Understanding Inverse Problems

Harvard EPS55

Explore the fascinating world of working backwards from effects to causes

What are Inverse Problems?

An inverse problem is when we have observed data and need to determine the model parameters that produced it. Unlike forward problems where we know the parameters and predict the observations, inverse problems require us to work backwards from data to parameters.

Forward Problem

Model Parameters → Forward Model → Observed Data

Example: Drop height/angle → Physics → Landing position

⬇️

Known: Model parameters (θ)
Find: Observed data (d)

Inverse Problem

Observed Data → Inverse Model → Model Parameters

Example: Landing position → Physics → Drop height/angle

⬆️

Known: Observed data (d)
Find: Model parameters (θ)

Mathematical Formulation

Forward Problem

d = G(θ) + ε

d: Observed data (measurements)
G: Forward model (physics/process)
θ: Model parameters (unknowns)
ε: Measurement noise/errors

Inverse Problem

θ_est = argmin ||G(θ) - d||² + λR(θ)

θ_est: Estimated parameters
||G(θ) - d||²: Data misfit (how well parameters explain data)
λ: Regularization parameter
R(θ): Regularization term (prior knowledge/constraints)

Real-World Examples

🏥

Medical Imaging

Data: X-ray measurements
Parameters: Tissue densities

🌍

Seismology

Data: Seismic waves
Parameters: Rock velocities

🔍

Computer Vision

Data: 2D images
Parameters: 3D structure

🌡️

Heat Transfer

Data: Temperature now
Parameters: Initial temp

Interactive Example: Parameter Estimation from Noisy Data

Observe how we estimate model parameters (signal frequency and amplitude) from noisy observed data

True Model Parameters

Frequency: 0.02, Amplitude: 50

Estimated Parameters

Calculating...

General Approach to Solving Inverse Problems

1

Model the Forward Problem

Understand how model parameters produce observed data. Develop mathematical model: d = G(θ) where d is data and θ are parameters

2

Formulate the Inverse

Set up equation to find parameters θ given data d. Often involves solving: θ = G⁻¹(d)

3

Address Ill-Posedness

Handle non-uniqueness (multiple parameters fit data), instability (small data changes cause large parameter changes)

4

Apply Regularization

Add constraints or prior knowledge about parameters to stabilize the solution

5

Validate Results

Check if estimated parameters produce observed data when run through forward model

Key Challenges

⚠️

Non-uniqueness: Multiple causes can produce the same effect

📊

Noise sensitivity: Small measurement errors cause large solution errors

💻

Computational cost: Solutions often require iterative algorithms

🎯

Incomplete data: Limited measurements constrain reconstruction quality

Solution Methods

📐

Least Squares

Minimize ||Ax - b||²

🎛️

Tikhonov Regularization

Add penalty term λ||x||²

🎲

Bayesian Methods

Use probability and priors

🔄

Iterative Methods

Gradual refinement