An [ordinary differential equation](/page/Ordinary%20Differential%20Equation) describes a rule for instantaneous motion, but a rule alone does not choose a path. The equation $\dot{x}=f(x)$ says that a moving point should have velocity $f(x)$ when it is at position $x$; it does not say where the point starts. An initial value problem adds that missing datum. The central question is whether the velocity rule and the starting point determine a solution, whether that solution is unique, and how long it survives. The first warning is that a plausible-looking velocity field may fail to determine the future. Consider a particle on the real line whose speed is proportional to the square root of its distance from the origin. If it starts at the origin, does it stay there, or can it begin moving later? The differential equation allows both behaviours, so the initial condition alone is not enough unless the vector field has the right regularity. [example: Non-Unique Departure from Rest] Let $f:\mathbb{R}\to\mathbb{R}$ be $f(x)=2\sqrt{|x|}$, and impose $x(0)=0$. The constant curve $x(t)=0$ is a solution because $\dot{x}(t)=0$ for every $t$, while \begin{align*} f(x(t))=f(0)=2\sqrt{|0|}=0. \end{align*} Now fix $a\ge 0$ and define $x_a(t)=0$ for $t\le a$, while $x_a(t)=(t-a)^2$ for $t>a$. Since $0\le a$, we have $x_a(0)=0$. The curve is continuous at $t=a$ because the left value is $0$ and \begin{align*} \lim_{t\downarrow a}(t-a)^2=0. \end{align*} It is also differentiable at $t=a$: for $h<0$, \begin{align*} \frac{x_a(a+h)-x_a(a)}{h}=\frac{0-0}{h}=0, \end{align*} and for $h>0$, \begin{align*} \frac{x_a(a+h)-x_a(a)}{h}=\frac{h^2-0}{h}=h, \end{align*} so both one-sided difference quotients tend to $0$. For $t<a$, $\dot{x}_a(t)=0$ and $f(x_a(t))=f(0)=0$. At $t=a$, the calculation above gives $\dot{x}_a(a)=0$, and again $f(x_a(a))=0$. For $t>a$, differentiating $x_a(t)=(t-a)^2$ gives \begin{align*} \dot{x}_a(t)=2(t-a). \end{align*} Since $t>a$, we have $t-a>0$, so \begin{align*} f(x_a(t))=2\sqrt{|(t-a)^2|}=2\sqrt{(t-a)^2}=2(t-a). \end{align*} Thus every $x_a$ solves $\dot{x}=2\sqrt{|x|}$ on $\mathbb{R}$ with $x_a(0)=0$, and the parameter $a\ge 0$ gives infinitely many distinct solutions. The vector field itself is continuous, but it is not Lipschitz on any interval containing $0$. Indeed, for $h>0$, \begin{align*} \frac{|f(h)-f(0)|}{|h-0|}=\frac{|2\sqrt{h}-0|}{h}=\frac{2}{\sqrt{h}}, \end{align*} and this quantity becomes arbitrarily large as $h\downarrow 0$. The initial condition is therefore not the source of non-uniqueness; the failure is the missing Lipschitz control at the origin. [/example] This example sets the agenda for the subject. Initial value problems are not merely a notation for differential equations with starting data. They are the place where analysis decides whether deterministic laws really determine motion, how regularity assumptions enter, and what kinds of singular behaviour can appear even in finite-dimensional systems. ## Definition To define the problem precisely, we must specify the time interval, the state space, the vector field, the initial time, and the initial state. The state space is usually an open subset of Euclidean space, because solutions should be allowed to move slightly in every coordinate direction while remaining inside the domain where the vector field is defined. [definition: Initial Value Problem] Let $E \subset \mathbb{R}^n$ be open, let $I \subset \mathbb{R}$ be an interval, let $t_0 \in I$, let $x_0 \in E$, and let $f: I \times E \to \mathbb{R}^n$ be a vector field. The initial value problem with vector field $f$ and initial condition $(t_0,x_0)$ is the problem of finding an interval $J \subset I$ with $t_0 \in J$ and a continuous map $x: J \to E$ such that $x$ is differentiable on the interior $J^\circ$, satisfies $\dot{x}(t)=f(t,x(t))$ for all $t \in J^\circ$, and satisfies $x(t_0)=x_0$. [/definition] This page uses the non-autonomous formulation as the parent notion: the velocity may depend on both time and state. The autonomous equation $\dot{x}=V(x)$ is recovered by taking $f(t,x)=V(x)$, so the geometric phase-space picture is a special case rather than a competing definition. ## Solution Intervals and Basic Forms The interval $I$ is part of the data when asking for a solution on a prescribed time domain. In many applications, however, the interval is not known in advance; the question is to find an interval around $t_0$ on which a solution exists. To make that question meaningful, we need a separate definition of a solution on whatever interval is under consideration. [definition: Solution of an Initial Value Problem] Let $E \subset \mathbb{R}^n$ be open, let $I \subset \mathbb{R}$ be an interval, let $t_0 \in I$, let $x_0 \in E$, and let $f: I \times E \to \mathbb{R}^n$. A solution of the initial value problem on an interval $J \subset I$ with $t_0 \in J$ is a continuous map $x: J \to E$ such that $x$ is differentiable on $J^\circ$, satisfies $\dot{x}(t)=f(t,x(t))$ for all $t \in J^\circ$, and satisfies $x(t_0)=x_0$. [/definition] When $f$ is continuous, any solution has continuous derivative on the interior of its solution