Autonomous Systems and Phase Space

What does it mean to say a system's future depends only on where it is now? It never depends on what time it is. This question separates autonomous systems from the general time-dependent mess of differential equations. It turns out to be the single most useful simplifying idea in the qualitative study of dynamics. When the rules governing change do not themselves change with time, the state space of the system becomes a geometric object in its own right — a phase space. Through it, solutions flow like particles carried by an invisible current, a vector field. This shift in viewpoint, from solving equations symbolically to picturing flows geometrically, is the origin of modern nonlinear dynamics. Henri Poincaré realized at the end of the nineteenth century that most differential equations arising in nature cannot be solved in closed form. Yet their qualitative behavior — where trajectories go, whether they spiral in or fly off, whether they cycle forever — can often be determined without ever writing down a formula for the solution. This is the birth of the "geometric theory of differential equations." It underlies everything from planetary orbits to neuron firing to epidemic curves. The big picture is this: an autonomous system of first-order ODEs assigns to every point in some space a velocity vector. A solution curve is a path that at every instant moves with exactly the velocity prescribed at its current location. Because the rule never changes, the whole future (and past) trajectory through a point is completely determined by that point alone. No clock is needed. This lets us draw the phase portrait: the entire family of trajectories sketched in the state space. It is a single static picture encoding the dynamics for all time and all initial conditions simultaneously. This article builds the rigorous foundation: precise definitions of autonomous systems, trajectories, orbits, the flow map, and the phase-space picture. It also covers the key structural theorem — the impossibility of trajectory crossing — that makes phase portraits meaningful. It sets up the vocabulary needed for the next natural questions. These are where a system can come to rest and how it behaves near such points.

Key takeaways
  • An autonomous system is x˙=f(x)\dot{\mathbf{x}} = \mathbf{f}(\mathbf{x}): the vector field f\mathbf{f} depends only on the state, not on time.
  • Phase space is the domain of f\mathbf{f}; a trajectory (or orbit) is the image of a solution curve traced there.
  • The flow ϕt\phi_t is the time-tt map sending an initial state to the state reached after time tt; it satisfies ϕ0=id\phi_0 = \mathrm{id} and ϕs+t=ϕs∘ϕt\phi_{s+t} = \phi_s \circ \phi_t.
  • Time-translation invariance means solutions can be shifted in time freely: if x(t)\mathbf{x}(t) solves the system, so does x(t−t0)\mathbf{x}(t-t_0).
  • Under standard smoothness hypotheses, distinct trajectories in phase space never cross — this is what makes a phase portrait a coherent, unambiguous picture.
  • Phase portraits reveal qualitative behavior (equilibria, cycles, separatrices) even when no explicit formula for solutions exists.
  • Key idea: time drops out of the picture, so all the dynamical information lives in the geometry of a single vector field on phase space.

Where this lives today

The autonomous/phase-space viewpoint is the standard language of essentially all of applied dynamics. In physics, the state of a mechanical system (positions and momenta) evolves autonomously under Hamilton's equations whenever the Hamiltonian has no explicit time-dependence, and phase space there literally means the space of positions and momenta central to classical and statistical mechanics. In population biology and epidemiology, models like Lotka–Volterra predator-prey equations and SIR epidemic models are autonomous systems whose phase portraits (predator vs. prey, susceptible vs. infected) reveal oscillations, extinction thresholds, and epidemic peaks without ever solving the equations explicitly. In electrical engineering, autonomous circuit equations (e.g., the van der Pol oscillator) are analyzed entirely through their phase portraits to predict self-sustained oscillations. In neuroscience, the Hodgkin–Huxley and FitzHugh–Nagumo models are studied as autonomous systems in a voltage-recovery phase plane to understand action potential generation. In control theory and robotics, autonomous state-space representations x˙=f(x)\dot{\mathbf x} = \mathbf f(\mathbf x) (with feedback control folded into f\mathbf f) underlie stability analysis via Lyapunov functions, a direct descendant of the "decreasing VV along trajectories" argument used above. In numerical analysis and scientific computing, the flow map ϕt\phi_t is exactly the object that time-stepping algorithms (Runge–Kutta, symplectic integrators) approximate. Even in pure mathematics, ergodic theory and topological dynamics take the flow {ϕt}\{\phi_t\} generated by an autonomous vector field as their primary object of study, extending the ideas here to measure-theoretic and topological settings far beyond Rn\mathbb{R}^n.

