xref: /petsc/src/tao/util/tao_util.c (revision d71ae5a4db6382e7f06317b8d368875286fe9008)

#include <petsc/private/petscimpl.h>
#include <petsctao.h> /*I "petsctao.h" I*/
#include <petscsys.h>

static inline PetscReal Fischer(PetscReal a, PetscReal b)
{
  /* Method suggested by Bob Vanderbei */
  if (a + b <= 0) return PetscSqrtReal(a * a + b * b) - (a + b);
  return -2.0 * a * b / (PetscSqrtReal(a * a + b * b) + (a + b));
}

/*@
   VecFischer - Evaluates the Fischer-Burmeister function for complementarity
   problems.

   Logically Collective on vectors

   Input Parameters:
+  X - current point
.  F - function evaluated at x
.  L - lower bounds
-  U - upper bounds

   Output Parameter:
.  FB - The Fischer-Burmeister function vector

   Notes:
   The Fischer-Burmeister function is defined as
$        phi(a,b) := sqrt(a*a + b*b) - a - b
   and is used reformulate a complementarity problem as a semismooth
   system of equations.

   The result of this function is done by cases:
+  l[i] == -infinity, u[i] == infinity  -- fb[i] = -f[i]
.  l[i] == -infinity, u[i] finite       -- fb[i] = phi(u[i]-x[i], -f[i])
.  l[i] finite,       u[i] == infinity  -- fb[i] = phi(x[i]-l[i],  f[i])
.  l[i] finite < u[i] finite -- fb[i] = phi(x[i]-l[i], phi(u[i]-x[i], -f[u]))
-  otherwise l[i] == u[i] -- fb[i] = l[i] - x[i]

   Level: developer

@*/
PetscErrorCode VecFischer(Vec X, Vec F, Vec L, Vec U, Vec FB)
{
  const PetscScalar *x, *f, *l, *u;
  PetscScalar       *fb;
  PetscReal          xval, fval, lval, uval;
  PetscInt           low[5], high[5], n, i;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(X, VEC_CLASSID, 1);
  PetscValidHeaderSpecific(F, VEC_CLASSID, 2);
  if (L) PetscValidHeaderSpecific(L, VEC_CLASSID, 3);
  if (U) PetscValidHeaderSpecific(U, VEC_CLASSID, 4);
  PetscValidHeaderSpecific(FB, VEC_CLASSID, 5);

  if (!L && !U) {
    PetscCall(VecAXPBY(FB, -1.0, 0.0, F));
    PetscFunctionReturn(0);
  }

  PetscCall(VecGetOwnershipRange(X, low, high));
  PetscCall(VecGetOwnershipRange(F, low + 1, high + 1));
  PetscCall(VecGetOwnershipRange(L, low + 2, high + 2));
  PetscCall(VecGetOwnershipRange(U, low + 3, high + 3));
  PetscCall(VecGetOwnershipRange(FB, low + 4, high + 4));

  for (i = 1; i < 4; ++i) PetscCheck(low[0] == low[i] && high[0] == high[i], PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Vectors must be identically loaded over processors");

  PetscCall(VecGetArrayRead(X, &x));
  PetscCall(VecGetArrayRead(F, &f));
  PetscCall(VecGetArrayRead(L, &l));
  PetscCall(VecGetArrayRead(U, &u));
  PetscCall(VecGetArray(FB, &fb));

  PetscCall(VecGetLocalSize(X, &n));

  for (i = 0; i < n; ++i) {
    xval = PetscRealPart(x[i]);
    fval = PetscRealPart(f[i]);
    lval = PetscRealPart(l[i]);
    uval = PetscRealPart(u[i]);

    if (lval <= -PETSC_INFINITY && uval >= PETSC_INFINITY) {
      fb[i] = -fval;
    } else if (lval <= -PETSC_INFINITY) {
      fb[i] = -Fischer(uval - xval, -fval);
    } else if (uval >= PETSC_INFINITY) {
      fb[i] = Fischer(xval - lval, fval);
    } else if (lval == uval) {
      fb[i] = lval - xval;
    } else {
      fval  = Fischer(uval - xval, -fval);
      fb[i] = Fischer(xval - lval, fval);
    }
  }

  PetscCall(VecRestoreArrayRead(X, &x));
  PetscCall(VecRestoreArrayRead(F, &f));
  PetscCall(VecRestoreArrayRead(L, &l));
  PetscCall(VecRestoreArrayRead(U, &u));
  PetscCall(VecRestoreArray(FB, &fb));
  PetscFunctionReturn(0);
}

static inline PetscReal SFischer(PetscReal a, PetscReal b, PetscReal c)
{
  /* Method suggested by Bob Vanderbei */
  if (a + b <= 0) return PetscSqrtReal(a * a + b * b + 2.0 * c * c) - (a + b);
  return 2.0 * (c * c - a * b) / (PetscSqrtReal(a * a + b * b + 2.0 * c * c) + (a + b));
}

/*@
   VecSFischer - Evaluates the Smoothed Fischer-Burmeister function for
   complementarity problems.

