Discretization, Time-Stepping, and the Bridge to Control
discretization, time stepping, sampled-data control, flow map, zero-order hold
1 Role
This is the sixth page of the ODEs and Dynamical Systems module.
Its job is to turn the continuous-time story into the language of updates:
a trajectory over one time interval becomes a map from the current state to the next state
That is the bridge both to numerical simulation and to digital control.
2 First-Pass Promise
Read this page after Lyapunov Functions, Invariant Sets, and Long-Time Behavior, together with Time-Stepping for ODEs and Stability.
If you stop here, you should still understand:
- how an ODE induces an exact step-to-step flow map
- how numerical time stepping approximates that map
- how sampled inputs turn continuous-time control into a discrete update law
- why continuous-time stability and discrete-time stability are related but not identical
3 Why It Matters
Continuous-time models are written as differential equations, but computers and digital controllers act in steps.
So one of the most important translations in applied mathematics is:
\[ \text{continuous law of change} \quad\longrightarrow\quad \text{discrete state update}. \]
This translation matters because:
- simulation evaluates the state only at finitely many times
- digital controllers sample sensors and update actuators on a clock
- optimization and estimation routines often run on discrete models even when the physical system is continuous
- stability can look different after discretization if the step size is too large or the scheme is a poor approximation
This page is where the ODE module meets both Numerical Methods and the later Control layer.
4 Prerequisite Recall
- linear systems evolve by the matrix exponential
e^{tA} - time stepping approximates a continuous trajectory by repeated local updates
- Lyapunov and stability arguments care about long-time behavior, not only one local derivative
- a state-space model is organized around the current state and what rule moves it forward
5 Intuition
5.1 Every ODE Has An Exact One-Step Map
For a differential equation
\[ x' = f(x), \]
starting from a state x_k and waiting one time interval h produces a new state.
That means the continuous dynamics induce an exact map
\[ x_{k+1} = \Phi_h(x_k). \]
This map is not an approximation. It is the true flow sampled every h units of time.
5.2 Numerical Methods Build Approximate Step Maps
A numerical method replaces the exact sampled flow \Phi_h by a computable update \Psi_h.
So the basic numerical question is:
how well does the approximate step map mimic the exact one?
5.3 Inputs Turn ODEs Into Controlled Update Laws
For a controlled system
\[ x' = f(x,u), \]
the next state depends not only on the current state but also on the input applied during the step.
If the input is held constant over each interval, then every time step becomes a state-update rule involving both x_k and u_k.
5.4 Discretization Is Where Control Becomes Digital
The physical system lives in continuous time, but the controller often sees:
- a sampled state
- a chosen input for the next interval
- an update map that predicts the next sampled state
This is the core sampled-data bridge.
5.5 Stability Can Change Under Discretization
A continuous-time system may be stable, but an explicit numerical scheme with a large step size can still create an unstable update rule.
So the right first-pass lesson is:
discretization is not harmless bookkeeping; it changes the dynamical object we are actually iterating
6 Formal Core
Definition 1 (Definition: Exact Flow Map) For an autonomous ODE
\[ x' = f(x), \]
the exact time-h flow map \Phi_h sends an initial state x_k to the true state reached after time h:
\[ x_{k+1}=\Phi_h(x_k). \]
Definition 2 (Definition: One-Step Discretization) A one-step discretization replaces the exact flow map by a computable update rule
\[ x_{k+1}=\Psi_h(x_k). \]
For controlled systems, the update depends on the input as well:
\[ x_{k+1}=\Psi_h(x_k,u_k). \]
Definition 3 (Definition: Sampled-Data / Zero-Order-Hold Input) For a controlled ODE
\[ x' = f(x,u), \]
a zero-order-hold input means the control is held constant on each interval:
\[ u(t)=u_k,\qquad t\in [kh,(k+1)h). \]
This is the simplest and most common bridge from continuous-time models to digital control.
Theorem 1 (Theorem Idea: Exact Discretization Of An LTI System) For the linear time-invariant controlled system
\[ x' = Ax + Bu, \]
with u(t)=u_k on [kh,(k+1)h), the sampled states satisfy
\[ x_{k+1}=e^{hA}x_k+\Gamma_h u_k, \]
where
\[ \Gamma_h=\int_0^h e^{(h-\tau)A}B\,d\tau. \]
This is the exact discrete-time model induced by the continuous dynamics under held inputs.
Theorem 2 (Theorem Idea: Forward Euler Gives The Simplest Approximate Control Update) Applying forward Euler to
\[ x' = Ax + Bu \]
gives
\[ x_{k+1}=(I+hA)x_k + hBu_k. \]
This is computationally simple, but it is only an approximation to the exact sampled-data model.
