Discretize Continuous State Space Model

hardPROVision MambaMath, Signal ProcessingAdvanced Architectures

Implement a discretization method for State Space Models originally defined in continuous time to accommodate discrete token sequences. This process is crucial for applying continuous-time models to real-world problems that involve discrete data.

The Continuous State Space Model is defined by the equations $x'(t) = Ax(t) + Bu(t)$ and $y(t) = Cx(t)$ , where $x(t)$ is the state vector, $u(t)$ is the input vector, and $y(t)$ is the output vector. To discretize this model, we can use the Zero-Order Hold method, which assumes the input is constant between sampling times.

Here are the steps to discretize the model:

Define the continuous-time model parameters $A$ , $B$ , and the sampling time $\Delta$ .
Compute the discretized state transition matrix $\bar{A}$ and input coefficient $\bar{B}$ .

\bar{A} = e^{A \cdot \Delta}

\bar{B} = \frac{e^{A \cdot \Delta} - 1}{A} \cdot B

When $A = 0$ the fraction for $\bar{B}$ is undefined; use its limit instead, $\bar{B} = B \cdot \Delta$ (and $\bar{A} = 1$ ).

Return the tuple $(\bar{A}, \bar{B})$ with both values rounded to 4 decimal places.

This technique is widely used in digital control systems.

Example

Input

A=-1.0, B=1.0, delta=0.1

Output

(0.9048, 0.0952)

Reasoning

Inputs: A $=-1$ is the continuous decay rate of the state, B $=1$ scales how strongly the input pushes the state, and delta $\Delta = 0.1$ is the time between two tokens.
State transition: over one step the state decays by the factor $\bar{A} = e^{A\Delta} = e^{-0.1} \approx 0.904837$ , which rounds to $0.9048$ .
Input coefficient: with the input held constant for the whole step, $\bar{B} = \frac{e^{-0.1} - 1}{-1} \cdot 1 = 1 - 0.904837 = 0.095163$ , which rounds to $0.0952$ .
Output: the pair (0.9048, 0.0952), so one step of the discrete model reads $x_k = 0.9048\,x_{k-1} + 0.0952\,u_k$ : about 9.5% of the old state fades and roughly that much input is absorbed.
Notice: $\bar{B}$ is close to $B\Delta = 0.1$ . For small $A\Delta$ the fraction $\frac{e^{A\Delta}-1}{A}$ approaches $\Delta$ , and that limit is exactly the value $\bar{B}/B$ takes when $A = 0$ .

Constraints

$A$ : Continuous state transition (scalar), $-10 \leq A \leq 0$
$B$ : Input coefficient (scalar), $0 \leq B \leq 10$
$\Delta$ : Discretization step size, $0.01 \leq \Delta \leq 2.0$

More from Advanced Architectures

Complete Mixer Block with Residualshard Channel Mixing with GELUmedium Token Mixing Layermedium Bidirectional Scan for Visionmedium State Space Model vs Attention Complexitymedium

Concept primerPREVIEW

Discretize Continuous State Space Model

Background Knowledge

State Space Models originate from continuous-time dynamical systems in control theory. To use them for discrete sequences (text, images), we must discretize the continuous equations.

Continuous-Time SSM

The continuous system is defined by:

dh/dt = Ah(t) + Bx(t) # State derivative
y(t) = Ch(t) # Output

Where:

$h(t) \in \mathbb{R}^S$ is the hidden state
$A \in \mathbb{R}^{S \times S}$ is the state transition matrix
$B \in \mathbb{R}^{S \times 1}$ is the input projection
$C \in \mathbb{R}^{1 \times S}$ is the output projection

Zero-Order Hold (ZOH) Discretization

The Zero-Order Hold method assumes the input $x(t)$ is constant between samples. This yields:

Ā = exp(ΔA)
B̄ = (exp(ΔA) - I) A⁻¹ B = (Ā - I) A⁻¹ B

PRO

Continue the full explanation

You're reading the free preview. Unlock the complete walkthrough, the code editor, test runner and reference solution with Premium.

PRO

Editor locked

The code editor is locked for Pro problems. It is only available for free problems. Please upgrade to gain access to the code editor for all problems.

solution.py

Python 3.13

Test results