   Logically Collective on vectors

   Input Parameters:
+  X - current point
.  F - function evaluated at x
.  L - lower bounds
.  U - upper bounds
-  mu - smoothing parameter

   Output Parameter:
.  FB - The Smoothed Fischer-Burmeister function vector

   Notes:
   The Smoothed Fischer-Burmeister function is defined as
$        phi(a,b) := sqrt(a*a + b*b + 2*mu*mu) - a - b
   and is used reformulate a complementarity problem as a semismooth
   system of equations.

   The result of this function is done by cases:
+  l[i] == -infinity, u[i] == infinity  -- fb[i] = -f[i] - 2*mu*x[i]
.  l[i] == -infinity, u[i] finite       -- fb[i] = phi(u[i]-x[i], -f[i], mu)
.  l[i] finite,       u[i] == infinity  -- fb[i] = phi(x[i]-l[i],  f[i], mu)
.  l[i] finite < u[i] finite -- fb[i] = phi(x[i]-l[i], phi(u[i]-x[i], -f[u], mu), mu)
-  otherwise l[i] == u[i] -- fb[i] = l[i] - x[i]

   Level: developer

.seealso `VecFischer()`
@*/
PetscErrorCode VecSFischer(Vec X, Vec F, Vec L, Vec U, PetscReal mu, Vec FB)
{
  const PetscScalar *x, *f, *l, *u;
  PetscScalar       *fb;
  PetscReal          xval, fval, lval, uval;
  PetscInt           low[5], high[5], n, i;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(X, VEC_CLASSID, 1);
  PetscValidHeaderSpecific(F, VEC_CLASSID, 2);
  PetscValidHeaderSpecific(L, VEC_CLASSID, 3);
  PetscValidHeaderSpecific(U, VEC_CLASSID, 4);
  PetscValidHeaderSpecific(FB, VEC_CLASSID, 6);

  PetscCall(VecGetOwnershipRange(X, low, high));
  PetscCall(VecGetOwnershipRange(F, low + 1, high + 1));
  PetscCall(VecGetOwnershipRange(L, low + 2, high + 2));
  PetscCall(VecGetOwnershipRange(U, low + 3, high + 3));
  PetscCall(VecGetOwnershipRange(FB, low + 4, high + 4));

  for (i = 1; i < 4; ++i) PetscCheck(low[0] == low[i] && high[0] == high[i], PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Vectors must be identically loaded over processors");

  PetscCall(VecGetArrayRead(X, &x));
  PetscCall(VecGetArrayRead(F, &f));
  PetscCall(VecGetArrayRead(L, &l));
  PetscCall(VecGetArrayRead(U, &u));
  PetscCall(VecGetArray(FB, &fb));

  PetscCall(VecGetLocalSize(X, &n));

  for (i = 0; i < n; ++i) {
    xval = PetscRealPart(*x++);
    fval = PetscRealPart(*f++);
    lval = PetscRealPart(*l++);
    uval = PetscRealPart(*u++);

    if ((lval <= -PETSC_INFINITY) && (uval >= PETSC_INFINITY)) {
      (*fb++) = -fval - mu * xval;
    } else if (lval <= -PETSC_INFINITY) {
      (*fb++) = -SFischer(uval - xval, -fval, mu);
    } else if (uval >= PETSC_INFINITY) {
      (*fb++) = SFischer(xval - lval, fval, mu);
    } else if (lval == uval) {
      (*fb++) = lval - xval;
    } else {
      fval    = SFischer(uval - xval, -fval, mu);
      (*fb++) = SFischer(xval - lval, fval, mu);
    }
  }
  x -= n;
  f -= n;
  l -= n;
  u -= n;
  fb -= n;

  PetscCall(VecRestoreArrayRead(X, &x));
  PetscCall(VecRestoreArrayRead(F, &f));
  PetscCall(VecRestoreArrayRead(L, &l));
  PetscCall(VecRestoreArrayRead(U, &u));
  PetscCall(VecRestoreArray(FB, &fb));
  PetscFunctionReturn(0);
}

static inline PetscReal fischnorm(PetscReal a, PetscReal b)
{
  return PetscSqrtReal(a * a + b * b);
}

static inline PetscReal fischsnorm(PetscReal a, PetscReal b, PetscReal c)
{
  return PetscSqrtReal(a * a + b * b + 2.0 * c * c);
}

/*@
   MatDFischer - Calculates an element of the B-subdifferential of the
   Fischer-Burmeister function for complementarity problems.

