Initialization from an ordering vector for csmoothEM

Initializes csmoothEM parameters from a 1D ordering score by:

1. Discretizing ordering_vec into K ordered groups (equal/quantile/kmeans).

2. Computing component means mu.

3. Estimating diagonal variances sigma2 under either:

   • modelName="homoskedastic": one variance per coordinate shared across clusters (length d).

   • modelName="heteroskedastic": one variance per coordinate per cluster (d x K).

This function mirrors the discretization and mean construction logic of make_init(), but returns sigma2 (diagonal variances) instead of a list of full covariance matrices.

make_init_csmooth(
  X,
  ordering_vec,
  K,
  modelName = c("homoskedastic", "heteroskedastic"),
  nugget = 0,
  discretization = c("equal", "quantile", "kmeans"),
  na_action = c("drop", "error"),
  eps = 1e-12
)

Arguments

X

Numeric matrix (n x d).

ordering_vec

Numeric vector of length n (can contain NA).

K

Integer >= 2; number of mixture components.

modelName

Either "homoskedastic" or "heteroskedastic".

nugget

Nonnegative scalar added to variance estimates.

discretization

One of "equal", "quantile", "kmeans".

na_action

How to handle NA in ordering_vec: "drop" or "error".

eps

Small positive floor for pi and sigma2.

Value

A list with:

• pi: length-K mixing proportions.

• mu: list of length K, each a length-d mean vector.

• sigma2: diagonal variances; length-d vector (homoskedastic) or d x K matrix (heteroskedastic).

• keep_idx: row indices kept after NA handling.

• cluster_rank: integer vector in 1..K for each kept row.