Global

Members

(constant) DEFAULT_MAX_TIME_MS

Description:

• Default maximum fitting time: one hour, expressed in milliseconds (Rust uses chrono::Duration::seconds(3600)).

Source:

• kriging-friedrich/gaussian_process.mjs, line 68

(constant) Gaussian

Description:

• Gaussian kernel — alias of SquaredExp (Rust pub type Gaussian = SquaredExp).

Source:

• kriging-friedrich/kernels.mjs, line 294

Methods

addKernels(k1, k2) → {KernelSum}

Description:

• Convenience constructor for k1 + k2 (Rust KernelArith Add operator).

Source:

• kriging-friedrich/kernels.mjs, line 965

Parameters:

Returns:

Type

KernelSum

addRowsCholeskyCovMatrix(cholesky, allInputs, nbNewInputs, kernel, diagonalNoise)

Description:

• Incrementally add the last nbNewInputs rows of allInputs to an existing Cholesky decomposition (in place), one row at a time. For each new row:

  • compute the new column (covariances with all rows up to and including itself, length col_index+1),
  • add noise² to the final (self) entry,
  • call cholesky.insertColumn(col_index, column).

Source:

• kriging-friedrich/algebra.mjs, line 789

Parameters:

addToDiagonal(A, value) → {Array.<Array.<number>>}

Description:

• Return a copy of A with value added to its diagonal (Aᵢᵢ += value). Does not mutate the input.

Source:

• kriging-friedrich/algebra.mjs, line 436

Parameters:

Returns:

Type

Array.<Array.<number>>

addVec(a, b) → {Array.<number>}

Description:

• Element-wise addition a + b.

Source:

• kriging-friedrich/algebra.mjs, line 51

Parameters:

Returns:

Type

Array.<number>

assert(condition, message)

Source:

• kriging-friedrich/test.mjs, line 30

Parameters:

cholesky(A, epsilonopt) → {Array.<Array.<number>>}

Description:

• Cholesky decomposition of a symmetric positive-definite matrix A(n×n), returning the lower-triangular factor L such that A = L·Lᵀ.

  Failure (non-positive-definite) behaviour:

  • epsilon === null → throw Error (matches Rust .expect() panic).
  • epsilon a finite positive number → mimic nalgebra Cholesky::new_with_substitute: whenever a diagonal pivot dⱼ ≤ 0, substitute epsilon for that pivot and keep going. If even the substitute fails to keep things real (cannot happen for epsilon > 0), throw.

  Reads only the lower triangle of A (the upper triangle is ignored), matching nalgebra's behaviour of using only the lower-triangular part.

Source:

• kriging-friedrich/algebra.mjs, line 273

Parameters:

Returns:

Type

Array.<Array.<number>>

choleskyInverse(L) → {Array.<Array.<number>>}

Description:

• Compute A⁻¹ = (L·Lᵀ)⁻¹ from the lower-triangular Cholesky factor L, by solving A·X = I. Matches nalgebra Cholesky::inverse.

Source:

• kriging-friedrich/algebra.mjs, line 420

Parameters:

Returns:

Type

Array.<Array.<number>>

choleskySolve(L, B) → {Array.<number>|Array.<Array.<number>>}

Description:

• Solve A·X = B given the lower-triangular Cholesky factor L (A = L·Lᵀ): first solve L·Y = B (forward), then Lᵀ·X = Y (back). Matches nalgebra Cholesky::solve.

Source:

• kriging-friedrich/algebra.mjs, line 407

Parameters:

Returns:

Type

Array.<number> | Array.<Array.<number>>

classifyFeature(feature, result)

Description:

• 將單一 Feature 依幾何類型分類到結果物件中

Source:

• splitAndProcGeoJSON.mjs, line 79

Parameters:

columnDot(A, B, c) → {number}

Description:

• Dot product of column c of A with column c of B (both number[][]).

Source:

• kriging-friedrich/gaussian_process.mjs, line 788

Parameters:

Returns:

Type

number

columnNormSquared(M, c) → {number}

Description:

• Squared Euclidean norm of column c of matrix M (number[][]).

