the accumulated forcing. [example: Integrating Factor as a One-Dimensional Fundamental Matrix] Let $a,b:I\to \mathbb{R}$ be continuous and consider the scalar initial value problem \begin{align*} \dot{x}(t)=a(t)x(t)+b(t), \qquad x(t_0)=x_0. \end{align*} Define \begin{align*} \Phi(t)=\exp\left(\int_{t_0}^{t}a(s)\,d\mathcal{L}^1(s)\right). \end{align*} Since $a$ is continuous, the fundamental theorem of calculus gives \begin{align*} \frac{d}{dt}\int_{t_0}^{t}a(s)\,d\mathcal{L}^1(s)=a(t). \end{align*} The chain rule therefore gives \begin{align*} \dot{\Phi}(t)=\exp\left(\int_{t_0}^{t}a(s)\,d\mathcal{L}^1(s)\right)a(t)=a(t)\Phi(t). \end{align*} Also $\Phi(t)>0$ for every $t\in I$, so $\Phi(t)$ is invertible as a one-dimensional matrix, with inverse $1/\Phi(t)$. Now set \begin{align*} x(t)=\Phi(t)x_0+\int_{t_0}^{t}\frac{\Phi(t)}{\Phi(s)}b(s)\,d\mathcal{L}^1(s). \end{align*} Because $\Phi(t)$ does not depend on the integration variable $s$, this is equivalently \begin{align*} x(t)=\Phi(t)\left(x_0+\int_{t_0}^{t}\frac{b(s)}{\Phi(s)}\,d\mathcal{L}^1(s)\right). \end{align*} At the initial time, \begin{align*} \Phi(t_0)=\exp\left(\int_{t_0}^{t_0}a(s)\,d\mathcal{L}^1(s)\right)=\exp(0)=1. \end{align*} Also the integral from $t_0$ to $t_0$ is $0$, so \begin{align*} x(t_0)=1\cdot(x_0+0)=x_0. \end{align*} Differentiate the product form. The product rule and the fundamental theorem of calculus give \begin{align*} \dot{x}(t)=\dot{\Phi}(t)\left(x_0+\int_{t_0}^{t}\frac{b(s)}{\Phi(s)}\,d\mathcal{L}^1(s)\right)+\Phi(t)\frac{b(t)}{\Phi(t)}. \end{align*} Using $\dot{\Phi}(t)=a(t)\Phi(t)$, this becomes \begin{align*} \dot{x}(t)=a(t)\Phi(t)\left(x_0+\int_{t_0}^{t}\frac{b(s)}{\Phi(s)}\,d\mathcal{L}^1(s)\right)+b(t). \end{align*} The expression multiplying $a(t)$ is exactly $x(t)$, so \begin{align*} \dot{x}(t)=a(t)x(t)+b(t). \end{align*} Thus the one-dimensional fundamental matrix $\Phi$ turns the linear equation into the usual integrating-factor formula, with $\Phi(t)/\Phi(s)$ transporting the forcing term from time $s$ to time $t$. [/example] ## Stability and Qualitative Behaviour Solving an equation exactly is often less important than knowing whether nearby solutions stay nearby. Stability theory asks whether small errors in the initial state remain small, decay, or grow. This is the language that turns ODEs into a tool for understanding equilibria in dynamical systems. [definition: Lyapunov Stability] Let $U\subset \mathbb{R}^n$ be open, let $f:U\to \mathbb{R}^n$ be locally Lipschitz, and let $x^*\in U$ be an equilibrium point of $\dot{x}=f(x)$. The equilibrium $x^*$ is Lyapunov stable if for every $\varepsilon>0$ there exists $\delta>0$ with $B(x^*,\delta)\subset U$ such that every solution with $x(0)\in U$ and $|x(0)-x^*|<\delta$ is defined for all $t\ge 0$ and satisfies \begin{align*} |x(t)-x^*|<\varepsilon \end{align*} for all $t\ge 0$. [/definition] This page uses the autonomous-ODE convention that [Lyapunov stability](/page/Lyapunov%20Stability) includes forward existence of all sufficiently nearby trajectories for all $t\ge 0$. Some texts use a local formulation instead: nearby solutions must remain close for as long as they are defined, and forward existence is then proved separately from compact containment. When this definition is imported into a setting with possible finite escape time, the built-in forward-existence requirement should be kept in view. Stability does not require convergence. A harmonic oscillator stays on a nearby ellipse forever, but it does not settle to the origin. For convergence, the definition must add attraction. [definition: Asymptotic Stability] Let $U\subset \mathbb{R}^n$ be open, let $f:U\to \mathbb{R}^n$ be locally Lipschitz, and let $x^*\in U$ be an equilibrium point of $\dot{x}=f(x)$. The equilibrium $x^*$ is asymptotically stable if it is Lyapunov stable and there exists $\rho>0$ such that every solution with $|x(0)-x^*|<\rho$ is defined