Curse of the dimension

• Humans have difficulty imagining beyond 2-4 dimensions.

• A lot of unexpected phenomena happen when you get dimensional.

• In higher dimensional space, almost all points are far from the center

  • A point within distance 1 from the origin in one dimension is half of a point within distance 2
  • 1/4 in 2 dimensions
  • 1/8 in 3D
  • ...and the "percentage of close points" decreases exponentially as the dimension increases.

• The number of samples required for sampling increases exponentially.

  • In the case of machine learning, increasing the dimensionality of the machine deteriorates the accuracy.
  • Because the effect of insufficient sample size is more overwhelming than the improvement in accuracy due to the additional dimension.

Chi-square distribution - Wikipedia

• For 3 or more dimensions, the vector length mode is non-zero.

  • Condition that each axis follows a standard normal distribution with mode 0

    • chi-square distribution

  • This is related to "most points are far from the center."

  • Almost all vectors are orthogonal

  • [How similar are vectors with high cosine similarity (from Iwanami Data Science publication event) - Mi manca qualche giovedi`?](http://d.hatena.ne.jp/n_shuyo/20160401/cosine_similarity)

  • 1000000 If you want to find the percentage of samples with a cosine similarity greater than 1/2,

  • 0.06 (about 1/17) in 10 dimensions,

  • 0.01 (about 1/100) in 20 dimensions,

  • 0.0021 (about 1/480) in 30 dimensions,

  • 0.00042 (about 1/2400) in 40 dimensions

  • 100 In 100 dimensions, there were no points in the 10,000,000 sampled points where the cosine similarity was greater than 1/2.

  • Of course, in two dimensions, 33%.

  • Cosine similarity of 0.2 is extremely rare in higher dimensions

• relevance

  • https://twitter.com/nishio/status/1258610796969340928?s=21

    • diversity

  • If you take two random vectors in a high-dimensional space, the probability that they are nearly the same direction is very small compared to the probability that they are nearly orthogonal

  • As the number of dimensions (number of evaluation axes) increases, the probability of a state of complete superiority of one person's skills over another's decreases.

    • 100% in 1D, 50% in 2D, 25% in 3D
    • 

  • In high-dimensional space, the normal distribution is almost uniformly distributed on the hypersphere

• Almost every stop is a [saddle point

  • 99.8% in 10 dimensions

• There are few cases where only one particular axis is larger than the other.

• blind spot card 19

This page is auto-translated from /nishio/次元の呪い using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.