Add rotation cost model #461
Conversation
|
|
||
|
|
||
| @frozen | ||
| class LogarithmicModel(RotationCostModel): |
There was a problem hiding this comment.
This name makes sense here, where you can clearly see that it's an instance of RotationCostModel but I suspect if it shows up outside of the class definition it will lose this critical context. Should probably have rotation in the name somewhere
|
|
||
|
|
||
| @frozen | ||
| class ConstantCostWithPreparation(RotationCostModel): |
There was a problem hiding this comment.
again, in actual use the fact that this is for rotations is lost; but we don't want it to be too long.
ConstOverheadRotationCost or something like that
|
|
||
| Attributes: | ||
| bitsize: Number of digits of accuracy for approximating a rotation. | ||
| slope: The coefficient of $log_2{budget}$. |
There was a problem hiding this comment.
why is there a slope? isn't this supposed to be constant?
There was a problem hiding this comment.
for computing the cost of the overhead rotation
There was a problem hiding this comment.
I changed the signature to make this clearer
| ) | ||
|
|
||
|
|
||
| BeverlandEtAl = LogarithmicModel( |
There was a problem hiding this comment.
Outside of this definition, this variable will appear without the context that it's a rotation cost model
Can you speak more about this? |
|
@mpharrigan to use TComplexity I will have to convert the toffolis to Ts The rotations are implemented in Toffolis so I think the runtime estimate will defer between an architectures that synthesize magic T state vs magic CCZ states. |
|
oh for some reason I thought we had a toffoli/ccz field in the |
This abstraction covers both standarad models which have a cost of$A \log_2(1/\epsilon) + B$ Where $A$ $B$ depend on the gateset, approximation protocol and hardware parameters and are obtained by fitting regression models. It also covers models with constant per rotation of cost which require a circuit that gets prepared once which gives preparation overhead.
and
I'm reporting the cost as an
AlgorithmSummaryobject rather than aTComplexityobject because it's more natural to report the cost of the second model in terms of Toffoli count rather than T count.Related to #370 and is part of #269