Source:

• kriging-friedrich/gaussian_process.mjs, line 772

Parameters:

Returns:

Type

number

createEmptyResult() → {Object}

Description:

• 建立空的分類結果物件

Source:

• splitAndProcGeoJSON.mjs, line 19

Returns:

Type

Object

diagonal(A) → {Array.<number>}

Description:

• Diagonal of a square (or rectangular) matrix.

Source:

• kriging-friedrich/algebra.mjs, line 231

Parameters:

Returns:

Type

Array.<number>

dot(a, b) → {number}

Description:

• Dot product Σ aᵢ·bᵢ.

Source:

• kriging-friedrich/algebra.mjs, line 27

Parameters:

Returns:

Type

number

ensureRingClosed(ring) → {Array}

Description:

• 修復 Polygon ring 閉合（首尾座標相同）

Source:

• splitAndProcGeoJSON.mjs, line 34

Parameters:

Returns:

Type

Array

fitAmplitudeVar(outputs) → {number}

Description:

• Best-guess amplitude = variance of the outputs.

  Matches nalgebra variance(): Σ(yᵢ − mean)² / N (divides by N, the sample count — NOT N−1). Confirmed from nalgebra 0.34.2 statistics.rs (PORT_SPEC §4.1).

Source:

• kriging-friedrich/kernels.mjs, line 58

Parameters:

Returns:

Type

number

fitBandwidthMean(inputs) → {number}

Description:

• Rough bandwidth estimate: mean Euclidean distance between distinct samples. Σ_{i<j} ‖xᵢ − xⱼ‖ / (n(n−1)/2)

Source:

• kriging-friedrich/kernels.mjs, line 36

Parameters:

Returns:

Type

number

fixCloseMultiPolygonCoords(coordinates) → {Array}

Description:

• 修復 MultiPolygon 的所有 ring 閉合

Source:

• splitAndProcGeoJSON.mjs, line 68

Parameters:

Returns:

Type

Array

fixClosePolygonCoords(coordinates) → {Array}

Description:

• 修復 Polygon 的所有 ring 閉合

Source:

• splitAndProcGeoJSON.mjs, line 57

Parameters:

Returns:

Type

Array

fromVector(v, single) → {number|Array.<number>}

Description:

• Convert an internal result vector back to the user-facing output type.

  single	Result	Rust analogue (from_dvector)
  true	v[0] (number)	Vec<f64> impl: assert_eq!(nrows,1); v[0]
  false	v.slice()	Vec<Vec<f64>> impl: v.iter().cloned().collect()

  Note: predictCovariance always returns number[][] and must NOT call this function (PORT_SPEC §9, §6.7).

Source:

• kriging-friedrich/conversion.mjs, line 126

Parameters:

Returns:

Type

number | Array.<number>

gradientMarginalLikelihood(gp) → {Array.<number>}

Description:

• Computes the gradient of the marginal likelihood for the current value of each parameter. The produced vector contains the gradient per kernel parameter followed by the gradient for the noise parameter.

  Per-parameter formula (optimizer.rs:24): ½ ( alphaᵀ · dp · alpha − trace(K⁻¹ · dp) ) where K = cov(train, train) alpha = K⁻¹ · output dp = ∂K/∂parameter

  Noise gradient (gradient(K, noise) = 2·noise·Id): noise · ( alpha·alpha − trace(K⁻¹) )

Source:

• kriging-friedrich/optimizer.mjs, line 55

Parameters:

Returns:

Type

Array.<number>

hypot(a, b) → {number}

Description:

• Numerically stable sqrt(a² + b²) (matches Rust f64::hypot).

Source:

• kriging-friedrich/algebra.mjs, line 95

Parameters:

Returns:

Type

number

identity(n) → {Array.<Array.<number>>}

Description:

• n×n identity matrix.

Source:

• kriging-friedrich/algebra.mjs, line 204

Parameters:

Returns:

Type

Array.<Array.<number>>

isSingle(input) → {boolean}

Description:

• Detect whether input represents a single multidimensional point.

  Rules (mirrors Rust trait dispatch):

  • number[] (all elements are numbers) → true (single point, like Vec<f64>)
  • number[][] (first element is an array) → false (multiple points, like Vec<Vec<f64>>)

  An empty array is treated as multi-point (falsy single, matches the assert_ne!(nb_rows, 0) guard in the Rust Vec<Vec<f64>> impl — empty inputs are rejected upstream anyway).

