core Estimated learning time: 9 h

PE2-2.3 Stochastic Modelling & Compression

You can build AR-1 and Markov chain models and solve the Yule-Walker equations, and apply correlation, predictive coding, transform domain compression and multiresolution coding to images.

The Yule-Walker equations are the practical payoff of AR modelling: measured autocorrelations in, model coefficients out, by solving a linear system. Predictive coding is the idea underneath most compression — send the error of a prediction rather than the value, because the error is smaller and less correlated. Transform and multiresolution coding are where JPEG and its successors come from, and seeing them derived rather than described makes image compression stop being a black box.

Work through these

  • Stochastic models; the AR-1 process

    The simplest model where the present depends on its own past, simulated rather than derived. Seeing the sample paths makes the parameter's meaning obvious.

  • Yule-Walker equations

    The equations that recover the model parameters from the measured correlation. It is the practical estimation method and it is a linear system, which is why this model dominates.

  • Markov chains

    Processes that move between states with fixed probabilities. They reappear in the speech and coding material and are worth solid attention.

  • Analog to digital conversion

    Turning a continuous signal into numbers, revisited here as a source coding problem rather than a circuit one. The quantisation error becomes a distortion measure.

  • K means

    Grouping data by nearness, used here to design a quantiser. It is the same algorithm as in machine learning, which is a connection worth noticing.

  • Correlation and predictive coding

    Using the correlation from earlier to predict the next sample and send only the difference. It is the whole idea behind the compression that follows.

  • Image compression

    Applying the ideas above to images, where the correlation is two-dimensional. It is the first real application of the course's machinery.

  • Transform domain compression

    Compressing in a transformed domain instead, which concentrates the energy into fewer coefficients. This is what actual image standards do.

  • Multi resolution coding

    Coding at several scales at once, which allows progressive transmission. It closes the compression sequence and connects to wavelet ideas.

Sign in to keep your progress.

Free resources

Links last checked 30 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.