Functional Representation
A functional representation describes an object not through its stored data, but through a computational rule that delivers the appropriate value on request. So instead of storing a list of finished values, one stores a rule that generates these values.
There are two very different ways to capture something in a computer. One can write down and store all the values individually. Or one can store a computational rule that recalculates each value on demand. The second variant is called a functional representation. A simple example is a circle: one can store millions of pixels, or just the center point, the radius, and the rule for how a circle emerges from them. Both approaches describe the same thing, but require vastly different amounts of space.
Why a rule is often worth more than a table
A table of values is rigid. It knows exactly the points someone entered into it, and nothing in between. A computational rule, by contrast, also delivers a result for intermediate values. That’s why one can zoom in arbitrarily far on a functional representation without the image becoming pixelated.
Then there’s the space advantage. A high-resolution 3D scene as a point cloud can take up many gigabytes. The same scene as a computational rule sometimes fits into just a few megabytes. The price for this is computing time: every value must first be generated upon retrieval. So one trades storage space for computational effort, and which side is cheaper depends on the use case.
In AI research, the term is also important because a trained model is itself a functional representation. It has not stored the training data as a copy. It has learned a function that computes a matching output for a given input. This explains why a language model can answer questions that never appeared in the training data in that exact form.
From the list of points to the learned function
Technically, a function is simply a mapping: you put in an input and get out an output. In an image scene, the input might be a location in space, the output a color. The art lies in finding a function that delivers the correct outputs for all desired inputs.
In the past, people wrote such functions by hand as formulas. That works well for circles and curves, but not for a face or a patch of forest. This is where machine learning comes in: a neural network, i.e. a computational model with many adjustable numbers, is fitted to examples. It learns the function instead of being handed it in advance.
These adjustable numbers are called parameters. They are the actual stored content of a learned functional representation. A well-known example is neural radiance fields, or NeRF for short: there, a small network memorizes an entire 3D scene and can use it to compute new viewing angles that were never photographed. A common misconception is that the model has stored the photos. It has only learned a rule that fits them.
Where this way of thinking shows up in products
It is most visible in fonts. Letters on a screen are not images, but curve descriptions. That’s why they stay sharp whether you set the font size to 8 or 200. Vector graphics in logos and map apps use the same principle.
In the 3D world, the term comes up in games, film effects, and photogrammetry. Companies like Nvidia or Apple advertise reconstructing real spaces from smartphone footage. The scene exists as a learned function, not as a huge file full of measurement points. You regularly come across such terms in reports on augmented reality.
And finally, the idea affects every discussion about language models. When there’s a debate over whether a model has memorized texts verbatim or recognizes patterns, this is exactly the distinction at stake. A database retrieves stored entries. A functional representation computes a new answer. The difference seems technical, but it decides copyright questions and how reliably a system responds.