Source:

• kriging-friedrich/conversion.mjs, line 47

Parameters:

Returns:

Type

boolean

lstsqSolve(A, b) → {Array.<number>}

Description:

• Least-squares solve of min ‖A·x − b‖ (used by LinearPrior.fit). A may be non-square (m×n). Implemented via the normal equations (AᵀA)·x = Aᵀb solved by Cholesky; falls back to a regularized solve if AᵀA is (numerically) singular, approximating SVD-with-threshold-0 well enough for the LinearPrior integration tests.

  Rust uses SVD solve (threshold 0); PORT_SPEC §12.12 permits the normal- equations approximation here.

Source:

• kriging-friedrich/algebra.mjs, line 457

Parameters:

Returns:

Type

Array.<number>

makeCholeskyCovMatrix(inputs, kernel, diagonalNoise, epsilonopt) → {CholeskyDecomposition}

Description:

• Covariance matrix of inputs with diagonal noise², returned as its CholeskyDecomposition.

  cov[i][j] = kernel.kernel(inputs[i], inputs[j]) cov[i][i] += diagonalNoise² (note: the SQUARE of the noise)

  epsilon is forwarded to the Cholesky (null → throw on failure; positive → substitute mode).

Source:

• kriging-friedrich/algebra.mjs, line 761

Parameters:

Returns:

Type

CholeskyDecomposition

makeCovarianceMatrix(m1, m2, kernel) → {Array.<Array.<number>>}

Description:

• Covariance matrix between rows of m1 and rows of m2 using kernel. out[r][c] = kernel.kernel(m1[r], m2[c]). Shape (m1.length × m2.length).

Source:

• kriging-friedrich/algebra.mjs, line 732

Parameters:

Returns:

Type

Array.<Array.<number>>

makeGp() → {GaussianProcess}

Description:

• Helper: build the same simple 1-D GP used by every Rust test (integration.rs::make_gp).

Source:

• kriging-friedrich/test.mjs, line 41

Returns:

Type

GaussianProcess

makeGradientCovarianceMatrices(inputs, kernel) → {Array.<Array.<Array.<number>>>}

Description:

• For each kernel hyper-parameter, build the (symmetric) gradient matrix ∂K/∂param. Returns number[][][] of length kernel.nbParameters().

  mats[p][r][c] = mats[p][c][r] = kernel.gradient(inputs[r], inputs[c])[p]

Source:

• kriging-friedrich/algebra.mjs, line 819

Parameters:

Returns:

Type

Array.<Array.<Array.<number>>>

matMul(A, B) → {Array.<Array.<number>>}

Description:

• Matrix product A(m×k) · B(k×n) → (m×n).

Source:

• kriging-friedrich/algebra.mjs, line 109

Parameters:

Returns:

Type

Array.<Array.<number>>

matSub(A, B) → {Array.<Array.<number>>}

Description:

• Element-wise matrix subtraction A − B.

Source:

• kriging-friedrich/algebra.mjs, line 187

Parameters:

Returns:

Type

Array.<Array.<number>>

matTransposeVec(A, x) → {Array.<number>}

Description:

• Transposed matrix-vector product Aᵀ(n×m) · x(m) → number. Equivalent to transpose(A)·x without materializing the transpose.

Source:

• kriging-friedrich/algebra.mjs, line 152

Parameters:

Returns:

Type

Array.<number>

matVec(A, x) → {Array.<number>}

Description:

• Matrix-vector product A(m×n) · x(n) → number.

Source:

• kriging-friedrich/algebra.mjs, line 133

Parameters:

Returns:

Type

Array.<number>

mulKernels(k1, k2) → {KernelProd}

Description:

• Convenience constructor for k1 · k2 (Rust KernelArith Mul operator).

Source:

• kriging-friedrich/kernels.mjs, line 975

Parameters:

Returns:

Type

KernelProd

mulberry32(seed)

Description:

• Seedable 32-bit PRNG (mulberry32).

  Returns a closure () => number ∈ [0, 1) that advances the PRNG state on each call. Pass the returned function as the rng argument to standardNormal and MultivariateNormal.sample.

