experiments/sb_reference_validation_20261004/continuous_kernel_checks.py

"""Independent continuous h-transform checks, including typed mixture marks.

This is a second implementation of Gaussian formulas, not a call into the
production emission/teacher functions. The gap oracle explicitly enumerates
current-to-terminal subsequences, including the rank/binomial factor.
"""
import json
from pathlib import Path
import numpy as np
from scipy.special import logsumexp

from oracle_kernel import gap_remaining_loglik


def logn(x, mean, variance):
    x,mean=np.asarray(x),np.asarray(mean)
    return float(-.5*(x.size*np.log(2*np.pi*variance)+np.square(x-mean).sum()/variance))


def run_checks():
    rng=np.random.default_rng(60341)
    t=.43; horizon=1.; delta=horizon-t; sigma2=1.7; v0=2.3; beta=.9
    # Real atom inventories: A=5, G=4; all matrices consist of actual atoms.
    target_types=['A','G','A']
    target=[rng.normal(0,.4,(5,3)),rng.normal(0,.4,(4,3)),rng.normal(0,.4,(5,3))]
    pi={'A':.6,'G':.4}; mu={aa:np.zeros((count,3)) for aa,count in [('A',5),('G',4)]}
    current_types=['A']; current=[rng.normal(0,.4,(5,3))]
    birth=np.array([np.log(pi[a])+logn(y,mu[a],v0+horizon*sigma2)
                    for a,y in zip(target_types,target)])

    def likelihood(types,coords):
        emissions=np.full((len(types),3),-np.inf)
        for i,(a,x) in enumerate(zip(types,coords)):
            for j,(b,y) in enumerate(zip(target_types,target)):
                if a==b: emissions[i,j]=logn(y,x,delta*sigma2)
        return float(gap_remaining_loglik(emissions,birth,beta,delta))

    h=likelihood(current_types,current)
    # Enumerate which terminal A descends from the single present A.
    assignment_logs=[]
    for existing in [0,2]:
        assignment_logs.append(logn(target[existing],current[0],delta*sigma2)
                               +sum(birth[j] for j in range(3) if j!=existing))
    weights=np.exp(assignment_logs-logsumexp(assignment_logs))
    drift=sum(w*(target[j]-current[0])/delta for w,j in zip(weights,[0,2]))
    numerical=np.zeros_like(drift); eps=1e-5
    for idx in np.ndindex(drift.shape):
        plus=current[0].copy();minus=plus.copy();plus[idx]+=eps;minus[idx]-=eps
        numerical[idx]=sigma2*(likelihood(['A'],[plus])-likelihood(['A'],[minus]))/(2*eps)
    drift_error=float(abs(drift-numerical).max())
    flux_errors=[]; ratios=[]
    for rank in [0,1]:
        for aa in ['A','G']:
            # Several positions can contribute to one slot's marked intensity.
            for _ in range(8):
                mark=rng.normal(0,.4,mu[aa].shape)
                types=current_types.copy();coords=current.copy()
                types.insert(rank,aa);coords.insert(rank,mark)
                ref=np.log(beta/2)+np.log(pi[aa])+logn(mark,mu[aa],v0+t*sigma2)
                lhs=np.exp(ref+likelihood(types,coords)-h)
                rhs=0.
                for w,existing in zip(weights,[0,2]):
                    for j in range(3):
                        if j==existing or target_types[j]!=aa or int(j>existing)!=rank:continue
                        vt=v0+t*sigma2;vT=v0+horizon*sigma2
                        mean=mu[aa]+vt/vT*(target[j]-mu[aa]);var=vt-vt*vt/vT
                        rhs+=w/delta*np.exp(logn(mark,mean,var))
                flux_errors.append(abs(lhs-rhs)/max(lhs,rhs,1e-200))
                if rhs>0:ratios.append(lhs/rhs)

    # Substitution A -> W: current backbone retained, sidechain dimension 1 -> 10 atoms.
    q=.7;s_b=.35/19
    xb=rng.normal(size=(4,3));yb=rng.normal(size=(4,3))
    yw=rng.normal(0,.4,(10,3));mw=np.zeros_like(yw);mark=rng.normal(0,.4,yw.shape)
    pending=np.log(s_b/q)+np.log(-np.expm1(-q*delta))+logn(yb,xb,delta*sigma2)+logn(yw,mw,v0+horizon*sigma2)
    locked=logn(yb,xb,delta*sigma2)+logn(yw,mark,delta*sigma2)
    physical_density=np.exp(np.log(s_b)+logn(mark,mw,v0+t*sigma2)+locked-pending)
    vt=v0+t*sigma2;vT=v0+horizon*sigma2
    expected_density=q/(-np.expm1(-q*delta))*np.exp(logn(mark,mw+vt/vT*(yw-mw),vt-vt*vt/vT))
    substitution_error=abs(physical_density-expected_density)/max(physical_density,expected_density)
    # Zero-dimensional reset for Gly uses density exactly one.
    gly_log_density=logn(np.empty((0,3)),np.empty((0,3)),vt)
    # Deletion h is independent of coordinates; its transformed rate agrees.
    d=.35;pD=d/q*(-np.expm1(-q*delta))
    deletion_error=abs(d/pD-q/(-np.expm1(-q*delta)))
    result=dict(scope='Independent typed all-atom gap and substitution marked kernels; no neural sampler or ligand rotations.',
                gap_current_A_target_AGA=True, gap_drift_max_abs_error=drift_error,
                gap_marked_intensity_max_relative_error=float(max(flux_errors)),
                gap_positive_mark_cases=len(ratios),
                substitution_A_to_W_marked_intensity_relative_error=float(substitution_error),
                gly_zero_dimensional_log_density=gly_log_density,
                deletion_hazard_abs_error=float(deletion_error))
    assert drift_error<1e-7,result
    assert max(flux_errors)<1e-10,result
    assert substitution_error<1e-10 and gly_log_density==0 and deletion_error<1e-12,result
    return result


if __name__=='__main__':
    result=run_checks();p=Path(__file__).resolve().parent/'results/continuous_kernel_checks.json'
    p.parent.mkdir(exist_ok=True);p.write_text(json.dumps(result,indent=2)+'\n');print(json.dumps(result,indent=2))