Ask the tutor: How do I locate and classify the equilibrium points of an autonomous system once I have its phase portrait?

DefinitionAutonomous system, trajectory, and flow

Let D⊆RnD \subseteq \mathbb{R}^n be an open set and let

f:D→Rn\mathbf{f} : D \to \mathbb{R}^n

be a continuous vector field. An autonomous system of ordinary differential equations is

x˙=f(x),x∈D,\dot{\mathbf{x}} = \mathbf{f}(\mathbf{x}), \qquad \mathbf{x} \in D,

where x˙=dx/dt\dot{\mathbf{x}} = d\mathbf{x}/dt and f\mathbf{f} has no explicit dependence on the independent variable tt. The set DD is called phase space (or the state space). A solution is a differentiable curve x:I→D\mathbf{x} : I \to D, defined on an open interval I⊆RI \subseteq \mathbb{R}, satisfying x′(t)=f(x(t))\mathbf{x}'(t) = \mathbf{f}(\mathbf{x}(t)) for every t∈It \in I. Given an initial condition x(t0)=x0∈D\mathbf{x}(t_0) = \mathbf{x}_0 \in D, we call x(t)\mathbf{x}(t) the solution through x0\mathbf{x}_0. The trajectory (or orbit) through x0\mathbf{x}_0 is the set

γ(x0)={ x(t):t∈I }⊆D,\gamma(\mathbf{x}_0) = \{\, \mathbf{x}(t) : t \in I \,\} \subseteq D,

i.e. the image of the solution curve in phase space, stripped of its time parametrization. A phase portrait is the collection of all trajectories of the system, sketched in DD, typically with arrows indicating the direction of increasing tt. When f\mathbf{f} is locally Lipschitz on DD (in particular if f∈C1(D)\mathbf{f} \in C^1(D)), the solution through each point exists and is unique on a maximal interval of existence. The flow generated by the system is the family of maps

ϕt:Dt→D,ϕt(x0)=x(t;x0),\phi_t : D_t \to D, \qquad \phi_t(\mathbf{x}_0) = \mathbf{x}(t; \mathbf{x}_0),

where x(⋅;x0)\mathbf{x}(\cdot; \mathbf{x}_0) is the maximal solution with x(0)=x0\mathbf{x}(0) = \mathbf{x}_0, and DtD_t is the (open) set of points whose solution survives to time tt.

Intuition

Picture phase space as a landscape and f(x)\mathbf{f}(\mathbf{x}) as an arrow attached to every point, indicating the instantaneous velocity a particle would have if it were sitting there. Because f\mathbf{f} does not know what time it is — only where the particle is — the arrow at a given location is fixed forever. A trajectory is then simply the path a particle traces out as it is swept along by this frozen field of arrows, like a leaf carried by a steady (time-independent) current in a river. This is the crucial contrast with non-autonomous systems x˙=f(x,t)\dot{\mathbf{x}} = \mathbf{f}(\mathbf{x}, t), where the current itself changes over time: the arrow at a fixed location today may point differently tomorrow. For autonomous systems, one picture — one arrow diagram — captures the entire dynamical law for all time. This is why we can draw a single phase portrait and read off the qualitative fate of every initial condition at a glance, without solving anything. A second key intuition is time-translation invariance: since the rule never changes, delaying an initial condition just delays the resulting motion without altering its shape. If x(t)\mathbf{x}(t) solves the system, then y(t)=x(t−t0)\mathbf{y}(t) = \mathbf{x}(t - t_0) solves it too, for any constant t0t_0. Consequently, many different solution curves (parametrized differently in tt) can trace out the same trajectory in phase space, merely shifted in time. This is precisely why we distinguish a solution (a time-parametrized curve) from a trajectory (its time-independent geometric shadow) — the phase portrait records only the latter.

Figure
A phase portrait for x˙=y, y˙=−x\dot{x} = y,\ \dot{y} = -x: circular trajectories traced by the flow
x y (0,0) flow direction
TheoremExistence, uniqueness, and non-crossing of trajectories

Let D⊆RnD \subseteq \mathbb{R}^n be open and let f∈C1(D)\mathbf{f} \in C^1(D) (or more generally, let f\mathbf{f} be locally Lipschitz on DD). Consider the autonomous system

x˙=f(x).\dot{\mathbf{x}} = \mathbf{f}(\mathbf{x}).

