Changes in version 1.0.3                        

Performance

Simulated annealing

  - O(base::ceiling(base::log(stop.temp / start.temp) /
    base::log(cool.fact)) · n) time, O(n^2) space →
    O(base::ceiling(base::log(stop.temp / start.temp) /
    base::log(cool.fact)) · k) time, O(n^2) space. Swap evaluation now
    computes delta updates via unchanged indices (unch.idx), replacing
    the maintained link vector and matching the brute force update
    pattern.

                 Changes in version 1.0.2 (2026-04-14)                  

Arguments and value

Renamed arguments

  - adj.mat → n.mat. The first argument of interest is n.mat, not
    necessarily a symmetric adjacency matrix of numeric, but a symmetric
    matrix of numeric corresponding to the symmetric matrix of
    associations (e.g., of zero-order polychoric correlation
    coefficients) computed from the set of items comprised in the
    original version of the said psychometric instrument.
  - method → algorithm. The third argument of interest is algorithm
    (rather than method), a character matching either “brute.force” or
    “simulated.annealing” corresponding to the combinatorial search
    algorithm.

Removed arguments

  - absolute now always TRUE.
  - max.iter now always base::ceiling(base::log(stop.temp / start.temp)
    / base::log(cool.fact)).

Renamed value

  - optimal.S → solution.character.
  - optimal.function.S → solution.numeric.

Robustness and correctness

  - base::set.seed() no longer permanently alters the global RNG state.
    .Random.seed is now saved and restored via base::on.exit().
  - Progress bars are now closed via base::on.exit(), ensuring cleanup
    even on early exit or error.
  - The diagonal of the input matrix is now explicitly set to zero
    before computation, preventing self-associations from inflating the
    objective.
  - NA/NaN values in the input matrix now trigger a warning with a count
    of affected entries before being replaced with zeros.
  - Input validation now checks for Inf values, non-numeric entries,
    non-square matrices, and matrices smaller than 2 × 2.
  - Scalar parameter validation now checks length, type, and NA status
    consistently across all parameters.
  - algorithm now uses base::match.arg() instead of manual %in%
    checking, supporting partial matching.
  - SA-specific parameters (start.temp, cool.fact, stop.temp, n.runs,
    seed) are now validated only when algorithm = "simulated.annealing".
  - Error messages now use call. = FALSE for cleaner console output.
  - k = 0 and k = n are now rejected during validation rather than
    handled as special cases inside the algorithm.
  - seed now explicitly accepts NULL to skip setting the RNG state.

Performance

Brute force

  - Subsets are now enumerated incrementally with delta updates to the
    objective function, replacing the previous approach of generating
    all candidate subsets in memory via utils::combn(). This reduces
    memory from O(C(n,k) × k) to O(n) and avoids recomputing the full
    objective at each step.
  - When k > n / 2, enumeration is performed over the smaller
    complement, reducing the number of combinations to evaluate.
  - The absolute value matrix is now precomputed once, rather than
    recomputed inside the objective function at every evaluation.
  - Brute force now detects intractable problems (C(n,k) exceeding
    .Machine$integer.max) and exits with a recommendation to use
    simulated annealing, rather than attempting to allocate the full
    combination matrix.

Simulated annealing

  - Swap evaluation now runs in O(n) via a maintained link vector,
    replacing the previous full matrix subsetting at each step.
  - The complement set is now maintained incrementally via a logical
    membership vector, replacing repeated base::setdiff() calls.
  - Removed the hash environment that stored all visited subsets as
    character keys, eliminating unbounded memory growth over long runs.

                 Changes in version 1.0.1 (2025-04-15)                  

  - Initial CRAN release.