Quantum clustering

Content sourced from Wikipedia, licensed under CC BY-SA 3.0.

Quantum clustering (QC) is a set of clustering methods that borrow ideas from quantum physics to find groups in data. Like other density-based methods, QC looks for regions in the data where points are concentrated.

How QC works
- Imagine each data point as the center of a small Gaussian blob. All these blobs are added together to make one overall density, which behaves like a quantum wave function for the data.
- From this wave function, a quantum potential is built. The landscape of this potential has valleys in places where data are dense.
- Points then move downhill in this landscape. As they slide toward the low spots, nearby points gather together, revealing clusters.

The main knob you tune in QC is sigma, the width of the Gaussian blobs. If sigma is very small, every point makes its own little dent in the landscape, so nothing clusters. If sigma is very large, the landscape becomes a single big bowl and all points end up in one cluster. By trying different sigma values, you can see different levels of structure: small sigma shows fine detail, large sigma shows the bigger picture.

Dynamic Quantum Clustering (DQC)
- DQC, developed in 2009, keeps the same quantum landscape but changes how points move. Instead of simple downhill movement, each point evolves over time according to the time-dependent Schrödinger equation.
- Each point stays represented by its own Gaussian wave function, and its evolution depends on the whole landscape. This nonlocal evolution means points can feel distant parts of the landscape, not just the slope right under them.
- Because of this, points can “tunnel” through barriers and avoid getting stuck in many small, unhelpful spots. DQC uses two extra knobs: the time step and the mass of each point (which controls tunneling). Often reasonable defaults work well.

Computational challenges and a practical solution
- If you treat every point with full quantum evolution, the calculations can be very slow for large data sets.
- A common fix is to use a smaller subset of points as a basis. Choose b points that spread out through the data space (the basis). Build b quantum states from these basis points and use them to describe all other points. This makes the problem tractable, though accuracy depends on sigma and the chosen basis size.
- In practice, with modern computing, a few thousand basis points are often manageable.

What you can see and how to use it
- DQC can produce trajectories showing how each point moves over time. When you visualize these paths (often in a 3D view built from the first few principal components), you can see channels that guide points toward clusters, like riverbeds feeding lakes.
- These channels can reveal subclusters or patterns related to extra information (metadata) about the data.

Applications and takeaways
- QC and DQC have been applied in biology, geology, physics, finance, engineering, and economics.
- The methods give a flexible way to explore structure at different scales by changing sigma, and they make it possible to see dynamic movement toward clusters through visualization of point trajectories.
- The underlying math also supports analysis of the quantum potential that shapes the clustering landscape.