  Example: const rng = mulberry32(42); rng(); // reproducible pseudo-random number

Source:

• kriging-friedrich/multivariate_normal.mjs, line 41

Parameters:

Returns:

norm(a) → {number}

Description:

• Euclidean norm sqrt(Σ aᵢ²).

Source:

• kriging-friedrich/algebra.mjs, line 74

Parameters:

Returns:

Type

number

normSquared(a) → {number}

Description:

• Squared Euclidean norm Σ aᵢ².

Source:

• kriging-friedrich/algebra.mjs, line 83

Parameters:

Returns:

Type

number

optimizeParameters(gp, maxIter, convergenceFraction, maxTime)

Description:

• Fit parameters using the ADAM gradient-ascent algorithm.

  Runs for at most maxIter iterations. Stops early if every component of the update step delta is ≤ convergenceFraction in magnitude (no significant progress), or if the runtime exceeds maxTime (milliseconds).

  The noise parameter is fitted in log-scale, since its magnitude matters more than its precise value.

  ADAM constants (optimizer.rs:79): beta1=0.9, beta2=0.999, epsilon=1e-8, learningRate=0.1.

Source:

• kriging-friedrich/optimizer.mjs, line 125

Parameters:

procSpecPolygon(polygonsFC) → {Object}

Description:

• 處理特殊 Polygon 數據，讓 MapLibre GL JS 可正確渲染

  此函數的設計目的是處理 splitGeoJSON 回傳的 polygons FeatureCollection 中， 含有「多層套疊 ring」的 Polygon / MultiPolygon 數據。

  Leaflet 利用 SVG 的 evenodd fill rule 可直接繪製多層套疊的 ring， 但 MapLibre GL JS 僅依賴 winding order（外環 CCW、洞環 CW）來決定填色/挖洞。 因此，必須將多層套疊結構轉換為符合 RFC 7946 的標準 MultiPolygon 格式。

  處理邏輯：

  1. 遍歷所有 Feature
  2. 對於 Polygon 類型：
     • ring 數量 <= 2 時，使用 flattenMultiPolygon 修正 winding order
     • ring 數量 > 2 時（可能為多層套疊），使用 flattenMultiPolygon 搭配 supposeType='ringStrings' 模式，透過 XOR 運算將套疊結構轉為標準 MultiPolygon
  3. 對於 MultiPolygon 類型：
     • 將每個子 polygon 各自透過 flattenMultiPolygon 處理後合併
  4. 處理完畢後，若 MultiPolygon 只含一個 polygon，降級為 Polygon 類型
  5. 深拷貝輸出，不汙染原始資料
  6. 保留所有 properties

Source:

• splitAndProcGeoJSON.mjs, line 301

Example

// 三層套疊 Polygon
let input = {
    type: 'FeatureCollection',
    features: [{
        type: 'Feature',
        properties: { name: '三層' },
        geometry: {
            type: 'Polygon',
            coordinates: [
                [[0, 0], [20, 0], [20, 20], [0, 20], [0, 0]],
                [[4, 4], [16, 4], [16, 16], [4, 16], [4, 4]],
                [[8, 8], [12, 8], [12, 12], [8, 12], [8, 8]],
            ],
        },
    }],
}
let result = procSpecPolygon(input)
console.log(result.features[0].geometry.type) // 'MultiPolygon'

Parameters:

Returns:

Type

Object

processPolygonCoords(coordinates) → {Object}

Description:

• 處理單一 Polygon 的 coordinates

Source:

• splitAndProcGeoJSON.mjs, line 378

Parameters:

Returns:

Type

Object

scaleVec(a, s) → {Array.<number>}

Description:

• Scalar multiplication s·a.

Source:

• kriging-friedrich/algebra.mjs, line 63

Parameters:

Returns:

Type

Array.<number>

scaledGradientMarginalLikelihood(gp)

Description:

• Returns [scale, gradients]: the optimal scale for the kernel+noise (used to optimize the noise) plus the gradient per kernel parameter (NOT including the noise gradient).

  Per-parameter formula (optimizer.rs:150): ½ ( alphaᵀ · dp · alpha / scale − trace(K⁻¹ · dp) ) scale = outputᵀ · K⁻¹ · output / n NOTE: the data-fit term is divided by scale, unlike the unscaled gradient.

