Curse of Dimensionality
The family of ways intuition breaks in high dimensions — data becomes sparse, distances stop being meaningful, and volume hides in the corners.
Reviewed July 17, 2026Stable
When not to use it
- As a blanket reason to avoid high-dimensional models. Deep learning works in vast dimensions by exploiting low-dimensional structure.
- In genuinely low-dimensional problems. Below a handful of dimensions the effects are negligible and the warning is noise.
- As the explanation when the real issue is sample size or label quality. The curse is about dimension specifically, not data problems generally.
Reach for something else instead
- Dimensionality reduction (PCA, UMAP) directly attacks the curse by finding the low-dimensional structure.
- Feature selection avoids adding harmful dimensions in the first place.
- Manifold-learning methods model the low-dimensional surface the data actually lies on.
Sources & further reading
- Bellman (1957), Dynamic Programming — coined "curse of dimensionality" for the exponential blow-up of state spaces.
- Beyer et al. (1999), When Is "Nearest Neighbor" Meaningful? — formalised distance concentration in high dimensions.
- Goodfellow, Bengio & Courville, Deep Learning — the manifold hypothesis and why deep networks escape the worst of the curse.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Believing more features always help. Past a point they hurt, because the space outgrows the data.
- Concluding high dimensions are always hopeless. The manifold hypothesis is why deep learning works despite them.
- Blaming the curse for what's really too little data or noisy labels. It's a dimension-specific effect, not a catch-all.
At a glance
FieldMachine Learning
Coined byRichard Bellman, 1957
Symptomssparsity, distance concentration, volume in the shell
Escape routethe manifold hypothesis
DifficultyIntermediate