Then: 1. (Existence and uniqueness) For every x0∈D\mathbf{x}_0 \in D there exists ε>0\varepsilon > 0 and a unique solution x(t)\mathbf{x}(t) defined on (−ε,ε)(-\varepsilon, \varepsilon) with x(0)=x0\mathbf{x}(0) = \mathbf{x}_0; this solution extends uniquely to a maximal open interval of existence I(x0)I(\mathbf{x}_0). 2. (Time-translation invariance) If x(t)\mathbf{x}(t) is a solution on interval II, then for any constant t0∈Rt_0 \in \mathbb{R}, the curve y(t):=x(t−t0)\mathbf{y}(t) := \mathbf{x}(t - t_0) is also a solution, on the shifted interval I+t0I + t_0. 3. (Non-crossing of trajectories) If two trajectories γ(a)\gamma(\mathbf{a}) and γ(b)\gamma(\mathbf{b}) share a common point, then γ(a)=γ(b)\gamma(\mathbf{a}) = \gamma(\mathbf{b}) as sets. Equivalently: distinct trajectories in the phase portrait never intersect.

Proof

We prove each part in turn. The technique throughout is the Picard–Lindelöf existence-uniqueness theorem together with elementary manipulation of solution curves; part 3 is proved by a uniqueness + time-shift argument.

Part 1 (Existence and uniqueness). Since f\mathbf{f} is locally Lipschitz on the open set DD and x0∈D\mathbf{x}_0 \in D, choose a closed ball B‾(x0,r)⊂D\overline{B}(\mathbf{x}_0, r) \subset D on which f\mathbf{f} is Lipschitz with constant LL and bounded, ∥f∥≤M\|\mathbf{f}\| \le M. The Picard–Lindelöf theorem (proved via the contraction mapping principle applied to the integral operator Tx(t)=x0+∫0tf(x(s)) dsT\mathbf{x}(t) = \mathbf{x}_0 + \int_0^t \mathbf{f}(\mathbf{x}(s))\,ds on the space of continuous curves with the sup norm) yields existence and uniqueness of a solution on some interval (−ε,ε)(-\varepsilon,\varepsilon) with ε=min⁡(r/M,1/(2L))\varepsilon = \min(r/M, 1/(2L)), or any standard variant of this estimate. This is a foundational result assumed as prerequisite machinery, not itself proved here, since it does not depend on the autonomous structure. To extend to a maximal interval: let I(x0)I(\mathbf{x}_0) be the union of all open intervals containing 00 on which a solution through x0\mathbf{x}_0 exists. Any two such solutions agree on the overlap of their domains, because if they disagreed at some first point t1t_1, the local uniqueness statement applied at t1t_1 (with initial data taken at t1t_1 instead of 00, valid since f\mathbf f has no explicit tt-dependence) would contradict their agreement immediately before t1t_1. Hence all these local solutions glue into a single well-defined solution on I(x0)I(\mathbf{x}_0), and by construction no further extension is possible, so I(x0)I(\mathbf{x}_0) is maximal.

Part 2 (Time-translation invariance). Suppose x(t)\mathbf{x}(t) satisfies x′(t)=f(x(t))\mathbf{x}'(t) = \mathbf{f}(\mathbf{x}(t)) for t∈It \in I. Define y(t)=x(t−t0)\mathbf{y}(t) = \mathbf{x}(t - t_0) for t∈I+t0:={t+t0:t∈I}t \in I + t_0 := \{t + t_0 : t \in I\}. By the chain rule,

y′(t)=ddtx(t−t0)=x′(t−t0)⋅1=f(x(t−t0))=f(y(t)).\mathbf{y}'(t) = \frac{d}{dt}\mathbf{x}(t - t_0) = \mathbf{x}'(t - t_0) \cdot 1 = \mathbf{f}\big(\mathbf{x}(t-t_0)\big) = \mathbf{f}(\mathbf{y}(t)).

This computation uses only that f\mathbf f does not depend explicitly on tt: the factor of 11 from the chain rule is what would instead become, e.g., f(x(t−t0),t−t0)\mathbf f(\mathbf x(t-t_0), t-t_0) if f\mathbf f depended on time, breaking the identity. Hence y\mathbf{y} solves the same autonomous equation, proving the claim.