Source:

• kriging-friedrich/optimizer.mjs, line 194

Parameters:

Returns:

scaledOptimizeParameters(gp, maxIter, convergenceFraction, maxTime)

Description:

• Fit parameters using ADAM gradient ascent; additionally, at each step the kernel and noise are rescaled by the optimal magnitude scale.

  Runs for at most maxIter iterations. Stops early on no significant progress (all delta ≤ convergenceFraction) or if the runtime exceeds maxTime (milliseconds).

  ADAM constants identical to optimizeParameters.

Source:

• kriging-friedrich/optimizer.mjs, line 258

Parameters:

solveLowerTri(L, B) → {Array.<number>|Array.<Array.<number>>}

Description:

• Solve L·X = B where L is lower-triangular (forward substitution). B may be a vector (number[]) → solves one RHS, or a matrix (number[][]) → solves column by column (each column is one RHS), returning the same shape. Matches nalgebra solve_lower_triangular.

Source:

• kriging-friedrich/algebra.mjs, line 320

Parameters:

Returns:

Type

Array.<number> | Array.<Array.<number>>

solveLowerTriVec(L, b) → {Array.<number>}

Description:

• Forward substitution for a single RHS vector: solve L·x = b.

Source:

• kriging-friedrich/algebra.mjs, line 343

Parameters:

Returns:

Type

Array.<number>

solveUpperTri(U, B) → {Array.<number>|Array.<Array.<number>>}

Description:

• Solve U·X = B where U is upper-triangular (back substitution). In friedrich the upper-triangular system is U = Lᵀ. B may be number[] or number[][] (column-wise), returning the same shape.

Source:

• kriging-friedrich/algebra.mjs, line 364

Parameters:

Returns:

Type

Array.<number> | Array.<Array.<number>>

solveUpperTriVec(U, b) → {Array.<number>}

Description:

• Back substitution for a single RHS vector: solve U·x = b.

Source:

• kriging-friedrich/algebra.mjs, line 386

Parameters:

Returns:

Type

Array.<number>

splitAndProcGeoJSON(geoIn) → {Object}

Description:

• 將任意 GeoJSON 資料拆分為依幾何類型分類的多個 FeatureCollection， 並對其中的 Polygon 數據進行特殊處理（多層套疊 ring 轉換為標準 MultiPolygon）

  此函數結合了 splitGeoJSON 與 procSpecPolygon 的功能：

  1. splitGeoJSON：將混合幾何類型的 GeoJSON 拆分為 points / lines / polygons
  2. procSpecPolygon：將 polygons 中含有多層套疊 ring 的 Polygon/MultiPolygon 轉換為符合 RFC 7946 規範的標準格式，確保 winding order 正確

  此函數設計為方便日後 MapLibre GL JS 渲染使用的一站式前處理函數。

Source:

• splitAndProcGeoJSON.mjs, line 463

Example

let input = {
    type: 'FeatureCollection',
    features: [
        { type: 'Feature', properties: { name: 'pt' }, geometry: { type: 'Point', coordinates: [121, 25] } },
        { type: 'Feature', properties: { name: 'ls' }, geometry: { type: 'LineString', coordinates: [[121, 25], [122, 26]] } },
        {
            type: 'Feature',
            properties: { name: '三層' },
            geometry: {
                type: 'Polygon',
                coordinates: [
                    [[0, 0], [20, 0], [20, 20], [0, 20], [0, 0]],
                    [[4, 4], [16, 4], [16, 16], [4, 16], [4, 4]],
                    [[8, 8], [12, 8], [12, 12], [8, 12], [8, 8]],
                ],
            },
        },
    ],
}
let result = splitAndProcGeoJSON(input)
console.log(result.points.features.length)   // 1
console.log(result.lines.features.length)    // 1
console.log(result.polygons.features.length) // 1（polygon 已被處理為 MultiPolygon）

Parameters:

Returns:

Type

Object

splitGeoJSON(geoIn) → {Object}

Description:

• 將 GeoJSON 資料拆分成依幾何類型分類的多個 FeatureCollection

  此函數的設計目的是針對 MapLibre GL JS 的 layer type 限制， 將混合不同幾何類型的 GeoJSON 資料預先拆分成：

  • points：包含 Point / MultiPoint 的 FeatureCollection
  • lines：包含 LineString / MultiLineString 的 FeatureCollection
  • polygons：包含 Polygon / MultiPolygon 的 FeatureCollection

