Theorem 6: Causal IPOMDP's: Any infra-POMDP with transition kernel  which fulfills the niceness conditions and always produces crisp infradistributions, and starting infradistribution , produces a causal belief function via .

So, first up, we can appeal to Theorem 4 on Pseudocausal IPOMDP's, to conclude that all the  are well-defined, and that this  we defined fulfills all belief function conditions except maybe normalization.

The only things to check are that the resulting  is normalized, and that the resulting  is causal. The second is far more difficult and makes up the bulk of the proof, so we'll just dispose of the first one.

T6.1 First, , since it's crisp by assumption, always has  for all constants . The same property extends to all the  (they're all crisp too) and then from there to  by induction. Then we can just use the following argument, where  is arbitrary.

The last two inequalities were by  mapping constants to the same constant regardless of , and by  being normalized since it's an infradistribution. This same proof works if you swap out the 1 with a 0, so we have normalization for 

T6.2 All that remains is checking the causality condition. We'll be using the alternate rephrasing of causality. We have  defined as

and want to derive our rephrasing of causality, that for any  and  and  and family , we have

Accordingly, we get to fix a , and family of functions , and assume

Our proof target is

Applying our reexpression of  in terms of  and  and projection, we get

Undoing the projections on both sides, we get

Unpacking the semidirect product, we get

Now, we can notice something interesting here. By concavity for infradistributions, we have

So, our new proof target becomes

Because if we hit that, we can stitch the inequalities together. The obvious move to show this new result is to apply monotonicity of infradistributions. If the inner function on the left hand side was always above the inner function on the right hand side, we'd have our result by infradistribution monotonicity. So, our new proof target becomes

At this point, now that we've gotten rid of , we can use the full power of crisp infradistributions. Crisp infradistributions are just a set of probability distributions! Remember that the semidirect product, , can be viewed as taking a probability distribution from , and for each , picking a conditional probability distribution from , and all the probability distributions of that form make up .

In particular, the infinitely iterated semidirect product here can be written as a repeated process where, for each history of the form  Murphy gets to select a probability distribution over the next observation and state from  (as a set of probability distributions) to fill in the next part of the action-observation-state tree. Accordingly every single probability distribution in  can be written as  interacting with an expanded probabilistic environment  of type  (which gives Murphy's choice wherever it needs to pick the next observation and state), subject to the constraints that

and

We'll abbreviate this property that  always picks stuff from the appropriate  set and starts off picking from  as 

Our old proof goal of

can be rephrased as

and then rephrased as

And now, by our earlier discussion, we can rephrase this statement in terms of picking expanded probabilistic environments.

Now, on the right side, since the set we're selecting the expanded environments from is the same each time and doesn't depend on , we can freely move the inf outside of the expectation. It is unconditionally true that

So now, our proof goal shifts to

because if that were true, we could pair it up with the always-true inequality to hit our old proof goal. And we'd be able to show this proof goal if we showed the following, more ambitious result.

Note that we're not restricting to compatibility with  and , we're trying to show it for any expanded probabilistic environment of type . We can rephrase this new proof goal slightly, viewing ourselves as selecting an action and observation history.

Now, for any expanded probabilistic environment , there is a unique corresponding probabilistic environment  of type  where, for any policy, . This is just the classical result that any POMDP can be viewed as a MDP with a state space consisting of finite histories, it still behaves the same. Using this swap, our new proof goal is:

Where  is an arbitrary probabilistic environment. Now, you can take any probabilistic environment  and view it as a mixture of deterministic environments . So we can rephrase  as, at the start of time, picking a  from  for some particular . This is our mixture of deterministic environments that makes up . This mixture isn't necessarily unique, but there's always some mixture of deterministic enivironments that works to make  where  is arbitrary. So, our new proof goal becomes:

Since  is just a dirac-delta distribution on a particular history, we can substitute that in to get the equivalent

Now we just swap the two expectations.