Part 3 (Non-crossing). Suppose γ(a)\gamma(\mathbf{a}) and γ(b)\gamma(\mathbf{b}) share a point p∈D\mathbf{p} \in D. Let x(t)\mathbf{x}(t) be the maximal solution with x(0)=a\mathbf{x}(0) = \mathbf{a}, and z(t)\mathbf{z}(t) the maximal solution with z(0)=b\mathbf{z}(0) = \mathbf{b}. By hypothesis there exist times t1,t2t_1, t_2 with

x(t1)=p=z(t2).\mathbf{x}(t_1) = \mathbf{p} = \mathbf{z}(t_2).

Define w(t):=x(t+t1−t2)\mathbf{w}(t) := \mathbf{x}(t + t_1 - t_2). By Part 2 (with shift t0=t2−t1t_0 = t_2 - t_1), w\mathbf{w} is a solution of the same system, and

w(t2)=x(t2+t1−t2)=x(t1)=p=z(t2).\mathbf{w}(t_2) = \mathbf{x}(t_2 + t_1 - t_2) = \mathbf{x}(t_1) = \mathbf{p} = \mathbf{z}(t_2).

So w\mathbf{w} and z\mathbf{z} are both solutions of the autonomous system agreeing at time t2t_2. By the uniqueness statement in Part 1 (applied with initial time t2t_2 instead of 00, which is legitimate since autonomy means the choice of initial time is immaterial to the theorem's applicability), w(t)=z(t)\mathbf{w}(t) = \mathbf{z}(t) for all tt in the common part of their maximal intervals of existence, and in fact their maximal intervals must coincide (else one could be used to extend the other, contradicting maximality). Therefore, for every tt, z(t)=w(t)=x(t+t1−t2)\mathbf{z}(t) = \mathbf{w}(t) = \mathbf{x}(t + t_1 - t_2), so

γ(b)={z(t)}={x(t+t1−t2)}={x(s):s}=γ(a),\gamma(\mathbf{b}) = \{\mathbf{z}(t)\} = \{\mathbf{x}(t + t_1 - t_2)\} = \{\mathbf{x}(s) : s\} = \gamma(\mathbf{a}),

where the last equality holds because s=t+t1−t2s = t + t_1 - t_2 ranges over the same maximal interval as t1t_1's solution does (the shift is a bijection of the real line onto itself and the two maximal intervals were just shown to coincide up to this shift). Thus γ(a)=γ(b)\gamma(\mathbf{a}) = \gamma(\mathbf{b}) as subsets of DD, completing the proof. ■\blacksquare

∎
CorollaryTrajectories partition phase space

Under the hypotheses of the theorem, the trajectories {γ(x):x∈D}\{\gamma(\mathbf{x}) : \mathbf{x} \in D\} form a partition of DD: every point of DD lies on exactly one trajectory. This follows immediately, since every point lies on its own trajectory (reflexivity, taking t=0t=0), and Part 3 shows any two trajectories sharing a point are identical.

Worked examples

Example 1: A nonlinear planar system with an explicit invariant. Consider on D=R2D = \mathbb{R}^2 the autonomous system

x˙=−y−x3,y˙=x−y3.\dot{x} = -y - x^3, \qquad \dot{y} = x - y^3.

Here f(x,y)=(−y−x3, x−y3)\mathbf{f}(x,y) = (-y - x^3,\ x - y^3) is a polynomial vector field, hence C1(R2)C^1(\mathbb{R}^2), so existence, uniqueness, and non-crossing all apply globally on D=R2D = \mathbb{R}^2. We do not attempt to solve this system in closed form (it has no elementary solution). Instead we extract qualitative phase-space information directly. Consider V(x,y)=x2+y2V(x,y) = x^2 + y^2, the squared distance from the origin. Along any trajectory,

ddtV(x(t),y(t))=2xx˙+2yy˙=2x(−y−x3)+2y(x−y3)=−2x4−2y4.\frac{d}{dt}V(x(t),y(t)) = 2x\dot x + 2y\dot y = 2x(-y-x^3) + 2y(x - y^3) = -2x^4 - 2y^4.

This is ≤0\le 0 everywhere, and strictly negative except at the origin. So VV is strictly decreasing along every trajectory not already at (0,0)(0,0). Geometrically: every trajectory moves monotonically inward across circles x2+y2=c2x^2+y^2 = c^2, spiraling (or moving directly) toward the origin as t→∞t \to \infty. No trajectory can be periodic (a closed loop would require VV to return to its starting value, contradicting strict monotonic decrease), and by the non-crossing theorem, distinct trajectories fill out R2∖{0}\mathbb{R}^2\setminus\{0\} without ever touching. This single computation — no explicit solution formula needed — already tells us the entire qualitative phase portrait: all trajectories funnel into the origin.

