SPB Git

Repository Public

stats.ts 3.6 KB 78 lines typescript

1/** Small numerical helpers shared by the simulations. Pure functions on plain arrays. */23export function sum(xs: readonly number[]): number { let s = 0; for (const x of xs) s += x; return s; }4export function mean(xs: readonly number[]): number { return xs.length ? sum(xs) / xs.length : NaN; }56/** Sample variance (divisor n − 1). NaN below two observations. */7export function variance(xs: readonly number[]): number {8  const n = xs.length;9  if (n < 2) return NaN;10  const m = mean(xs);11  let s = 0;12  for (const x of xs) s += (x - m) * (x - m);13  return s / (n - 1);14}15export function sd(xs: readonly number[]): number { return Math.sqrt(variance(xs)); }1617/** Quantile by linear interpolation of the order statistics (R type 7, the default of R and NumPy). */18export function quantile(xs: readonly number[], p: number): number {19  if (!xs.length) return NaN;20  const s = [...xs].sort((a, b) => a - b);21  const h = (s.length - 1) * Math.min(1, Math.max(0, p));22  const lo = Math.floor(h), hi = Math.ceil(h);23  return s[lo] + (h - lo) * (s[hi] - s[lo]);24}2526export interface OlsFit {27  n: number; intercept: number; slope: number;28  /** Residual standard error, sqrt(SSR / (n − 2)). */29  s: number; r2: number; xBar: number; sxx: number; seIntercept: number; seSlope: number;30}31/** Simple least squares y = a + b·x. */32export function ols(x: readonly number[], y: readonly number[]): OlsFit {33  const n = Math.min(x.length, y.length);34  const xBar = mean(x.slice(0, n)), yBar = mean(y.slice(0, n));35  let sxx = 0, sxy = 0, syy = 0;36  for (let i = 0; i < n; i++) { const dx = x[i] - xBar, dy = y[i] - yBar; sxx += dx * dx; sxy += dx * dy; syy += dy * dy; }37  const slope = sxx > 0 ? sxy / sxx : NaN;38  const intercept = yBar - slope * xBar;39  let ssr = 0;40  for (let i = 0; i < n; i++) { const e = y[i] - intercept - slope * x[i]; ssr += e * e; }41  const s = n > 2 ? Math.sqrt(ssr / (n - 2)) : NaN;42  const r2 = syy > 0 ? Math.max(0, 1 - ssr / syy) : 1;43  return { n, intercept, slope, s, r2, xBar, sxx, seSlope: s / Math.sqrt(sxx), seIntercept: s * Math.sqrt(1 / n + (xBar * xBar) / sxx) };44}4546/** Pearson correlation coefficient. */47export function pearson(x: readonly number[], y: readonly number[]): number {48  const n = Math.min(x.length, y.length);49  const xBar = mean(x.slice(0, n)), yBar = mean(y.slice(0, n));50  let sxx = 0, sxy = 0, syy = 0;51  for (let i = 0; i < n; i++) { const dx = x[i] - xBar, dy = y[i] - yBar; sxx += dx * dx; sxy += dx * dy; syy += dy * dy; }52  return sxy / Math.sqrt(sxx * syy);53}5455/** Softmax of logits / temperature, computed stably (the maximum is subtracted first). −Infinity logits get 0. */56export function softmax(logits: readonly number[], temperature = 1): number[] {57  const t = Math.max(temperature, 1e-9);58  let m = -Infinity;59  for (const l of logits) if (l > m) m = l;60  const e = logits.map((l) => Math.exp((l - m) / t));61  const z = sum(e);62  return e.map((v) => v / z);63}6465/** Indices of the k largest values, largest first; ties go to the lower index. */66export function topK(values: readonly number[], k: number): number[] {67  return values.map((v, i) => [v, i] as const).sort((a, b) => b[0] - a[0] || a[1] - b[1]).slice(0, Math.max(0, Math.floor(k))).map((p) => p[1]);68}6970/** Shannon entropy in nats; 0·log 0 = 0. */71export function entropy(p: readonly number[]): number { let h = 0; for (const v of p) if (v > 0) h -= v * Math.log(v); return h; }7273/** n evenly spaced values from a to b inclusive. */74export function linspace(a: number, b: number, n: number): number[] {75  if (n <= 1) return [a];76  return Array.from({ length: n }, (_, i) => a + ((b - a) * i) / (n - 1));77}78
Files main
Menu

Theme