7 A Small Worked Example
Consider the scalar controlled system
\[ x' = -2x + u. \]
Assume the input is held constant on each interval:
\[ u(t)=u_k,\qquad t\in [kh,(k+1)h). \]
7.1 Step 1: Write The Exact Sampled Update
The homogeneous part decays like e^{-2t}, so the exact one-step sampled model is
\[ x_{k+1}=e^{-2h}x_k + \frac{1-e^{-2h}}{2}\,u_k. \]
This is the true discrete-time system seen at sampling times.
7.2 Step 2: Write The Forward-Euler Update
Forward Euler gives
\[ x_{k+1}=x_k+h(-2x_k+u_k) =(1-2h)x_k + h u_k. \]
7.3 Step 3: Compare Stability When u_k=0
For the continuous-time system with zero input,
\[ x'=-2x, \]
the origin is asymptotically stable.
For the exact sampled system, the multiplier is
\[ e^{-2h}, \]
which always satisfies |e^{-2h}|<1 for every h>0.
So the exact sampled model is stable for any positive sampling interval.
For forward Euler, the multiplier is
\[ 1-2h. \]
This is stable only when
\[ |1-2h|<1, \]
which means
\[ 0<h<1. \]
So a continuous-time stable system can produce an unstable explicit update if the step is too large.
7.4 Step 4: Read The Control Meaning
The state update is now in the form
\[ x_{k+1}=a_h x_k + b_h u_k. \]
That is exactly the language of discrete-time control:
- current state
x_k - chosen control input
u_k - predicted next state
x_{k+1}
8 Computation Lens
This page is where three different notions finally need to be separated:
- the continuous model itself
- the exact sampled flow map
- the numerical method used to approximate that map
In practice, that means asking:
- is
ha physical sampling interval, a solver step size, or both? - am I simulating the exact sampled model or only approximating it?
- does my discrete update preserve the qualitative behavior the continuous system should have?
- if I design a controller on the discrete model, how sensitive is that design to the discretization choice?
These are the questions that turn ODEs into engineering systems.
9 Application Lens
9.1 Digital Control
Modern controllers usually measure, compute, and actuate at discrete times, so sampled-data models are the practical bridge from continuous physics to real implementation.
9.2 Simulation And Estimation
Kalman-style filtering, model predictive control, and simulation pipelines all depend on getting the state-update map right enough for repeated use.
9.3 Continuous-Time ML
Neural ODEs, probability-flow ODEs, and continuous-time optimization views only become computable after choosing a discretization and stepping rule.
10 Stop Here For First Pass
If you can now explain:
- what the exact flow map
\Phi_hmeans - how a numerical update differs from the exact sampled dynamics
- how held inputs produce a discrete-time controlled model
- why continuous-time stability does not automatically guarantee stability of an explicit discretization
then this page has done its job.
11 Go Deeper
After this page, the next natural steps are:
- Time-Stepping for ODEs and Stability
- Research Bridges: Reverse-Time SDEs, Probability-Flow ODEs, Flow Matching, and Control
State-Space ModelsControllability and Observability
The strongest adjacent pages are:
12 Optional Deeper Reading After First Pass
The strongest current references connected to this page are:
- MIT 18.03SC Unit IV: First-order Systems - official systems unit connecting matrix methods, phase portraits, and state evolution. Checked
2026-04-25. - MIT 18.03SC Matrix Exponentials - official page for the exact linear-system evolution map and variation-of-parameters viewpoint. Checked
2026-04-25. - MIT 18.086 Lecture 1: Difference Methods for Ordinary Differential Equations - official MIT lecture resource for turning continuous ODEs into stepwise updates. Checked
2026-04-25. - Stanford EE263 bulletin - official course description connecting matrix exponentials, stability, control, reachability, and observability. Checked
2026-04-25. - Stanford EE363 bulletin - official course description for the more advanced linear-dynamical-systems and control-facing arc. Checked
2026-04-25. - Stanford ENGR155A bulletin - official ODE course description connecting analytic methods, numerical methods, and engineering models. Checked
2026-04-25.
13 Sources and Further Reading
- MIT 18.03SC Unit IV: First-order Systems -
First pass- official systems unit anchoring the state-space and sampled-evolution viewpoint. Checked2026-04-25. - MIT 18.03SC Matrix Exponentials -
First pass- official page for the exact evolution operator of linear systems. Checked2026-04-25. - MIT 18.086 Lecture 1: Difference Methods for Ordinary Differential Equations -
First pass- official MIT lecture resource for the finite-difference and update-rule viewpoint on ODEs. Checked2026-04-25. - Stanford EE263 bulletin -
Second pass- official course description linking linear dynamical systems to reachability, observability, and control. Checked2026-04-25. - Stanford EE363 bulletin -
Second pass- official advanced linear-systems course description linking matrix exponential, state transfer, and control. Checked2026-04-25. - Stanford ENGR155A bulletin -
Second pass- official engineering ODE course description linking analytic ODE methods with numerical stepping. Checked2026-04-25.