Example 2: Distinguishing trajectory from solution via time-shift, asymmetric initial data. Consider the linear autonomous system

x˙=y,y˙=−x,\dot x = y, \qquad \dot y = -x,

on D=R2D = \mathbb{R}^2, i.e. f(x,y)=(y,−x)\mathbf{f}(x,y) = (y,-x), clearly C1C^1 (in fact linear), so the theorem's hypotheses hold globally. Take the asymmetric initial condition (x(0),y(0))=(3,−1)(x(0), y(0)) = (3, -1) (chosen off any axis of symmetry to avoid masking errors). The system is equivalent to x¨=−x\ddot{x} = -x, with the given data giving x(0)=3x(0) = 3, x˙(0)=y(0)=−1\dot x(0) = y(0) = -1. The general solution of x¨=−x\ddot x = -x is x(t)=Acos⁡t+Bsin⁡tx(t) = A\cos t + B \sin t; matching data:

A=3,x˙(0)=B=−1.A = 3, \qquad \dot x(0) = B = -1.

So

x(t)=3cos⁡t−sin⁡t,y(t)=x˙(t)=−3sin⁡t−cos⁡t.x(t) = 3\cos t - \sin t, \qquad y(t) = \dot x(t) = -3\sin t - \cos t.

Check: x(0)=3x(0) = 3, y(0)=−3(0)−1=−1y(0) = -3(0) - 1 = -1. Good. Also verify the ODE directly: y˙=−3cos⁡t+sin⁡t=−(3cos⁡t−sin⁡t)=−x\dot y = -3\cos t + \sin t = -(3\cos t - \sin t) = -x. Correct. Now consider the time-shifted solution x^(t)=x(t−π/2)\hat x(t) = x(t - \pi/2), per the theorem's Part 2. Compute

x^(t)=3cos⁡(t−π2)−sin⁡(t−π2)=3sin⁡t+cos⁡t,\hat x(t) = 3\cos(t - \tfrac{\pi}{2}) - \sin(t - \tfrac{\pi}{2}) = 3\sin t + \cos t,

using cos⁡(t−π/2)=sin⁡t\cos(t - \pi/2) = \sin t and sin⁡(t−π/2)=−cos⁡t\sin(t-\pi/2) = -\cos t. This is manifestly a different function of tt from x(t)=3cos⁡t−sin⁡tx(t) = 3\cos t - \sin t — different solution curve, different initial point at t=0t=0 (namely x^(0)=1≠3\hat x(0) = 1 \ne 3). Yet both curves trace out the same trajectory in the phase plane: both are points on the circle

x2+y2=9cos⁡2t−6cos⁡tsin⁡t+sin⁡2t+9sin⁡2t+6sin⁡tcos⁡t+cos⁡2t=9+1=10,x^2 + y^2 = 9\cos^2 t - 6\cos t \sin t + \sin^2 t + 9\sin^2 t + 6\sin t\cos t + \cos^2 t = 9 + 1 = 10,

(a direct computation confirms x2+y2≡10x^2+y^2 \equiv 10 along the original solution, since x2+y2x^2+y^2 is conserved for this system, as ddt(x2+y2)=2xx˙+2yy˙=2xy−2yx=0\frac{d}{dt}(x^2+y^2) = 2x\dot x + 2y \dot y = 2xy - 2yx = 0). So γ((3,−1))\gamma((3,-1)) is exactly the circle of radius 10\sqrt{10} centered at the origin, traversed clockwise, and every time-shift of the solution stays on this same circle — illustrating concretely that infinitely many distinct solutions (parametrized differently in tt) share one and the same trajectory.