interval, because $t \mapsto f(t,x(t))$ is continuous there. Thus solutions are automatically $C^1$ away from endpoints in the standard existence theory. A basic organizing question is whether the velocity rule changes with the clock itself or only with the current state. If the same state always determines the same velocity, then the initial time can be normalized to $0$ and the problem is governed by a vector field on state space alone. In that situation the unknown interval is not just bookkeeping: it records how long the solution curve can be followed while it remains inside the state space $E$. The formal problem therefore asks for both a curve and a time interval containing the normalized initial time, with the curve's tangent at each interior time prescribed by the vector field at its current position. [definition: Autonomous Initial Value Problem] Let $E \subset \mathbb{R}^n$ be open, let $V: E \to \mathbb{R}^n$ be a vector field, and let $x_0 \in E$. The autonomous initial value problem with initial state $x_0$ is the problem of finding an interval $I \subset \mathbb{R}$ with $0 \in I$ and a continuous map $x: I \to E$ such that $x$ is differentiable on $I^\circ$, satisfies $\dot{x}(t)=V(x(t))$ for all $t \in I^\circ$, and satisfies $x(0)=x_0$. [/definition] The autonomous case matches the geometric picture of a vector field placing an arrow at each point of the state space. A solution is a curve whose tangent vector agrees with the arrow it meets. Time-dependent forcing, seasonal effects, and control inputs break this picture because the arrow at a fixed state may vary with time, so we need a separate formulation that keeps time as an explicit input. [definition: Non-Autonomous Initial Value Problem] Let $E \subset \mathbb{R}^n$ be open, let $I \subset \mathbb{R}$ be an interval, let $f: I \times E \to \mathbb{R}^n$ be a vector field, and let $(t_0,x_0) \in I \times E$. The non-autonomous initial value problem is the problem $\dot{x}(t)=f(t,x(t))$ on $J^\circ$ and $x(t_0)=x_0$ for an unknown continuous map $x: J \to E$ that is differentiable on $J^\circ$, where $J \subset I$ is an interval containing $t_0$. [/definition] A non-autonomous problem can be converted into an autonomous problem in one higher dimension by adjoining time as a state variable. This device is useful, but it does not erase the distinction: the original state variable and the clock play different roles in applications. Once solutions exist, we also need language for the geometric curve seen in state space. [definition: Trajectory] Let $E \subset \mathbb{R}^n$ be open, let $I \subset \mathbb{R}$ be an interval, and let $x: I \to E$ be a solution of an initial value problem. The trajectory of $x$ is the image set $x(I)=\{x(t):t\in I\}\subset E$. [/definition] The trajectory forgets the speed at which the curve is traversed. For autonomous equations, this is often natural: the phase portrait is a picture of possible paths in state space, while the parameter $t$ records how the motion moves along those paths. The simplest solvable example already shows the three questions of existence, uniqueness, and dependence on initial state. [example: Exponential Growth as an Initial Value Problem] For $\lambda \in \mathbb{R}$ and $x_0 \in \mathbb{R}$, consider $\dot{x}(t)=\lambda x(t)$ with $x(0)=x_0$. We first verify the proposed solution $x(t)=x_0e^{\lambda t}$: \begin{align*} \dot{x}(t)=x_0\lambda e^{\lambda t}=\lambda x_0e^{\lambda t}=\lambda x(t). \end{align*} At the initial time, \begin{align*} x(0)=x_0e^{0}=x_0. \end{align*} Thus $x(t)=x_0e^{\lambda t}$ is a solution on all of $\mathbb{R}$. To see uniqueness, let $u:\mathbb{R}\to\mathbb{R}$ be any differentiable solution with $u(0)=x_0$, and define $w(t)=e^{-\lambda t}u(t)$. By the product rule, \begin{align*} w'(t)=(-\lambda e^{-\lambda t})u(t)+e^{-\lambda t}u'(t). \end{align*} Since $u'(t)=\lambda u(t)$, \begin{align*} w'(t)=-\lambda e^{-\lambda t}u(t)+e^{-\lambda t}\lambda u(t)=0. \end{align*} Hence $w$ is constant on $\mathbb{R}$, and evaluating at $t=0$ gives \begin{align*} w(t)=w(0)=e^0u(0)=x_0. \end{align*} Multiplying by $e^{\lambda t}$ gives \begin{align*} u(t)=x_0e^{\lambda t}. \end{align*} Therefore the solution is unique. If $x_0\ne 0$, then $|x(t)|=|x_0|e^{\lambda t}$. For $\lambda>0$, this tends to $\infty$ as $t\to\infty$, so nonzero initial states move away from $0$ exponentially forward in time. For $\lambda<0$, this tends to $0$ as $t\to\infty$, so nonzero initial states decay to $0$. If $x_0=0$, then \begin{align*} x(t)=0\cdot e^{\lambda t}=0 \end{align*} for every $t$, so the solution remains at the equilibrium $0$ for every value of $\lambda$. The same formula measures dependence on the initial state. If $y(0)=y_0$, then uniqueness gives $y(t)=y_0e^{\lambda t}$, and for each $t\in\mathbb{R}$, \begin{align*} x(t)-y(t)=x_0e^{\lambda t}-y_0e^{\lambda t}=(x_0-y_0)e^{\lambda t}. \end{align*} Since $e^{\lambda t}>0$, \begin{align*} |x(t)-y(t)|=e^{\lambda t}|x_0-y_0|. \end{align*} Thus initial perturbations are amplified forward in time when $\lambda>0$ and damped forward in time when $\lambda<0$. [/example] The formula above