   Collective on jac

   Input Parameters:
+  jac - the jacobian of f at X
.  X - current point
.  Con - constraints function evaluated at X
.  XL - lower bounds
.  XU - upper bounds
.  t1 - work vector
-  t2 - work vector

   Output Parameters:
+  Da - diagonal perturbation component of the result
-  Db - row scaling component of the result

   Level: developer

.seealso: `VecFischer()`
@*/
PetscErrorCode MatDFischer(Mat jac, Vec X, Vec Con, Vec XL, Vec XU, Vec T1, Vec T2, Vec Da, Vec Db)
{
  PetscInt           i, nn;
  const PetscScalar *x, *f, *l, *u, *t2;
  PetscScalar       *da, *db, *t1;
  PetscReal          ai, bi, ci, di, ei;

  PetscFunctionBegin;
  PetscCall(VecGetLocalSize(X, &nn));
  PetscCall(VecGetArrayRead(X, &x));
  PetscCall(VecGetArrayRead(Con, &f));
  PetscCall(VecGetArrayRead(XL, &l));
  PetscCall(VecGetArrayRead(XU, &u));
  PetscCall(VecGetArray(Da, &da));
  PetscCall(VecGetArray(Db, &db));
  PetscCall(VecGetArray(T1, &t1));
  PetscCall(VecGetArrayRead(T2, &t2));

  for (i = 0; i < nn; i++) {
    da[i] = 0.0;
    db[i] = 0.0;
    t1[i] = 0.0;

    if (PetscAbsScalar(f[i]) <= PETSC_MACHINE_EPSILON) {
      if (PetscRealPart(l[i]) > PETSC_NINFINITY && PetscAbsScalar(x[i] - l[i]) <= PETSC_MACHINE_EPSILON) {
        t1[i] = 1.0;
        da[i] = 1.0;
      }

      if (PetscRealPart(u[i]) < PETSC_INFINITY && PetscAbsScalar(u[i] - x[i]) <= PETSC_MACHINE_EPSILON) {
        t1[i] = 1.0;
        db[i] = 1.0;
      }
    }
  }

  PetscCall(VecRestoreArray(T1, &t1));
  PetscCall(VecRestoreArrayRead(T2, &t2));
  PetscCall(MatMult(jac, T1, T2));
  PetscCall(VecGetArrayRead(T2, &t2));

  for (i = 0; i < nn; i++) {
    if ((PetscRealPart(l[i]) <= PETSC_NINFINITY) && (PetscRealPart(u[i]) >= PETSC_INFINITY)) {
      da[i] = 0.0;
      db[i] = -1.0;
    } else if (PetscRealPart(l[i]) <= PETSC_NINFINITY) {
      if (PetscRealPart(db[i]) >= 1) {
        ai = fischnorm(1.0, PetscRealPart(t2[i]));

        da[i] = -1.0 / ai - 1.0;
        db[i] = -t2[i] / ai - 1.0;
      } else {
        bi = PetscRealPart(u[i]) - PetscRealPart(x[i]);
        ai = fischnorm(bi, PetscRealPart(f[i]));
        ai = PetscMax(PETSC_MACHINE_EPSILON, ai);

        da[i] = bi / ai - 1.0;
        db[i] = -f[i] / ai - 1.0;
      }
    } else if (PetscRealPart(u[i]) >= PETSC_INFINITY) {
      if (PetscRealPart(da[i]) >= 1) {
        ai = fischnorm(1.0, PetscRealPart(t2[i]));

        da[i] = 1.0 / ai - 1.0;
        db[i] = t2[i] / ai - 1.0;
      } else {
        bi = PetscRealPart(x[i]) - PetscRealPart(l[i]);
        ai = fischnorm(bi, PetscRealPart(f[i]));
        ai = PetscMax(PETSC_MACHINE_EPSILON, ai);

        da[i] = bi / ai - 1.0;
        db[i] = f[i] / ai - 1.0;
      }
    } else if (PetscRealPart(l[i]) == PetscRealPart(u[i])) {
      da[i] = -1.0;
      db[i] = 0.0;
    } else {
      if (PetscRealPart(db[i]) >= 1) {
        ai = fischnorm(1.0, PetscRealPart(t2[i]));