for all $t\ge 0$ and satisfies \begin{align*} \lim_{t\to\infty}x(t)=x^*. \end{align*} [/definition] [Asymptotic stability](/page/Asymptotic%20Stability) is a statement about every nearby trajectory, but explicit formulas for those trajectories are rare. A Lyapunov function gives a substitute for solving the equation: it assigns an energy-like height to each nearby state, with the equilibrium as the unique minimum, and asks that this height not increase along the vector field. [definition: Lyapunov Function] Let $U\subset \mathbb{R}^n$ be open, let $f:U\to \mathbb{R}^n$ be continuous, let $x^*\in U$ be an equilibrium point, and let $W\subset U$ be an open neighbourhood of $x^*$. A Lyapunov function for $x^*$ on $W$ is a continuously differentiable function $V:W\to \mathbb{R}$ such that \begin{align*} V(x^*)=0, \end{align*} \begin{align*} V(x)>0 \quad \text{for all } x\in W\setminus\{x^*\}, \end{align*} and \begin{align*} \nabla V(x)\cdot f(x)\le 0 \quad \text{for all } x\in W. \end{align*} [/definition] The definition is useful only if decreasing height can be turned back into a statement about distance from the equilibrium. The basic Lyapunov theorem supplies that conversion: compact sublevel sets inside the region where $V$ is defined give controlled neighbourhoods of $x^*$, and monotonicity of $V$ prevents trajectories that start in a sufficiently small sublevel set from escaping a prescribed neighbourhood or leaving the local domain in finite forward time. [quotetheorem:7603] Quadratic Lyapunov functions are the first examples because they measure squared distance in a weighted Euclidean geometry. [example: Quadratic Lyapunov Function for a Linear System] Let $A\in \mathbb{R}^{n\times n}$ and suppose there exists a symmetric positive definite matrix $P\in \mathbb{R}^{n\times n}$ such that \begin{align*} A^\top P+PA=-I. \end{align*} For the linear system $\dot{x}=Ax$, define $V:\mathbb{R}^n\to \mathbb{R}$ by \begin{align*} V(x)=x^\top Px. \end{align*} Because $P$ is positive definite, $V(0)=0$ and $V(x)=x^\top Px>0$ whenever $x\ne 0$. Let $x(t)$ be a solution of $\dot{x}=Ax$. By the product rule, \begin{align*} \frac{d}{dt}V(x(t))=\dot{x}(t)^\top Px(t)+x(t)^\top P\dot{x}(t). \end{align*} Substituting $\dot{x}(t)=Ax(t)$ gives \begin{align*} \frac{d}{dt}V(x(t))=(Ax(t))^\top Px(t)+x(t)^\top PAx(t). \end{align*} Since $(Ax(t))^\top=x(t)^\top A^\top$, this becomes \begin{align*} \frac{d}{dt}V(x(t))=x(t)^\top A^\top Px(t)+x(t)^\top PAx(t). \end{align*} Factoring out the common left and right vectors gives \begin{align*} \frac{d}{dt}V(x(t))=x(t)^\top(A^\top P+PA)x(t). \end{align*} Using $A^\top P+PA=-I$, \begin{align*} \frac{d}{dt}V(x(t))=x(t)^\top(-I)x(t). \end{align*} Since $Ix(t)=x(t)$, \begin{align*} \frac{d}{dt}V(x(t))=-x(t)^\top x(t)=-|x(t)|^2. \end{align*} Thus $V$ decreases strictly at every nonzero point of a trajectory and is constant only at the origin, so the quadratic form gives a concrete Lyapunov certificate for stability of the equilibrium $0$. [/example] ## Dependence on Initial Data and Parameters A deterministic equation should not only have a unique trajectory; nearby experiments should produce nearby trajectories. This is where estimates such as Gronwall's inequality enter. They turn a differential or integral comparison into a bound that grows at a controlled exponential rate. [quotetheorem:872] Gronwall converts an integral inequality into an explicit bound, which is exactly what is missing when comparing two unknown solution curves. Existence and uniqueness say that a starting point determines a trajectory, but they do not yet justify using that trajectory as a stable prediction: a measured initial state is never exact. The next theorem supplies the local estimate that underlies continuous dependence. It is stated on a fixed interval where the relevant solutions are already known to exist and remain in a compact region with a common Lipschitz