-
What graph structures produce high-order factors that are difficult for ORBIT?
BNs produce high-order factors even for basic toy graphs because of the hidden interaction between two variables that share a child. The moralized graph gives the cliques generated by moralization that make the graph highly dense. Current representation of the energy doesn’t solve this problem -
What is the native maximum factor order of the intended backend?
2 -
If quadratization is required, how many auxiliary variables are introduced?
Depends on the exact algorithmic optimisation used but an upper bound is k (if is the order of the node), a lower bound is . -
What bias/coupling dynamic range is available?
Check note 5. -
How large does clamp strength need to be before leakage is negligible?
Check note 4. critical strength section. -
What is the relationship between epsilon smoothing and hardware saturation?
using a too small would set the energy level of the state to be too high. -
How should zero-probability CPTs be represented in energy form?
Infinite energy barrier solved via -smoothing -
Under what conditions does physical clamping match the graph-mutilated intervention reference?
never -
Under what conditions does physical clamping reduce to ordinary conditioning?
partially when lambda is finite, but as , it reproduces conditioning perfectly (see Prop. 2). -
Which queries should be rejected by the compiler because the hardware profile cannot support them faithfully?
queries that condition/try to draw causation from an impossible state. -
Does energy scaling affect the intended probability distribution?
-
When does approximate sampling introduce unacceptable inference error?
-
Which parts of DIRAC are causal-compiler problems versus hardware-compiler problems?
-
Can intervention reference compilation be automated cleanly?
-
Can the hardware dynamically reconfigure arbitrary couplings?
-
If not, how severe is virtual-mutilation leakage?
-
What is the minimal applied demonstrator topic? To be determined.
Commentary