jeremistadler / tdigest

tdigest: javascript implementation of Dunning's T-Digest for streaming quantile approximation

This is a fork of welch/tdigest using ESM

Install

npm install githup:rocicorp/tdigest

tdigest

Javascript implementation of Dunning's T-Digest for streaming quantile approximation

The T-Digest is a data structure and algorithm for constructing an approximate distribution for a collection of real numbers presented as a stream. The algorithm makes no guarantees, but behaves well enough in practice that implementations have been included in Apache Mahout and ElasticSearch for computing summaries and approximate order statistics over a stream.

For an overview of T-Digest's behavior, see Davidson-Pilon's blog post regarding a python implementation. For more details, there are the tdigest paper and reference implementation (Java). This javascript implementation is based on a reading of the paper, with some boundary and performance tweaks.

changes in 0.1.1:

1. percentile on an empty digest returns undefined or array of undefined instead of NaN

2. upgraded bintrees to get bugfix.

3. bugfix for discrete percentile and p_rank, make boundary conditions conform to standard definition.

changes in 0.1.0:

Discrete mode: when a TDigest is created with delta=false, the sample distribution is treated as discrete. TDigest behavior is disabled, differing samples are never merged (they needn't even be numeric), and percentiles are reported as nearest exact data values rather than interpolated.

Digest: distribution digest structure. Starts in exact histogram (discrete) mode, remains in exact mode for reasonable numbers of distinct values as sample size inreases, and automatically switches to TDigest mode for large samples that appear to be from a continuous distribution.

Renamed quantile() -> p_rank(), Percentile Rank.

percentile() and p_rank() now accept arrays or singleton arguments.

changes in 0.0.7:

A grunt dist task has been added to create a UMD-wrapped version of tdigest and dependencies for importing as a standalone module in client-side javascript.

bugfixes and speed improvements.

changes in 0.0.5:

API Overhaul:

• asArray() -> toArray()
• redigest() -> compress()
• digest() -> push()
• pushing an array no longer triggers compression

bugfixes and speed improvements.

quickstart

node.js:

npm install tdigest

import {TDigest} from 'tdigest';
let x = [],
  N = 100000;
for (let i = 0; i < N; i += 1) {
  x.push(Math.random() * 10 - 5);
}
td = new TDigest();
td.push(x);
td.compress();
console.log(td.summary());
console.log('median ~ ' + td.percentile(0.5));

See also example.js in this package.

In the browser:

The grunt dist task has been configured to generate a self-contained UMD-wrapped version of tdigest in dist/tdigest.js.

Embed it in HTML like this:

<script src="dist/tdigest.js"></script>
<script type="module">
    import {TDigest} from './tdigest.js';

    var td = new TDigest();
    for (var i=0; i < 1000000; i++) {
        td.push(Math.random());
    }
    td.compress();
    document.write(td.summary())
</script>

See also example.html in this package.

dependencies

bintrees: https://www.npmjs.com/package/bintrees