inline DynProgProbLim (const DynProgProbLim &dynProgProbLim_)
: DynProgProb (dynProgProbLim_),
d_probLost (dynProgProbLim_.getProbLost ())
{}
virtual inline void copy (const DynProgProbLim &dynProgProbLim_)
{
copy (dynProgProbLim_, dynProgProbLim_.getProbLost ());
}
void LocalMaxStatUtil::descendingLadderEpochRepeat (
size_t dimension_, // #(distinct values)
const Int4 *score_, // values
const double *prob_, // probability of corresponding value
double *eSumAlpha_, // expectation (sum [alpha])
double *eOneMinusExpSumAlpha_, // expectation [1.0 - exp (sum [alpha])]
bool isStrict_, // ? is this a strict descending ladder epoch
double lambda_, // lambda for repeats : default is lambda0_ below
size_t endW_, // maximum w plus 1
double *pAlphaW_, // probability {alpha = w} : pAlphaW_ [0, wEnd)
double *eOneMinusExpSumAlphaW_, // expectation [1.0 - exp (sum [alpha]); alpha = w] : eOneMinusExpSumAlphaW_ [0, wEnd)
double lambda0_, // lambda for flattened distribution (avoid recomputation)
double mu0_, // mean of flattened distribution (avoid recomputation)
double muAssoc0_, // mean of associated flattened distribution (avoid recomputation)
double thetaMin0_, // thetaMin of flattened distribution (avoid recomputation)
double rMin0_, // rMin of flattened distribution (avoid recomputation)
double time_, // get time for the dynamic programming computation
bool *terminated_) // ? Was the dynamic programming computation terminated prematurely ?
// assumes logarithmic regime
{
// Start dynamic programming probability calculation using notation in
//
// Mott R. and Tribe R. (1999)
// J. Computational Biology 6(1):91-112
//
// Karlin S. and Taylor H.M.(1981)
// A Second Course in Stochastic Processes, p. 480
//
// Note there is an error in Eq (6.19) there, which is corrected in Eq (6.20)
//
// This program uses departure into (-Inf, 0] not (-Inf, 0)
// avoid recomputation
double mu0 = 0.0 == mu0_ ? mu (dimension_, score_, prob_) : mu0_;
assert (mu0 < 0.0);
double lambda0 = 0.0 == lambda0_ ? lambda (dimension_, score_, prob_) : lambda0_;
assert (0.0 < lambda0);
if (lambda_ == 0.0) lambda_ = lambda0;
assert (0.0 < lambda_);
double muAssoc0 = 0.0 == muAssoc0_ ? muAssoc (dimension_, score_, prob_, lambda0) : muAssoc0_;
assert (0.0 < muAssoc0);
double thetaMin0 = 0.0 == thetaMin0_ ? thetaMin (dimension_, score_, prob_, lambda0) : thetaMin0_;
assert (0.0 < thetaMin0);
double rMin0 = 0.0 == rMin0_ ? rMin (dimension_, score_, prob_, lambda0, thetaMin0) : rMin0_;
assert (0.0 < rMin0 && rMin0 < 1.0);
const Int4 ITER_MIN = static_cast <Int4> ((log (REL_TOL * (1.0 - rMin0)) / log (rMin0)));
assert (0 < ITER_MIN);
const Int4 ITER = static_cast <Int4> (endW_) < ITER_MIN ? ITER_MIN : static_cast <Int4> (endW_);
assert (0 < ITER);
const Int4 Y_MAX = static_cast <Int4> (-log (REL_TOL) / lambda0);
Int4 entry = isStrict_ ? -1 : 0;
n_setParameters (dimension_, score_, prob_, entry);
double time0 = 0.0;
double time1 = 0.0;
if (time_ > 0.0) Sls::alp_data::get_current_time (time0);
DynProgProbLim dynProgProb (n_step, dimension_, prob_, score_ [0] - 1, Y_MAX);
if (pAlphaW_) pAlphaW_ [0] = 0.0;
if (eOneMinusExpSumAlphaW_) eOneMinusExpSumAlphaW_ [0] = 0.0;
dynProgProb.update (); // iterate random walk
Int4 value = 0;
if (eSumAlpha_) *eSumAlpha_ = 0.0;
if (eOneMinusExpSumAlpha_) *eOneMinusExpSumAlpha_ = 0.0;
for (size_t w = 1; w < static_cast <size_t> (ITER); w++) {
if (w < endW_) { // sum pAlphaW_ [w] and eOneMinusExpSumAlphaW_ [w]
if (pAlphaW_) pAlphaW_ [w] = 0.0;
if (eOneMinusExpSumAlphaW_) eOneMinusExpSumAlphaW_ [w] = 0.0;
for (value = score_ [0]; value <= entry; value++) {
if (pAlphaW_) pAlphaW_ [w] += dynProgProb.getProb (value);
if (eOneMinusExpSumAlphaW_) eOneMinusExpSumAlphaW_ [w] +=
dynProgProb.getProb (value) *
(1.0 - exp (lambda_ * static_cast <double> (value)));
}
}
for (value = score_ [0]; value <= entry; value++) {
if (eSumAlpha_) *eSumAlpha_ += dynProgProb.getProb (value) * static_cast <double> (value);
if (eOneMinusExpSumAlpha_) *eOneMinusExpSumAlpha_ += dynProgProb.getProb (value) *
(1.0 - exp (lambda_ * static_cast <double> (value)));
}
dynProgProb.setValueFct (n_bury);
dynProgProb.update (); // put probability into the morgue
dynProgProb.setValueFct (n_step);
dynProgProb.update (); // iterate random walk
if (time_ > 0.0)
{
Sls::alp_data::get_current_time (time1);
if (time1 - time0 > time_)
{
*terminated_ = true;
return;
}
}
}
for (value = score_ [0]; value <= entry; value++) {
if (eSumAlpha_) *eSumAlpha_ += dynProgProb.getProb (value) * static_cast <double> (value);
if (eOneMinusExpSumAlpha_) *eOneMinusExpSumAlpha_ += dynProgProb.getProb (value) *
(1.0 - exp (lambda_ * static_cast <double> (value)));
}
// check that not too much probability has been omitted
double prob = 0.0;
for (value = entry + 1; value < dynProgProb.getValueUpper (); value++) {
prob += dynProgProb.getProb (value);
}
prob += dynProgProb.getProbLost ();
const double FUDGE = 2.0;
assert (prob <= FUDGE * static_cast <double> (dimension_) * REL_TOL);
}