        ci = 1.0 / ai + 1.0;
        di = PetscRealPart(t2[i]) / ai + 1.0;
      } else {
        bi = PetscRealPart(x[i]) - PetscRealPart(u[i]);
        ai = fischnorm(bi, PetscRealPart(f[i]));
        ai = PetscMax(PETSC_MACHINE_EPSILON, ai);

        ci = bi / ai + 1.0;
        di = PetscRealPart(f[i]) / ai + 1.0;
      }

      if (PetscRealPart(da[i]) >= 1) {
        bi = ci + di * PetscRealPart(t2[i]);
        ai = fischnorm(1.0, bi);

        bi = bi / ai - 1.0;
        ai = 1.0 / ai - 1.0;
      } else {
        ei = Fischer(PetscRealPart(u[i]) - PetscRealPart(x[i]), -PetscRealPart(f[i]));
        ai = fischnorm(PetscRealPart(x[i]) - PetscRealPart(l[i]), ei);
        ai = PetscMax(PETSC_MACHINE_EPSILON, ai);

        bi = ei / ai - 1.0;
        ai = (PetscRealPart(x[i]) - PetscRealPart(l[i])) / ai - 1.0;
      }

      da[i] = ai + bi * ci;
      db[i] = bi * di;
    }
  }

  PetscCall(VecRestoreArray(Da, &da));
  PetscCall(VecRestoreArray(Db, &db));
  PetscCall(VecRestoreArrayRead(X, &x));
  PetscCall(VecRestoreArrayRead(Con, &f));
  PetscCall(VecRestoreArrayRead(XL, &l));
  PetscCall(VecRestoreArrayRead(XU, &u));
  PetscCall(VecRestoreArrayRead(T2, &t2));
  PetscFunctionReturn(0);
}

/*@
   MatDSFischer - Calculates an element of the B-subdifferential of the
   smoothed Fischer-Burmeister function for complementarity problems.

   Collective on jac

   Input Parameters:
+  jac - the jacobian of f at X
.  X - current point
.  F - constraint function evaluated at X
.  XL - lower bounds
.  XU - upper bounds
.  mu - smoothing parameter
.  T1 - work vector
-  T2 - work vector

   Output Parameters:
+  Da - diagonal perturbation component of the result
.  Db - row scaling component of the result
-  Dm - derivative with respect to scaling parameter

   Level: developer

.seealso `MatDFischer()`
@*/
PetscErrorCode MatDSFischer(Mat jac, Vec X, Vec Con, Vec XL, Vec XU, PetscReal mu, Vec T1, Vec T2, Vec Da, Vec Db, Vec Dm)
{
  PetscInt           i, nn;
  const PetscScalar *x, *f, *l, *u;
  PetscScalar       *da, *db, *dm;
  PetscReal          ai, bi, ci, di, ei, fi;

  PetscFunctionBegin;
  if (PetscAbsReal(mu) <= PETSC_MACHINE_EPSILON) {
    PetscCall(VecZeroEntries(Dm));
    PetscCall(MatDFischer(jac, X, Con, XL, XU, T1, T2, Da, Db));
  } else {
    PetscCall(VecGetLocalSize(X, &nn));
    PetscCall(VecGetArrayRead(X, &x));
    PetscCall(VecGetArrayRead(Con, &f));
    PetscCall(VecGetArrayRead(XL, &l));
    PetscCall(VecGetArrayRead(XU, &u));
    PetscCall(VecGetArray(Da, &da));
    PetscCall(VecGetArray(Db, &db));
    PetscCall(VecGetArray(Dm, &dm));

    for (i = 0; i < nn; ++i) {
      if ((PetscRealPart(l[i]) <= PETSC_NINFINITY) && (PetscRealPart(u[i]) >= PETSC_INFINITY)) {
        da[i] = -mu;
        db[i] = -1.0;
        dm[i] = -x[i];
      } else if (PetscRealPart(l[i]) <= PETSC_NINFINITY) {
        bi = PetscRealPart(u[i]) - PetscRealPart(x[i]);
        ai = fischsnorm(bi, PetscRealPart(f[i]), mu);
        ai = PetscMax(PETSC_MACHINE_EPSILON, ai);

        da[i] = bi / ai - 1.0;
        db[i] = -PetscRealPart(f[i]) / ai - 1.0;
        dm[i] = 2.0 * mu / ai;
      } else if (PetscRealPart(u[i]) >= PETSC_INFINITY) {
        bi = PetscRealPart(x[i]) - PetscRealPart(l[i]);
        ai = fischsnorm(bi, PetscRealPart(f[i]), mu);
        ai = PetscMax(PETSC_MACHINE_EPSILON, ai);