constant. [quotetheorem:8319] The estimate is useful because it turns uniqueness into quantitative robustness. On a fixed compact time interval where both solutions stay inside the same region, the common Lipschitz constant controls how fast an initial error can grow. Thus a small measurement error at the initial time remains bounded by a computable expression, rather than merely producing a different solution whose relation to the original one is unknown. The hypotheses are deliberately local. The bound depends on having the same Lipschitz control throughout the region visited by both trajectories, and it only applies on an interval where those trajectories are already known to exist. If a solution leaves the chosen compact set, approaches the boundary of the state space, or is followed beyond its established lifespan, the comparison has to be restarted with new data or may fail altogether. Within those limits, the theorem is the quantitative bridge to continuous dependence and explains why flow maps can be treated as stable objects under perturbation of the initial state. Before parameters enter, there is one more standard qualitative question: how does a one-dimensional autonomous equation move along the line? In dimension one, the sign of the vector field already determines the direction of motion, so equilibria divide the state space into intervals of monotone flow. [definition: Phase Line] Let $U\subset \mathbb{R}$ be open and let $f:U\to \mathbb{R}$ be continuous. The phase line of the autonomous equation $\dot{x}=f(x)$ is the real line, or the state interval $U$, marked with the equilibrium points of $f$ and the sign of $f$ on each component of $U\setminus\{x\in U:f(x)=0\}$. [/definition] The phase line is a compact way to record attraction and repulsion without solving the equation, but the informal arrow picture needs hypotheses. Local uniqueness prevents a trajectory from crossing an equilibrium, and isolation ensures the neighbouring intervals really have a fixed sign. [quotetheorem:8320] The theorem gives precise content to the arrows drawn on a phase line. A positive-to-negative sign change gives local attraction for forward-global trajectories that stay in the isolating interval, while a negative-to-positive sign change gives local repulsion. The remaining cases, such as a sign that does not change, record one-sided attraction or semistability rather than full attraction from both sides. This sign test depends on the order structure of the real line, so it has no direct analogue in higher dimensions. The next definition is needed to replace the unavailable left-versus-right sign picture by a local linear model: near an equilibrium in $\mathbb{R}^n$, the Jacobian gives the first-order system that controls the initial qualitative diagnostic. [definition: Linearisation at an Equilibrium] Let $U\subset \mathbb{R}^n$ be open, let $f\in C^1(U;\mathbb{R}^n)$, and let $x^*\in U$ be an equilibrium point of $\dot{x}=f(x)$. The linearisation of $\dot{x}=f(x)$ at $x^*$ is the linear autonomous system \begin{align*} \dot{y}(t)=Jf_{x^*}y(t), \end{align*} where $Jf_{x^*}\in \mathbb{R}^{n\times n}$ is the Jacobian matrix of $f$ at $x^*$. [/definition] This definition extracts the best linear predictor near a fixed equilibrium, but the definition alone does not tell us how to read that predictor. To use linearisation as a diagnostic, we first need the exact stability test for the frozen linear system: eigenvalues with negative real parts force every small perturbation to decay, while an eigenvalue with positive real part produces directions in which perturbations grow. The next theorem is the baseline result that makes later nonlinear stability tests credible rather than merely suggestive. [quotetheorem:8321] The eigenvalue criterion is deliberately linear: it describes the frozen model exactly, and it guides nonlinear intuition only when additional hypotheses