is familiar, but it also illustrates the structure of the theory. Existence gives a solution, uniqueness says there is no other solution with the same initial value, and dependence on $x_0$ says the formula varies predictably when the initial state is changed. ## Existence and Uniqueness ### Local Existence The fundamental analytical problem is local: given a point $(t_0,x_0)$ where the vector field is defined, can we solve the equation at least for a short time? Global existence is often too much to ask, because solutions may leave every bounded region in finite time. Local existence is the first threshold. A continuous vector field gives enough compactness to produce at least one solution locally. This is a powerful statement because it asks for no Lipschitz control, but the opening example warns that uniqueness may fail. In the local setting, Peano's existence principle says that if $f$ is continuous near $(0,X_0)$, then, after possibly shrinking the time interval, there is at least one curve $X$ satisfying the initial condition $X(0)=X_0$ and the differential equation $\dot X(t)=f(t,X(t))$. The principle is deliberately only an existence result: it produces a solution on a small interval, but it does not compare that solution with any other candidate curve. Peano's theorem says that continuity prevents immediate analytical breakdown. It does not say that the solution is determined by the initial condition. To obtain determinism, the vector field must not change too sharply with respect to the state variable, so the next notion measures controlled variation in that variable. [definition: Local Lipschitz Continuity in the State Variable] Let $E \subset \mathbb{R}^n$ be open and let $I \subset \mathbb{R}$ be an interval. A map $f: I \times E \to \mathbb{R}^n$ is locally Lipschitz in the state variable if for every compact set $K \subset I \times E$ there exists $L_K >0$ such that $|f(t,x)-f(t,y)|\le L_K|x-y|$ for all $t \in I$ and all $x,y \in E$ such that $(t,x),(t,y) \in K$. [/definition] This condition allows the vector field to vary with time and position, but on each compact rectangle it gives a uniform bound on how much changing the state can change the velocity. That is exactly the estimate needed to compare two candidate solutions. The resulting theorem is the standard determinism theorem for ordinary differential equations. [quotetheorem:8313] The last sentence is a local uniqueness statement. It says that two solutions with the same initial value cannot split apart while both are defined. This is the mathematical expression of determinism for ordinary differential equations. The point of the Lipschitz hypothesis is not just technical smoothness; it gives a quantitative comparison estimate, so differences between two solutions can be controlled before they have any chance to grow from zero. ### Failure of Uniqueness The square-root example shows that continuity alone is not enough. The next example shows that the Lipschitz condition is not a technical ornament; it is the boundary between uniqueness and branching. [example: Why Lipschitz Regularity Matters] Let $f(x)=2\sqrt{|x|}$. To test Lipschitz control at the origin, take $h>0$. Then $|h|=h$, $f(0)=2\sqrt{|0|}=0$, and $f(h)=2\sqrt{|h|}=2\sqrt{h}$. Hence \begin{align*} \frac{|f(h)-f(0)|}{|h-0|}=\frac{|2\sqrt{h}-0|}{h}. \end{align*} Since $\sqrt{h}>0$, we have $|2\sqrt{h}|=2\sqrt{h}$, so \begin{align*} \frac{|2\sqrt{h}-0|}{h}=\frac{2\sqrt{h}}{h}. \end{align*} Because $h=(\sqrt{h})^2$ and $\sqrt{h}>0$, \begin{align*} \frac{2\sqrt{h}}{h}=\frac{2\sqrt{h}}{(\sqrt{h})^2}=\frac{2}{\sqrt{h}}. \end{align*} Now let $J$ be any interval containing $0$ and suppose, for contradiction, that $f$ is Lipschitz on $J$ with constant $L$. For every sufficiently small $h>0$ with $h\in J$, the Lipschitz inequality gives \begin{align*} \frac{|f(h)-f(0)|}{|h-0|}\le L. \end{align*} The computation above gives $2/\sqrt{h}\le L$. Choosing $0<h<4/L^2$ when $L>0$ gives $2/\sqrt{h}>L$, a contradiction; if $L=0$, any $h>0$ already gives $2/\sqrt{h}>0=L$. Thus $f$ is not Lipschitz on any interval containing $0$. This is exactly the regularity failure used by the delayed solutions $x_a$: the equation permits the solution to remain at $0$ for an arbitrary waiting time and then depart along $x_a(t)=(t-a)^2$, so continuity of the vector field alone does not force a unique departure time. [/example] Uniqueness can also be obtained under hypotheses weaker or different from local Lipschitz continuity, but Picard-Lindelof is the standard theorem because it gives existence, uniqueness, and a method of construction in one argument. ## Integral Formulation and Fixed Points Differentiating a proposed solution is often less convenient than integrating the equation. The integral version includes the initial condition automatically and moves the problem into a function space where fixed-point theorems apply. A solution of $\dot{x}=f(t,x)$ should accumulate velocity over time. This leads to the integral equation below, which is equivalent to the differential equation under the usual continuity assumptions. In the theorem card, $d\mathcal{L}^1(s)$ denotes integration with respect to one-dimensional [Lebesgue measure](/page/Lebesgue%20Measure); for the ordinary ODEs considered here, it can be read as the usual $ds$ notation. [quotetheorem:8311] The integral formulation turns the initial value problem into the problem of finding a fixed point of a map on a space of curves. The map takes a candidate curve, reads off its velocity along that curve, and integrates that velocity from the initial state. [definition: Picard Operator] Let $E \subset \mathbb{R}^n$ be open, let $I \subset \mathbb{R}$ be an interval, let $f: I \times E \to \mathbb{R}^n$ be continuous, and let $(t_0,x_0) \in I \times E$. For a set $X \subset C(I;E)$ of continuous maps $u: I \to E$, the Picard operator associated to the initial value problem is the map $T: X \to C(I;\mathbb{R}^n)$ defined by \begin{align*} (Tu)(t)=x_0+\int_{t_0}^{t} f(s,u(s))\, ds. \end{align*} [/definition] When $T(X)\subset X$, the Picard operator is regarded as a self-map $T:X\to X$, and a fixed point of $T$ is exactly a curve satisfying the integral equation. The proof of Picard-Lindelof chooses a small interval and a closed ball of curves on which $T$ maps the ball into itself and contracts distances. This raises the constructive question of what happens if we keep applying $T$ to an initial guess, which motivates the iteration scheme. [definition: Picard Iteration] Let $E \subset \mathbb{R}^n$ be open, let $I \subset \mathbb{R}$ be an interval, let $f: I \times E \to \mathbb{R}^n$ be continuous, and let $(t_0,x_0) \in I \times E$. Picard iteration is a sequence of continuous maps $u_k:I\to E$ with $u_0:I\to E$ given by $u_0(t)=x_0$, and with each later iterate defined whenever the formula below takes values in $E$: \begin{align*} u_{k+1}(t)=x_0+\int_{t_0}^{t} f(s,u_k(s))\, ds. \end{align*} [/definition] The iteration creates successive approximations to the solution. Under the contraction estimates in Picard-Lindelof, these approximations converge uniformly on a sufficiently small interval to the unique fixed point. [example: Picard Iteration for Exponential Growth] For the initial value problem $\dot{x}(t)=\lambda x(t)$ with $x(0)=a$, Picard iteration starts from $u_0(t)=a$ and uses \begin{align*} u_{k+1}(t)=a+\int_0^t \lambda u_k(s)\,ds. \end{align*} For the first iterate, \begin{align*} u_1(t)=a+\int_0^t \lambda a\,ds=a+\lambda a t=a(1+\lambda t). \end{align*} For the second iterate, \begin{align*} u_2(t)=a+\int_0^t \lambda a(1+\lambda s)\,ds. \end{align*} Expanding the integrand gives \begin{align*} \lambda a(1+\lambda s)=\lambda a+\lambda^2as. \end{align*} Therefore \begin{align*} u_2(t)=a+\lambda at+\frac{\lambda^2at^2}{2}=a\left(1+\lambda t+\frac{\lambda^2t^2}{2}\right). \end{align*} The pattern is that \begin{align*} u_k(t)=a\sum_{j=0}^k\frac{(\lambda t)^j}{j!}. \end{align*} Indeed, if this formula holds for $u_k$, then \begin{align*} u_{k+1}(t)=a+\int_0^t \lambda a\sum_{j=0}^k\frac{(\lambda s)^j}{j!}\,ds. \end{align*} Using linearity of the integral, \begin{align*} u_{k+1}(t)=a+a\sum_{j=0}^k\frac{\lambda^{j+1}}{j!}\int_0^t s^j\,ds. \end{align*} Since $\int_0^t s^j\,ds=t^{j+1}/(j+1)$, \begin{align*} u_{k+1}(t)=a+a\sum_{j=0}^k\frac{\lambda^{j+1}t^{j+1}}{(j+1)!}. \end{align*} Reindexing with $m=j+1$ gives \begin{align*} u_{k+1}(t)=a\sum_{m=0}^{k+1}\frac{(\lambda t)^m}{m!}. \end{align*} Thus the Picard iterates are exactly the Taylor partial sums of $ae^{\lambda t}$, so the fixed-point construction recovers the exponential solution. [/example] The fixed-point viewpoint explains why local existence intervals often shrink when the vector field is large or the Lipschitz constant is large. The curve must remain inside a region where the estimates are valid, and the contraction factor depends on the length of the interval. ## Maximal Solutions and Blow-Up ### Maximal Continuation Local solutions are not the end of the story. Once a solution exists near $t_0$, we can ask whether it can be extended. If it can, we extend it; if it cannot, we want to know whether it hits the boundary of the domain, becomes unbounded, or reaches the edge of the prescribed time interval. The correct object is not merely a solution on some convenient interval, but a solution that has been continued as far as possible. [definition: Maximal Solution] Let $E \subset \mathbb{R}^n$ be open, let $I \subset \mathbb{R}$ be an interval, let $f: I \times E \to \mathbb{R}^n$, and let $(t_0,x_0) \in I \times E$. A solution $x: J \to E$ of the initial value problem is maximal if $J \subset I$ is an interval containing $t_0$, $J$ is open in the relative topology of $I$, and there is no solution $y: K \to E$ on an interval $K \subset I$ with $J \subsetneq K$ such that $y(t)=x(t)$ for all $t \in J$. [/definition] Maximality is a statement about non-extendability, not about size. Because the maximal domain is taken relative to the prescribed time interval $I$, it may look open, half-open, or closed when viewed inside $\mathbb{R}$; for instance, an interval beginning at the left endpoint of $I$ can still be open in the relative topology of $I$. When local uniqueness holds, all compatible local pieces should