  額外處理：

  1. GeometryCollection 會被遞迴拆解成獨立的 Feature，properties 繼承父 Feature
  2. Polygon / MultiPolygon 的 ring 自動修復閉合（首尾座標相同）
  3. 支援多種輸入格式：FeatureCollection、Feature、裸 Geometry、JSON 字串
  4. 深拷貝輸出，不汙染原始資料
  5. 空輸入 / 無效輸入回傳空的分類結果

Source:

• splitAndProcGeoJSON.mjs, line 186

Example

// 混合類型 FeatureCollection
let input = {
    type: 'FeatureCollection',
    features: [
        { type: 'Feature', properties: { name: 'pt' }, geometry: { type: 'Point', coordinates: [121, 25] } },
        { type: 'Feature', properties: { name: 'ls' }, geometry: { type: 'LineString', coordinates: [[121, 25], [122, 26]] } },
        { type: 'Feature', properties: { name: 'pg' }, geometry: { type: 'Polygon', coordinates: [[[0, 0], [4, 0], [4, 4], [0, 4], [0, 0]]] } },
    ],
}
let result = splitGeoJSON(input)
console.log(result.points.features.length)   // 1
console.log(result.lines.features.length)    // 1
console.log(result.polygons.features.length) // 1

Parameters:

Returns:

Type

Object

standardNormal(rng) → {number}

Description:

• Draw one standard-normal sample N(0,1) using the Box-Muller transform.

  Consumes two uniform draws from rng; retries if the first draw is exactly 0 (to avoid log(0) → -Infinity).

  Note: this uses only the cosine branch of Box-Muller (the sine branch is discarded), matching the PORT_SPEC §3 canonical implementation. Statistical correctness is maintained; the sine branch would give equally valid samples but is not required.

Source:

• kriging-friedrich/multivariate_normal.mjs, line 70

Parameters:

Returns:

Type

number

subVec(a, b) → {Array.<number>}

Description:

• Element-wise subtraction a − b.

Source:

• kriging-friedrich/algebra.mjs, line 39

Parameters:

Returns:

Type

Array.<number>

toMatrix(input) → {Array.<Array.<number>>}

Description:

• Normalise any recognised input form to number[][] (one row per sample).

  JS input	Rust analogue	Result
  [a, b, c]	Vec<f64>	[[a, b, c]] (1×d)
  [[a,b],[c,d],…]	Vec<Vec<f64>>	unchanged

  Matches Input::to_dmatrix (both impls):

  • Vec<f64> → DMatrix::from_row_slice(1, m.len(), m) — wraps in 1 row.
  • Vec<Vec<f64>> → DMatrix::from_fn(nb_rows, nb_cols, |r,c| m[r][c]) — direct.

Source:

• kriging-friedrich/conversion.mjs, line 71

Parameters:

Returns:

Type

Array.<Array.<number>>

toVector(output) → {Array.<number>}

Description:

• Normalise a training-output value to number[].

  JS output	Rust analogue	Result
  number	f64	[v]
  number[]	Vec<f64>	unchanged

  Matches Input::to_dvector:

  • Vec<f64> impl: DVector::from_element(1, *v) — wraps scalar in 1-element vector.
  • Vec<Vec<f64>> impl: DVector::from_column_slice(v) — direct slice.

Source:

• kriging-friedrich/conversion.mjs, line 100

Parameters:

Returns:

Type

Array.<number>

trace(A) → {number}

Description:

• Trace Σ Aᵢᵢ (also used as generic diagonal sum, e.g. trace(A⁻¹)).

Source:

• kriging-friedrich/algebra.mjs, line 243

Parameters:

Returns:

Type

number

transpose(A) → {Array.<Array.<number>>}

Description:

• Transpose of A.

Source:

• kriging-friedrich/algebra.mjs, line 169

Parameters:

Returns:

Type

Array.<Array.<number>>

zeros(m, n) → {Array.<Array.<number>>}

Description:

• m×n zero matrix.

Source:

• kriging-friedrich/algebra.mjs, line 220

Parameters:

Returns:

Type

Array.<Array.<number>>