CounterexampleDropping local Lipschitz continuity breaks non-crossing

The non-crossing theorem requires f\mathbf{f} locally Lipschitz (guaranteed, e.g., by f∈C1\mathbf f \in C^1). If we only assume f\mathbf f is continuous, uniqueness — and hence non-crossing — can fail. Take the scalar autonomous equation on D=RD = \mathbb{R}:

x˙=f(x)=3x2/3,\dot x = f(x) = 3x^{2/3},

where ff is continuous on all of R\mathbb{R} but not Lipschitz near x=0x=0 (its derivative f′(x)=2x−1/3f'(x) = 2x^{-1/3} blows up as x→0x \to 0). Consider the trajectory through a=0\mathbf{a} = 0: the constant function x(t)≡0x(t) \equiv 0 is a solution, since x˙=0=3(0)2/3\dot x = 0 = 3(0)^{2/3}. So γ(0)={0}\gamma(0) = \{0\}. But also consider z(t)=t3z(t) = t^3. Then z˙=3t2\dot z = 3t^2, while f(z(t))=3(t3)2/3=3t2f(z(t)) = 3(t^3)^{2/3} = 3t^2 (for t≥0t \ge 0; taking the real cube root and t2t^{2} nonnegative makes this consistent for all real tt since (t3)2/3=t2(t^3)^{2/3} = t^2). So z(t)=t3z(t) = t^3 is also a solution, and z(0)=0z(0) = 0. Now both x(t)≡0x(t)\equiv 0 and z(t)=t3z(t) = t^3 are solutions passing through the point 00 at time t=0t=0 — uniqueness has failed. Their trajectories both contain the point 00: γ(0)={0}\gamma(0) = \{0\} (from the constant solution) while γ\gamma of the point z(1)=1z(1)=1, traced backward, passes through 00 too, arriving at 00 in finite time and then — nothing forces it to stay, since another solution through 00 (the zero solution) also exists. In effect, the "trajectory" through the origin is no longer a single well-defined curve: multiple genuinely different solution curves emanate from the same point, and the clean statement "two trajectories sharing a point are identical" collapses, because there are now infinitely many distinct trajectories through 00 (e.g. zc(t)=(t−c)3z_c(t) = (t-c)^3 for t≥ct \ge c, glued to 00 for t≤ct \le c, for any constant cc). The phase portrait becomes ambiguous exactly at x=0x=0: arrows no longer determine a unique path, so drawing "the" trajectory through the origin is meaningless. This is precisely what local Lipschitz continuity is hypothesized to prevent.

Scope, edge cases & common traps

SituationWhat holdsWhat can fail
f∈C1(D)\mathbf f \in C^1(D) or locally LipschitzExistence, uniqueness, non-crossing all hold—
f\mathbf f merely continuousExistence holds (Peano)Uniqueness and non-crossing can fail (see counterexample)
f(x,t)\mathbf f(\mathbf x, t) genuinely time-dependentExistence/uniqueness as usualTrajectories (in x\mathbf x-space alone) can cross; must use (x,t)(\mathbf x,t)-extended phase space to restore non-crossing
Maximal interval I(x0)I(\mathbf x_0)May be a proper subset of R\mathbb R (finite-time blow-up)Phase portrait still well-defined on DD; just not all trajectories are "for all time"
Equilibrium points f(x∗)=0\mathbf f(\mathbf x^*) = 0Constant solution x(t)≡x∗\mathbf x(t) \equiv \mathbf x^*, a legitimate trajectory (a single point)Sometimes mistaken as "not a trajectory" — it is one, of the degenerate kind

The mathematician behind it

Origins · 1 minJules Henri Poincaré and how this was discovered
JH
Jules Henri Poincaré
1854–1912 · France, late 19th–early 20th century

Poincaré transformed the study of differential equations from a search for explicit formulas into a geometric and topological science. Investigating the three-body problem for celestial mechanics — for which no general closed-form solution exists — he developed the qualitative theory of dynamical systems: phase space, the classification of equilibria, limit cycles, and the beginnings of what would become chaos theory. His memoir on the three-body problem, submitted for King Oscar II of Sweden's prize competition, contained the first description of what we now call sensitive dependence on initial conditions.

History
- 1881–1886: Poincaré publishes "Mémoire sur les courbes définies par une équation différentielle," founding the geometric theory of ODEs and introducing phase-plane analysis.
- 1890: Poincaré's prize memoir on the three-body problem reveals the possibility of highly complicated (what we'd now call chaotic) trajectories in a deterministic autonomous system.
- 1892–1899: Poincaré's Les méthodes nouvelles de la mécanique céleste systematizes phase-space methods, periodic orbits, and stability for autonomous Hamiltonian systems.
- 1937–1950s: Andronov and collaborators formalize the modern theory of autonomous planar systems, structural stability, and bifurcations, building directly on Poincaré's phase-plane picture.
- 1960s–1970s: Smale, Lorenz, and others extend the phase-space viewpoint to higher-dimensional autonomous systems, discovering strange attractors and deterministic chaos.
What exactly makes a system "autonomous"?