glue together; the next theorem confirms that maximality becomes a unique canonical object. [quotetheorem:8375] After uniqueness has made the maximal solution canonical, the remaining question is what non-extendability can mean in this finite-dimensional setting. For an ordinary differential equation on an [open set](/page/Open%20Set) $E \subset \mathbb{R}^n$, the obstruction is geometric rather than mysterious: near a finite endpoint, a maximal solution must either run into the boundary of the domain where the vector field is defined or escape all bounded regions of space. [remark: Compact Escape for Finite-Dimensional ODEs] For a maximal solution $x:J\to E$ of $x'=f(t,x)$ with $E\subset\mathbb{R}^n$ open, failure to continue to a forward endpoint $b<\sup I$ means that the curve cannot remain in a compact part of the domain of the vector field as $t\uparrow b$. If $E=\mathbb{R}^n$ and $b$ is finite, this reduces to unbounded growth of the solution: $|x(t)|\to\infty$ as $t\uparrow b$. On a proper open subset $E$, a solution can also fail by approaching the boundary of $E$. [/remark] ### Blow-Up and Global Existence Finite-time breakdown can look mysterious if we only stare at the differential equation, because the formula for the vector field may remain smooth everywhere. The useful diagnostic is whether the solution itself escapes every bounded region before the endpoint. [definition: Finite-Time Blow-Up] Let $x: (a,b) \to \mathbb{R}^n$ be a solution of an initial value problem with $b < \infty$. The solution blows up at time $b$ if $\lim_{t \uparrow b}|x(t)|=\infty$. [/definition] Blow-up does not mean the differential equation stops making sense algebraically. It means the solution leaves every bounded region in finite time, so there is no value in $\mathbb{R}^n$ that can be assigned at the endpoint to continue the curve. [example: Finite-Time Blow-Up for a Smooth Vector Field] Consider the scalar initial value problem $\dot{x}(t)=x(t)^2$ with $x(0)=1$. Define \begin{align*} x(t)=\frac{1}{1-t} \end{align*} for $t<1$. Then \begin{align*} x(0)=\frac{1}{1-0}=1. \end{align*} Writing $x(t)=(1-t)^{-1}$, the chain rule gives \begin{align*} \dot{x}(t)=-(1-t)^{-2}\cdot(-1)=(1-t)^{-2}. \end{align*} On the other hand, \begin{align*} x(t)^2=\left(\frac{1}{1-t}\right)^2=\frac{1}{(1-t)^2}=(1-t)^{-2}. \end{align*} Thus $\dot{x}(t)=x(t)^2$ for every $t<1$. The formula also shows why the solution cannot be continued through $t=1$ as a real-valued solution. For $0<t<1$, \begin{align*} |x(t)|=\left|\frac{1}{1-t}\right|=\frac{1}{1-t}, \end{align*} because $1-t>0$. If $M>0$, choosing $t>1-1/M$ gives $1-t<1/M$, hence \begin{align*} |x(t)|=\frac{1}{1-t}>M. \end{align*} Therefore $\lim_{t\uparrow 1}|x(t)|=\infty$. The maximal interval containing the initial time is $(-\infty,1)$, so even the smooth vector field $f(x)=x^2$ can produce finite-time blow-up rather than a global solution. [/example] Global existence requires additional structure, such as linear growth bounds, conserved quantities, compact state space, or dissipative estimates. These are not merely technical hypotheses; they are what prevent trajectories from escaping in finite time. A common sufficient condition is a growth estimate that bounds the speed by an affine function of the distance from the origin. [quotetheorem:8315] This theorem explains why linear systems behave better than superlinear equations such as $\dot{x}=x^2$. Linear growth keeps solutions from racing to infinity in finite time. ## Dependence on Initial Data ### Flow Maps Solving one initial value problem is rarely enough. In modelling, [numerical analysis](/page/Numerical%20Analysis), and dynamics, we need to know what happens when the starting point or starting time is perturbed. A theory without dependence on data would be unstable: nearby experiments could have unrelated outcomes. For autonomous equations with uniqueness, each initial state determines a solution at later times, but not always for the same amount of time. The natural object is therefore a local flow whose domain records exactly which pairs of time and starting point are valid. When every starting point survives on a common interval, this local flow restricts to the simpler product-domain version used in many elementary examples. [definition: Flow Map] Let $E \subset \mathbb{R}^n$ be open and let $V: E \to \mathbb{R}^n$ be a vector field such that the autonomous initial value problem has a unique maximal solution $x_{x_0}:J_{x_0}\to E$ through each $x_0\in E$. The local flow domain is \begin{align*} \mathcal{D}=\{(t,x_0)\in \mathbb{R}\times E:t\in J_{x_0}\}. \end{align*} The flow map is the map $\varphi:\mathcal{D}\to E$ defined by $\varphi(t,x_0)=x_{x_0}(t)$. [/definition] For an autonomous system, the flow has a composition law: moving for time $s$ and then for time $t$ should match moving for time $s+t$, as long as all expressions are defined. This law is not extra structure imposed on the system; it is forced by uniqueness of the initial value problem. [quotetheorem:8376] The flow property says that autonomous initial value problems generate a local action of the additive time