QCaml/Utils/Davidson.ml

open Lacaml.D

type t

let make
    ?guess
    ?(n_states=1)
    ?(n_iter=8)
    ?(threshold=1.e-7)
    diagonal
    matrix_prod
  =

  (* Size of the matrix to diagonalize *)
  let n = Vec.dim diagonal in

  let m = n_states in

  (* Create guess vectors u, with unknown vectors initialized to unity. *)
  let init_vectors m =
    let init_vector k =
      Vector.sparse_of_assoc_list n [ (k,1.0) ]
    in
    Array.init m (fun i -> init_vector (i+1)) 
    |> Matrix.sparse_of_vector_array
    |> Matrix.to_mat
  in

  let guess =
    match guess with
    | Some vectors -> vectors 
    | None -> init_vectors m
  in

  let guess =
    if Mat.dim2 guess = n_states then guess
    else
      (Mat.to_col_vecs_list guess) @
        (Mat.to_col_vecs_list (init_vectors (m-(Mat.dim2 guess))) )
      |> Mat.of_col_vecs_list
  in

  let pick_new u =
    Mat.to_col_vecs_list u
    |> Util.list_pack m
    |> List.rev
    |> List.hd
  in

  let u_new = Mat.to_col_vecs_list guess in

  let rec iteration u u_new w iter macro =
    (* u is a list of orthonormal vectors, on which the operator has
       been applied : w = op.u
       u_new is a list of vectors which will increase the size of the
       space.
    *)
    (* Orthonormalize input vectors u_new *)
    let u_new_ortho =
      List.concat [u ; u_new]
      |> Mat.of_col_vecs_list
      |> Util.qr_ortho 
      |> pick_new
    in

    (* Apply the operator to the m last vectors *)
    let w_new =
      matrix_prod (
        u_new_ortho
        |> Mat.of_col_vecs_list
        |> Matrix.dense_of_mat )
      |> Matrix.to_mat
      |> Mat.to_col_vecs_list
    in             

    (* Data for the next iteration *)
    let u_next =
      List.concat [ u ; u_new_ortho ]
    and w_next = 
      List.concat [ w ; w_new ]
    in

    (* Build the small matrix h = <U_k | W_l> *)
    let m_U =
      Mat.of_col_vecs_list u_next
    and m_W =
      Mat.of_col_vecs_list w_next
    in
    let m_h =
      gemm ~transa:`T m_U m_W
    in

    (* Diagonalize h *)
    let y, lambda =
      Util.diagonalize_symm m_h
    in

    (* Express m lowest eigenvectors of h in the large basis *)
    let m_new_U =
      gemm  ~n:m  m_U  y
    and m_new_W =
      gemm  ~n:m  m_W  y
    in

    (* Compute the residual as proposed new vectors *)
    let u_proposed = 
      Mat.init_cols n m (fun i k ->
          let delta = lambda.{k} -. diagonal.{i} in
          let delta = 
            if abs_float delta > 1.e-2  then delta
            else if delta > 0. then 1.e-2
            else (-1.e-2)
          in
            (lambda.{k} *. m_new_U.{i,k} -. m_new_W.{i,k}) /. delta
          )
      |> Mat.to_col_vecs_list
    in


    let residual_norms = List.rev @@ List.rev_map nrm2 u_proposed in
    let residual_norm  =
      List.fold_left (fun accu i -> accu +. i *. i) 0. residual_norms
      |> sqrt
    in

    if Parallel.master then
      begin
        Printf.printf "%3d " iter;
        Vec.iteri (fun i x -> if (i<=m) then Printf.printf "%16.10f  " x) lambda;
        Printf.printf "%16.8e%!\n" residual_norm
      end;

    (* Make new vectors sparse *)

    let u_proposed = 
      Mat.of_col_vecs_list u_proposed
    in
    let maxu = lange u_proposed ~norm:`M in
    let thr = maxu *. 0.00001 in
    let u_proposed = 
      Mat.map (fun x -> if abs_float x < thr then 0. else x) u_proposed
      |> Mat.to_col_vecs_list
    in
    (*
    Format.printf "%a@." Matrix.pp_matrix @@ Matrix.dense_of_mat m_new_U;
    *)


    if residual_norm > threshold && macro > 0 then
      let u_next, w_next, iter, macro =
        if iter = n_iter then
          m_new_U |> pick_new, 
          m_new_W |> pick_new,
          0, (macro-1)
        else
          u_next, w_next, (iter+1), macro
      in
      (iteration [@tailcall]) u_next u_proposed w_next iter macro
    else
      (m_new_U |> pick_new |> Mat.of_col_vecs_list), lambda

  in
  iteration [] u_new [] 1 30