Original quadratic teaching model
A summary is a function.
Change only the external target. The cached child summary stays fixed, but both estimates move.
Cache once
EL(u,s) = ½(u − 1)² + ½(s − u)²
mL(s) = minu EL = ¼(s − 1)²
u*(s) = (s + 1) / 2
EL(u,s) = ½(u − 1)² + ½(s − u)²
mL(s) = minu EL = ¼(s − 1)²
u*(s) = (s + 1) / 2
New external evidence
ER(s;z) = ½(s − z)²
s* = arg mins [mL + ER]
s* = (1 + 2z) / 3
ER(s;z) = ½(s − z)²
s* = arg mins [mL + ER]
s* = (1 + 2z) / 3
separator s*1.00
recovered child u*1.00
Min-sum analogy, not a full iSAM2 implementation. The root has no separator; s is the child’s separator. No data is sent to a server.
Root decision → child reconstruction
cached mL(s)external ER(s;z)sum