Point Spread Function

No image is perfectly sharp. If you photograph a very small, bright point – say, a distant star – it never appears in the image as a single point. Instead, it is smeared into a small spot with soft edges. The point spread function is the precise description of this spot: it states how much light from the point ends up at which location in the image. You can think of it as a smearing pattern that a device imprints on every single point of light. Because an entire image consists of countless such points, this pattern determines the sharpness of the whole image.

The fingerprint of a lens

The point spread function is the most honest metric for the quality of an optical system. An expensive lens produces a tight, compact spot. A cheap or improperly focused one produces a wide, ragged one. Anyone wanting to compare lenses, microscopes, or telescopes therefore looks at this function and not just at the number of pixels.

It is even more important because it can be reversed. If you know how a device smears light, you can computationally undo part of the blur. This procedure is called deconvolution. In the 1990s, astronomers used it to rescue images from the Hubble Space Telescope, whose mirror had a grinding error. The images were blurry, but the error was known and could therefore be partially calculated out.

There is, however, a hard limit. Very fine details vanish completely into image noise during smearing – that is, into the random disturbances present in every sensor. This information is gone and cannot be recovered by any calculation. Anyone claiming the opposite is confusing reconstruction with invention – a point that is frequently glossed over in AI-based image enhancement.

Smearing as a computational operation

Mathematically, a real image arises by convolving the ideal subject with the point spread function. Convolution here means: at the location of each individual image point, you place the smearing pattern, weight it by the brightness of that point, and add everything up. The result is the blurred image that the sensor actually records. This computational operation is the same one that also underlies neural networks used for image recognition.

The shape of the spot has several causes. One of them cannot be eliminated: light is a wave and is diffracted at the round aperture of a lens. Even a flawlessly manufactured telescope therefore produces a tiny disc with faint rings around it, the so-called diffraction disc. On top of this come avoidable causes: errors in lens grinding, incorrect focus, camera movement, or turbulent air in the atmosphere.

The function can be determined in two ways. One can measure it directly by capturing an object that is as point-like as possible – in astronomy, simply a faint star; in microscopy, a tiny fluorescent bead. Or one can calculate it from the known design of the optics. Importantly, the function is rarely the same across the entire image. At the edges of the image, the spot is usually larger and more distorted than in the center.

From the Hubble Telescope to the smartphone camera

In everyday life, this principle is encountered with every smartphone photo. The lenses are tiny and optically limited, and the software compensates for this. Part of this computation uses knowledge of exactly how this particular camera model smears points. Even the artificially generated background blur effect in portrait mode is nothing other than a deliberately applied, very wide point spread function.

In research, the term is standard vocabulary. Fluorescence microscopy in biology constantly works with deconvolution to make cell structures visible. In astronomy, papers on new telescopes regularly report the measured point spread function. Techniques for high-resolution microscopy, which earned the 2014 Nobel Prize in Chemistry, are also based on exploiting the known shape of the spot.

In AI news, the topic comes up indirectly. Models that sharpen blurry images are often trained using images that have been artificially blurred. To do this, a simulated point spread function is applied to sharp source images. If this simulation does not match the real camera, the model fails in practice – a common reason why impressive demos disappoint in everyday use.