justify comparing the nonlinear equation with its linearisation. In applications, the vector field often changes with a mass, damping coefficient, forcing strength, or geometric constant, and then the real question is not only how one trajectory behaves, but how the whole family changes as the model parameter moves. The next definition keeps that parameter separate from time and state so that continuous dependence, perturbation, and bifurcation questions have a precise domain. [definition: Parameter-Dependent Ordinary Differential Equation] Let $I\subset \mathbb{R}$ be an interval, let $U\subset \mathbb{R}^n$ and $P\subset \mathbb{R}^m$ be open, and let $F:I\times U\times P\to \mathbb{R}^n$. A parameter-dependent ordinary differential equation is an equation of the form \begin{align*} \dot{x}(t)=F(t,x(t),p), \end{align*} where $p\in P$ is fixed and $x:J\to U$ is the unknown curve. [/definition] This formulation connects ODEs to perturbation theory and bifurcation theory. Before asking whether qualitative behaviour changes abruptly at a special parameter value, we need the ordinary regime: away from such a breakdown, small changes in the initial state and the parameter should produce small changes in the solution on a fixed compact time interval. [quotetheorem:8322] The theorem says that parameters do not automatically make a model fragile: on a time interval where the reference trajectory stays away from the boundary of the state space, nearby models track it uniformly. Bifurcation theory begins where this regular picture stops being enough, usually because equilibria collide, eigenvalues cross the imaginary axis, or the time interval of interest ceases to be compact. ## Beyond and Connected Topics Ordinary differential equations are the local language of motion. In dynamical systems, the emphasis shifts from solving individual equations to studying the geometry of orbits, invariant sets, recurrence, and bifurcation. The flow map introduced above is the bridge between the analytic initial value problem and the global phase-space picture. Linear equations lead naturally to spectral methods. The behaviour of $\dot{x}=Ax$ is controlled by the eigenvalues of $A$, and the same idea reappears in linearisation near equilibria for nonlinear systems. This is the entry point to stable and unstable manifolds, normal forms, and qualitative theory. Ordinary differential equations also sit inside geometry. A vector field on a manifold generates local flows, and those flows are used to define Lie derivatives, exponential maps for Lie groups, and the geometric meaning of transport. The Androma notes on [Cambridge III Differential Geometry](/page/Cambridge%20III%20Differential%20Geometry) and [Differential Forms and de Rham Cohomology](/page/Differential%20Forms%20and%20de%20Rham%20Cohomology) develop this viewpoint. For course-level continuation, [Cambridge IA Differential Equations](/page/Cambridge%20IA%20Differential%20Equations) is the natural internal reference for explicit solution methods and classical examples, while [Cambridge IB Analysis and Topology](/page/Cambridge%20IB%20Analysis%20and%20Topology) supplies the compactness and metric-space background used in existence and continuation arguments. ## References Androma, [Cambridge IA Differential Equations](/page/Cambridge%20IA%20Differential%20Equations). Androma, [Cambridge IB Analysis and Topology](/page/Cambridge%20IB%20Analysis%20and%20Topology). Androma, [Cambridge III Differential Geometry](/page/Cambridge%20III%20Differential%20Geometry). Androma, [Differential Forms and de Rham Cohomology](/page/Differential%20Forms%20and%20de%20Rham%20Cohomology). Vladimir I. Arnold, *Ordinary Differential Equations* (1992). Philip Hartman, *Ordinary Differential Equations* (1964). Morris W. Hirsch, Stephen Smale, and Robert L. Devaney, *Differential Equations, Dynamical Systems, and an Introduction to Chaos* (2013).