parameter. This is the reason autonomous ODEs form the foundation of finite-dimensional dynamics. For applications, the next question is whether this action changes continuously when the starting point changes. [quotetheorem:8377] The compactness of the time interval matters because local estimates must be made on a region where the relevant solutions remain controlled. Dependence can fail to be uniform near a blow-up time. ### Linearisation Along a Solution A sharper statement is available when the vector field has continuous first derivatives. Then the flow is differentiable with respect to the initial condition, and its derivative satisfies a linear initial value problem. This linear equation records how infinitesimal changes in the starting point are transported along the original trajectory. [definition: Variational Equation] Let $E \subset \mathbb{R}^n$ be open, let $V \in C^1(E;\mathbb{R}^n)$, and let $x: I \to E$ be a solution of $\dot{x}=V(x)$ with $0 \in I$. The variational equation along $x$ is the matrix-valued initial value problem for an unknown continuous map $Y: I \to \mathbb{R}^{n \times n}$ that is differentiable on $I^\circ$, satisfies $\dot{Y}(t)=JV_{x(t)}Y(t)$ for all $t \in I^\circ$, and satisfies $Y(0)=I_n$, where $JV_{x(t)} \in \mathbb{R}^{n \times n}$ is the Jacobian matrix of $V$ at $x(t)$. [/definition] The variational equation describes how an infinitesimal perturbation of the initial state evolves along a trajectory. It is the linearized dynamics seen by nearby solutions. This motivates asking whether the linearized equation is genuinely the derivative of the nonlinear flow, and the differentiable-dependence theorem answers that question. [quotetheorem:8378] Differentiable dependence is the entry point to stability theory, invariant manifolds, and bifurcation theory. It turns qualitative questions about nearby trajectories into linear algebra along a known solution. ## Linear Initial Value Problems ### Homogeneous Systems Linear systems are the model case where the theory becomes especially explicit. They also provide the local approximation to nonlinear systems through the variational equation. The homogeneous linear system evolves vectors by a time-dependent family of linear maps. The object that records all solutions at once is a fundamental matrix, but first we isolate the linear initial value problem itself. [definition: Homogeneous Linear Initial Value Problem] Let $I \subset \mathbb{R}$ be an interval, let $A: I \to \mathbb{R}^{n \times n}$ be continuous, let $t_0 \in I$, and let $x_0 \in \mathbb{R}^n$. The homogeneous linear initial value problem is the problem of finding a continuous map $x: I \to \mathbb{R}^n$ such that $x$ is differentiable on $I^\circ$, satisfies $\dot{x}(t)=A(t)x(t)$ for all $t \in I^\circ$, and satisfies $x(t_0)=x_0$. [/definition] Linearity means that the solution with initial state $x_0$ depends linearly on $x_0$. Rather than solve separately for every initial vector, we need a single matrix solution whose columns are the solutions starting from the standard basis vectors. In the initial value problem setting, it is useful to normalize this matrix at the initial time; many texts call any invertible matrix solution a fundamental matrix, while the normalized object below is the principal fundamental matrix based at $t_0$. [definition: Principal Fundamental Matrix] Let $I \subset \mathbb{R}$ be an interval, let $A: I \to \mathbb{R}^{n \times n}$ be continuous, and let $t_0 \in I$. The principal fundamental matrix for the system $\dot{x}=A(t)x$ based at $t_0$ is a continuous map $\Phi: I \to \mathbb{R}^{n \times n}$ such that $\Phi$ is differentiable on $I^\circ$, satisfies $\dot{\Phi}(t)=A(t)\Phi(t)$ for all $t \in I^\circ$, and satisfies $\Phi(t_0)=I_n$. [/definition] Once $\Phi$ is known, the solution with initial value $x_0$ is $x(t)=\Phi(t)x_0$. This reduces all homogeneous linear initial value problems for the same coefficient matrix to one matrix-valued problem. The existence theorem confirms that this packaging loses no information. [quotetheorem:8379] The important new point is that the principal fundamental matrix is not merely a convenient package of solutions. Its invertibility means that every state at one time corresponds to exactly one initial state at the base time, so the linear flow can be run backward as well as forward wherever the coefficients are continuous. This is the system-level analogue of the scalar fact that a nonzero solution of a homogeneous linear equation never crosses zero, but in systems it is encoded by a whole matrix rather than by a single integrating factor. The theorem does not solve forced equations by itself; instead, it supplies the reversible homogeneous evolution that the inhomogeneous theory will use as a moving coordinate frame. ### Inhomogeneous Systems Many linear models include forcing: an external input acts on the system independently of the current state. This produces an affine differential equation rather than a homogeneous linear one. [definition: Inhomogeneous Linear Initial Value Problem] Let $I \subset \mathbb{R}$ be an interval, let $A: I \to \mathbb{R}^{n \times n}$ and $g: I \to \mathbb{R}^n$ be continuous, let $t_0 \in I$, and let $x_0 \in \mathbb{R}^n$. The inhomogeneous linear initial value problem is the problem of finding a continuous map $x: I \to \mathbb{R}^n$ such that $x$ is differentiable on $I^\circ$, satisfies $\dot{x}(t)=A(t)x(t)+g(t)$ for all $t \in I^\circ$, and satisfies $x(t_0)=x_0$. [/definition] The forcing term is transported by the homogeneous dynamics from each past time to the current time. We therefore need a formula that combines the fundamental matrix with the accumulated forcing, and variation of constants is exactly that representation. [quotetheorem:8318] For constant coefficients, this formula reduces to the matrix exponential. That case connects ODEs directly to eigenvalues, stability, and the phase portrait of linear systems. [example: Constant-Coefficient Linear Systems] Let $A \in \mathbb{R}^{n \times n}$ and consider $\dot{x}(t)=Ax(t)$ with $x(0)=x_0$. Define the matrix exponential by \begin{align*} e^{tA}=\sum_{k=0}^{\infty}\frac{t^kA^k}{k!}. \end{align*} This series converges for every $t\in\mathbb{R}$, because any submultiplicative matrix norm gives \begin{align*} \left\|\frac{t^kA^k}{k!}\right\|\le \frac{|t|^k\|A\|^k}{k!}, \end{align*} and the scalar series $\sum_{k=0}^{\infty}|t|^k\|A\|^k/k!=e^{|t|\|A\|}$ converges. Now differentiate the [power series](/page/Power%20Series) term by term on bounded $t$-intervals, justified by [uniform convergence](/page/Uniform%20Convergence) of the differentiated series. Since the $k=0$ term is constant, \begin{align*} \frac{d}{dt}e^{tA}=\sum_{k=1}^{\infty}\frac{k t^{k-1}A^k}{k!}. \end{align*} Because $k/k!=1/(k-1)!$, \begin{align*} \frac{d}{dt}e^{tA}=\sum_{k=1}^{\infty}\frac{t^{k-1}A^k}{(k-1)!}. \end{align*} Writing $m=k-1$ gives \begin{align*} \frac{d}{dt}e^{tA}=\sum_{m=0}^{\infty}\frac{t^mA^{m+1}}{m!}. \end{align*} Since $A^{m+1}=AA^m$ and left multiplication by $A$ is linear, \begin{align*} \frac{d}{dt}e^{tA}=A\sum_{m=0}^{\infty}\frac{t^mA^m}{m!}=Ae^{tA}. \end{align*} Also, \begin{align*} e^{0A}=\sum_{k=0}^{\infty}\frac{0^kA^k}{k!}=I_n, \end{align*} where the $k=0$ term is $I_n$ and every $k\ge 1$ term is zero. Hence $\Phi(t)=e^{tA}$ is the principal fundamental matrix based at $0$. For $x(t)=e^{tA}x_0$, differentiating gives \begin{align*} \dot{x}(t)=\frac{d}{dt}(e^{tA}x_0)=(Ae^{tA})x_0. \end{align*} By associativity of matrix multiplication, \begin{align*} (Ae^{tA})x_0=A(e^{tA}x_0)=Ax(t). \end{align*} At the initial time, \begin{align*} x(0)=e^{0A}x_0=I_nx_0=x_0. \end{align*} Thus $x(t)=e^{tA}x_0$ solves the initial value problem. The eigenvalues determine the long-time behaviour of this formula. Over $\mathbb{C}$, putting $A$ into [Jordan normal form](/theorems/864) writes $e^{tA}$ as a [change of basis](/page/Change%20Of%20Basis) times blocks of the form $e^{\lambda t}$ multiplied by polynomials in $t$. If every eigenvalue $\lambda$ of $A$ has $\operatorname{Re}\lambda<0$, each factor $t^m e^{\operatorname{Re}\lambda t}$ tends to $0$ as $t\to\infty$, so $e^{tA}x_0\to 0$ for every initial state $x_0$. If $A$ has an eigenvalue $\lambda$ with $\operatorname{Re}\lambda>0$ and eigenvector $v\ne 0$, then \begin{align*} e^{tA}v=e^{\lambda t}v, \end{align*} so \begin{align*} |e^{tA}v|=|e^{\lambda t}|\,|v|=e^{\operatorname{Re}\lambda t}|v|. \end{align*} This tends to $\infty$ as $t\to\infty$, so some solutions grow exponentially. [/example] Linear systems are not only solvable examples. They are also the first approximation to nonlinear systems near a trajectory or equilibrium, which is why the variational equation plays such a central role. ## Dynamical Interpretation ### Equilibria and Phase Portraits An autonomous initial value problem can be read as a rule for evolving every point of a state space. This perspective changes the question from solving a single equation to understanding the geometry of all solutions at once. The stationary points of the vector field are the simplest solutions. They are where the velocity vanishes, so a trajectory beginning there does not move. [definition: Equilibrium] Let $E \subset \mathbb{R}^n$ be open and let $V: E \to \mathbb{R}^n$ be a vector field. A point $x^* \in E$ is an equilibrium of the autonomous equation $\dot{x}=V(x)$ if $V(x^*)=0$. [/definition] Equilibria are the anchors of phase-space analysis. Near an equilibrium, the first question is whether nearby solutions approach it, move away from it, or exhibit mixed behaviour. [example: Logistic Growth as an Initial Value Problem] Let $r>0$ and $K>0$, and consider \begin{align*} \dot{x}(t)=rx(t)\left(1-\frac{x(t)}{K}\right), \qquad x(0)=x_0. \end{align*} The equilibria are the constant solutions. If $x(t)\equiv q$, then $\dot{x}(t)=0$, so $q$ must satisfy \begin{align*} rq\left(1-\frac{q}{K}\right)=0. \end{align*} Since $r>0$, this is equivalent to \begin{align*} q=0 \quad \text{or} \quad 1-\frac{q}{K}=0. \end{align*} The second equation gives $q=K$, so the equilibria are $0$ and $K$. Now assume $x_0>0$ and set \begin{align*} c=\frac{K-x_0}{x_0}. \end{align*} For times where the denominator is nonzero, define \begin{align*} x(t)=\frac{K}{1+ce^{-rt}}. \end{align*} At $t=0$, \begin{align*} x(0)=\frac{K}{1+c}=\frac{K}{1+(K-x_0)/x_0}=\frac{K}{K/x_0}=x_0. \end{align*} Differentiating the formula gives \begin{align*} \dot{x}(t)=K\frac{rce^{-rt}}{(1+ce^{-rt})^2}. \end{align*} On the other hand, \begin{align*} rx(t)\left(1-\frac{x(t)}{K}\right)=r\frac{K}{1+ce^{-rt}}\left(1-\frac{1}{1+ce^{-rt}}\right). \end{align*} Since \begin{align*} 1-\frac{1}{1+ce^{-rt}}=\frac{ce^{-rt}}{1+ce^{-rt}}, \end{align*} we get \begin{align*} rx(t)\left(1-\frac{x(t)}{K}\right)=\frac{rKce^{-rt}}{(1+ce^{-rt})^2}=\dot{x}(t). \end{align*} Thus this formula solves the initial value problem whenever $x_0>0$. If $0<x_0<K$, then $c>0$, so $\dot{x}(t)>0$ for every $t\ge 0$, and \begin{align*} K-x(t)=K-\frac{K}{1+ce^{-rt}}=\frac{Kce^{-rt}}{1+ce^{-rt}}>0. \end{align*} Thus the solution increases and stays below $K$. Since $e^{-rt}\to 0$ as $t\to\infty$, \begin{align*} \lim_{t\to\infty}x(t)=\frac{K}{1+0}=K. \end{align*} If $x_0>K$, then $-1<c<0$, so $1+ce^{-rt}>0$ for $t\ge 0$ and $\dot{x}(t)<0$. Also $K-x(t)<0$, so $x(t)>K$, and the same limit computation gives $x(t)\to K$. For $x_0=0$ and $x_0=K$, the corresponding solutions are the constant equilibrium solutions $x(t)\equiv 0$ and $x(t)\equiv K$. Thus the initial value determines which trajectory is followed, while the vector field fixes $K$ as the carrying-capacity equilibrium approached by every positive non-equilibrium solution in forward time. [/example] A single explicit solution rarely reveals the organisation of every trajectory. To study the whole autonomous system, we need to collect all maximal trajectories together with their orientation in time; that qualitative collection is the phase portrait. [definition: Phase Portrait] Let $E \subset \mathbb{R}^n$ be open and let $V: E \to \mathbb{R}^n$ be a vector field for which the autonomous initial value problem has unique maximal solutions. The phase portrait of $\dot{x}=V(x)$ is the collection of trajectories of its maximal solutions in $E$, together with their time orientation. [/definition] In dimension one, the phase portrait is often determined by the signs of $V(x)$. In higher dimensions, geometry becomes richer: trajectories may spiral, approach invariant sets, or organise around stable and unstable manifolds. ### Non-Crossing A common geometric obstruction is that unique autonomous trajectories cannot cross. If two curves meet at the same time-independent state, they have the same future and past while both are defined. [quotetheorem:8317] This theorem explains why phase portraits have coherent geometry. Apparent crossings in a drawn picture either represent different times along the same trajectory, a projection from a higher-dimensional system, or a failure of uniqueness. ## Beyond and Connected Topics Initial value problems are the starting point for [Cambridge IA Differential Equations](/page/Cambridge%20IA%20Differential%20Equations), where explicit solution methods, phase-line analysis, second-order equations, and systems of ODEs are developed systematically. The present page supplies the existence-and-uniqueness language behind those computational techniques. The local theory also connects to [Cambridge IA Analysis Notes](/page/Cambridge%20IA%20Analysis%20Notes), especially through completeness, uniform convergence, compactness, and fixed-point arguments. Picard iteration is a concrete place where abstract analysis produces a solution to a differential equation. Autonomous initial value problems lead naturally to [Cambridge II Dynamical Systems](/page/Cambridge%20II%20Dynamical%20Systems). Once uniqueness gives a flow, the next questions concern equilibria, stability, periodic orbits, invariant sets, bifurcations, and long-time behaviour. There is also a complex-analytic direction. Holomorphic vector fields and complex differential equations appear in [Cambridge IB Complex Analysis](/page/Cambridge%20IB%20Complex%20Analysis), where analytic dependence and power-series methods provide stronger local structure than continuity or real differentiability alone. Partial differential equations have their own initial value problems, but the meaning changes. For a PDE, initial data may be a function on space rather than a point in $\mathbb{R}^n$, and existence theory often takes place in Banach or Sobolev spaces. The finite-dimensional IVP remains the guiding model: prescribe a state, evolve by a law, and study existence, uniqueness, stability, and continuation. ## References Androma, [Cambridge IA Analysis Notes](/page/Cambridge%20IA%20Analysis%20Notes). Androma, [Cambridge IB Complex Analysis](/page/Cambridge%20IB%20Complex%20Analysis). Androma, [Cambridge IA Differential Equations](/page/Cambridge%20IA%20Differential%20Equations). Androma, [Cambridge II Dynamical Systems](/page/Cambridge%20II%20Dynamical%20Systems). V. I. Arnold, *Ordinary Differential Equations* (1992). Morris W. Hirsch, Stephen Smale, and Robert L. Devaney, *Differential Equations, Dynamical Systems, and an Introduction to Chaos* (2013). Philip Hartman, *Ordinary Differential Equations* (1964). Jack K. Hale, *Ordinary Differential Equations* (1969).