A system x˙=f(x,t)\dot{\mathbf x} = \mathbf f(\mathbf x, t) is autonomous precisely when the right-hand side has no explicit occurrence of the independent variable tt, so it can be written x˙=f(x)\dot{\mathbf x} = \mathbf f(\mathbf x). The vector field assigns a velocity to each location in phase space, and that assignment never changes as time passes. Concretely, x˙=x2−1\dot x = x^2 - 1 is autonomous, while x˙=x2−sin⁡t\dot x = x^2 - \sin t is not, because the second system's velocity at the same state xx differs depending on when you look. Autonomy is what allows a single, static picture (the phase portrait) to encode the complete dynamics for all time.

Is every non-autonomous system secretly autonomous in a bigger space?

Yes, and this trick is used constantly. Given x˙=f(x,t)\dot{\mathbf x} = \mathbf f(\mathbf x, t), introduce an extra variable τ\tau with τ˙=1\dot \tau = 1, τ(0)=0\tau(0)=0, and consider the enlarged state (x,τ)(\mathbf x, \tau). The new system x˙=f(x,τ), τ˙=1\dot{\mathbf x} = \mathbf f(\mathbf x,\tau),\ \dot\tau = 1 has a right-hand side depending only on (x,τ)(\mathbf x,\tau), hence is autonomous on the extended phase space D×RD \times \mathbb R. This restores uniqueness and non-crossing in the extended space, even though trajectories projected down onto the original x\mathbf x-space alone may cross.

What is the difference between a solution and a trajectory?

A solution is a specific time-parametrized function x(t)\mathbf x(t) satisfying the ODE, complete with a "clock" telling you where the state is at each instant tt. A trajectory is the image of that curve in phase space — the geometric path traced out — with the time-labeling forgotten. Because of time-translation invariance, infinitely many different solutions (differing only by when they pass through each point) correspond to the very same trajectory. The phase portrait draws trajectories, not solutions, which is exactly why it can display the whole dynamics in one static picture.

Can a trajectory intersect itself?

Yes — this is exactly how periodic orbits appear in a phase portrait: the trajectory is a closed curve, and the solution returns to the same point after a fixed period TT, i.e., x(t+T)=x(t)\mathbf x(t+T) = \mathbf x(t) for all tt. This does not violate the non-crossing theorem, because non-crossing forbids two different trajectories from sharing a point; a single trajectory closing up on itself is one trajectory, traversed repeatedly, not two crossing curves. Non-closed, non-periodic trajectories, by contrast, can never self-intersect without becoming periodic, by the same uniqueness argument applied to two different times along the same solution.

Does existence and uniqueness hold for all time, or just locally?

Only locally, in general. The Picard–Lindelöf argument guarantees a unique solution on some small interval around the initial time, and this extends to a unique maximal interval of existence — but that maximal interval can be finite. A classic example is x˙=x2\dot x = x^2 with x(0)=1x(0)=1, whose solution x(t)=1/(1−t)x(t) = 1/(1-t) blows up as t→1−t \to 1^-, so the maximal interval is (−∞,1)(-\infty, 1), not all of R\mathbb R. The phase portrait and trajectory are still perfectly well-defined objects; one just cannot always follow a trajectory forever in time.

Why does the non-crossing property matter practically, beyond aesthetics?

It is what makes a phase portrait a legitimate predictive tool rather than just a pretty picture. Because trajectories partition phase space without crossing, once you know which trajectory a system is on, you know its entire past and future path is confined to that curve — it cannot "jump" to a different branch. This is the geometric backbone of stability arguments: proving a trajectory cannot escape a bounded region (because escaping would require crossing an already-established boundary trajectory) is a standard technique for proving global boundedness or convergence to an equilibrium, as used implicitly for Fixed Points and Equilibria.

How does this connect to what comes next in the theory?

Once phase space and trajectories are established, the natural next question is: where can a trajectory possibly come to permanent rest? These are the equilibrium points, where f(x∗)=0\mathbf f(\mathbf x^*) = 0, studied under Fixed Points and Equilibria. To understand the qualitative behavior of trajectories near such a point, one approximates the nonlinear vector field by its best linear approximation, the subject of Linearization and the Jacobian. Both of these build directly on the vocabulary — trajectory, flow, phase portrait — established here, and both rely on the Vector Fields perspective of treating f\mathbf f as an arrow attached to every point of phase space.

Where this leads