liveBook · Manning

Approximation error

As explained in Advanced Algorithms and Data Structures

Approximation error is a nonnegative factor that bounds how far an approximate nearest-neighbor result may be from the corresponding exact result.

Approximation error controls the accuracy–search-cost trade-off in approximate nearest-neighbor search. A parameter ϵ, where 0.0 ≤ ϵ ≤ 0.5, permits early termination while bounding the returned distance. section 10.4.3

For a query point P:

  • If the true nearest neighbor N is at distance d, the returned point Q is guaranteed to be no farther than (1+ϵ)d. section 10.4.3
  • Prune a branch only when its closest possible point is farther than the current nearest-neighbor distance divided by 1+ϵ. section 10.4.3
  • Apply the same guarantee to n-nearest-neighbor search by measuring the distance of the n-th neighbor. section 10.4.3

Missing the true neighbor

With ϵ approximately 0.25, a traversal can prune the branch containing the true nearest neighbor J when J lies outside the permitted approximate query region, returning a closer-traversed alternative instead.

How Advanced Algorithms and Data Structures covers it

How deep it goes
Defines the distance guarantee and derives the pruning threshold, but does not provide a formal proof.
Examples
Uses query points P and Z, candidates Q and N, pruning bounds, and a traversal with ϵ approximately 0.25.
What it assumes
Assumes nearest-neighbor search, query-to-point distance, branch traversal, and bounding envelopes.
What you can do after
Apply it via branch-pruning rules that guarantee an approximate nearest-neighbor distance within a factor of the exact result.
How it connects
Connects approximate nearest-neighbor search with early termination, branch pruning, bounding envelopes, and n-nearest-neighbor search.

Also explained in

Compare how 2 books explain approximation error →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Advanced Algorithms and Data Structures. The section links point to the parts of the book it draws on.