        da[i] = bi / ai - 1.0;
        db[i] = PetscRealPart(f[i]) / ai - 1.0;
        dm[i] = 2.0 * mu / ai;
      } else if (PetscRealPart(l[i]) == PetscRealPart(u[i])) {
        da[i] = -1.0;
        db[i] = 0.0;
        dm[i] = 0.0;
      } else {
        bi = PetscRealPart(x[i]) - PetscRealPart(u[i]);
        ai = fischsnorm(bi, PetscRealPart(f[i]), mu);
        ai = PetscMax(PETSC_MACHINE_EPSILON, ai);

        ci = bi / ai + 1.0;
        di = PetscRealPart(f[i]) / ai + 1.0;
        fi = 2.0 * mu / ai;

        ei = SFischer(PetscRealPart(u[i]) - PetscRealPart(x[i]), -PetscRealPart(f[i]), mu);
        ai = fischsnorm(PetscRealPart(x[i]) - PetscRealPart(l[i]), ei, mu);
        ai = PetscMax(PETSC_MACHINE_EPSILON, ai);

        bi = ei / ai - 1.0;
        ei = 2.0 * mu / ei;
        ai = (PetscRealPart(x[i]) - PetscRealPart(l[i])) / ai - 1.0;

        da[i] = ai + bi * ci;
        db[i] = bi * di;
        dm[i] = ei + bi * fi;
      }
    }

    PetscCall(VecRestoreArrayRead(X, &x));
    PetscCall(VecRestoreArrayRead(Con, &f));
    PetscCall(VecRestoreArrayRead(XL, &l));
    PetscCall(VecRestoreArrayRead(XU, &u));
    PetscCall(VecRestoreArray(Da, &da));
    PetscCall(VecRestoreArray(Db, &db));
    PetscCall(VecRestoreArray(Dm, &dm));
  }
  PetscFunctionReturn(0);
}

static inline PetscReal ST_InternalPN(PetscScalar in, PetscReal lb, PetscReal ub)
{
  return PetscMax(0, (PetscReal)PetscRealPart(in) - ub) - PetscMax(0, -(PetscReal)PetscRealPart(in) - PetscAbsReal(lb));
}

static inline PetscReal ST_InternalNN(PetscScalar in, PetscReal lb, PetscReal ub)
{
  return PetscMax(0, (PetscReal)PetscRealPart(in) + PetscAbsReal(ub)) - PetscMax(0, -(PetscReal)PetscRealPart(in) - PetscAbsReal(lb));
}

static inline PetscReal ST_InternalPP(PetscScalar in, PetscReal lb, PetscReal ub)
{
  return PetscMax(0, (PetscReal)PetscRealPart(in) - ub) + PetscMin(0, (PetscReal)PetscRealPart(in) - lb);
}

/*@
   TaoSoftThreshold - Calculates soft thresholding routine with input vector
   and given lower and upper bound and returns it to output vector.

   Input Parameters:
+  in - input vector to be thresholded
.  lb - lower bound
-  ub - upper bound

   Output Parameter:
.  out - Soft thresholded output vector

   Notes:
   Soft thresholding is defined as
   \[ S(input,lb,ub) =
     \begin{cases}
    input - ub  \text{input > ub} \\
    0           \text{lb =< input <= ub} \\
    input + lb  \text{input < lb} \\
   \]

   Level: developer

@*/
PetscErrorCode TaoSoftThreshold(Vec in, PetscReal lb, PetscReal ub, Vec out)
{
  PetscInt     i, nlocal, mlocal;
  PetscScalar *inarray, *outarray;

  PetscFunctionBegin;
  PetscCall(VecGetArrayPair(in, out, &inarray, &outarray));
  PetscCall(VecGetLocalSize(in, &nlocal));
  PetscCall(VecGetLocalSize(in, &mlocal));

  PetscCheck(nlocal == mlocal, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Input and output vectors need to be of same size.");
  PetscCheck(lb < ub, PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Lower bound needs to be lower than upper bound.");

  if (ub >= 0 && lb < 0) {
    for (i = 0; i < nlocal; i++) outarray[i] = ST_InternalPN(inarray[i], lb, ub);
  } else if (ub < 0 && lb < 0) {
    for (i = 0; i < nlocal; i++) outarray[i] = ST_InternalNN(inarray[i], lb, ub);
  } else {
    for (i = 0; i < nlocal; i++) outarray[i] = ST_InternalPP(inarray[i], lb, ub);
  }

  PetscCall(VecRestoreArrayPair(in, out, &inarray, &outarray));
  PetscFunctionReturn(0);
}