And then we remember that our starting assumption was

So we've managed to hit our proof target and we're done, this propagates back to show causality for .

 

Theorem 7: Pseudocausal to Causal Translation: The following diagram commutes when the bottom belief function  is pseudocausal. The upwards arrows are the causal translation procedures, and the top arrows are the usual morphisms between the type signatures. The new belief function  is causal. Also, if , and  is defined as  for any history containing Nirvana, and  for any nirvana-free history, then .

T7.0 Preliminaries. Let's restate what the translations are. The notation  is that "defends"  defends  if it responds to any prefix of  which ends in an action with either the next observation in , or Nirvana, and responds to any non-prefixes of  with Nirvana.

Also, we'll write  if  is "compatible" with  and . The criterion for this is that , and that any prefix of  which ends in an action and contains no Nirvana observations must either be a prefix of , or only deviate from  in the last action. An alternate way to state that is that  is capable of being produced by a copolicy which defends , interacting with .

With these two notes, we introduce the infrakernels , and .

 is total uncertainty over  such that  defends 

And  is total uncertainty over  such that  is compatible with  and 

Now, to briefly recap the three translations.For first-person static translation, we have

Where  is the original belief function.

For third-person static translation, we have

Where  is the original infradistribution over .

For first-person dynamic translation, the new infrakernel has type signature

Ie, the state space is destinies plus a single Nirvana state, and the observation state is the old space of observations plus a Nirvana observation.  is used for the Nirvana observation, and  is used for the Nirvana state. The transition dynamics are:

and, if , then 

and otherwise, 

The initial infradistribution is , the obvious injection of  from  to .

The obvious way to show this theorem is that we can assume the bottom part of the diagram commutes and  are all pseudocausal, and then show that all three paths to  produce the same result, and it's causal. Then we just wrap up that last part about the utility functions.

Fortunately, assuming all three paths produce the same  belief function, it's trivial to show that it's causal.  is an infradistribution since  is, and  is a crisp infrakernel, so we can just invoke Theorem 6 that crisp infrakernels produce causal belief functions to show causality for . So we only have to check that all three paths produce the same result, and show the result about the utility functions. Rephrasing one of the proofs is one phase, and then there are two more phases showing that the other two phases hit the same target, and the last phase showing our result about the utility functions.

 We want to show that

T7.1 First, let's try to rephrase  into a more handy form for future use. Invoking our definition of , we have:

Then, we undo the projection

And unpack the semidirect product

And unpack what  means

And then apply the definition of 

This will be our final rephrasing of . Our job is to derive this result for the other two paths.

T7.2 First up, the third-person static translation. We can start with definitions.

And then, we can recall that since  and  are assumed to commute, we can define  in terms of  via . Making this substitution, we get

We undo the first pushforward

And then the projection

And unpack the semidirect product

And remove the projection

and unpack the semidirect product and what  is

Now, we recall that  is all the  that defend , to get

At this point we can realize that for any  which defends , when  interacts with , it makes a  that is compatible with  and . Further, all  compatible with  and  have an  which produces them. If  has its nirvana-free prefix not being a prefix of , then the defending  which doesn't end  early, just sends deviations to Nirvana, is capable of producing it. If  looks like  but ended early with Nirvana, you can just have a defending  which ends with Nirvana at the same time. So we can rewrite this as

And this is exactly the form we got for .

T7.3 Time for the third and final translation. We want to show that

is what we want. Because the pseudocausal square commutes, we have that , so we can make that substitution to yield

Remove the projection, to get

For this notation, we're using  to refer to an element of , the state space.  is an element of , a full unrolling.  is the projection of this to .

We unpack the semidirect product to yield

and rewrite the injection pushforward

and remove the projection

And unpack the semidirect product along with .

Now we rewrite this as a projection to yield

And do a little bit of reindexing for convenience, swapping out the symbol  for , getting

And now, we can ask what  is. Basically, the starting state is the destiny , and then  interacts with it through the kernel . Due to how  is defined, it can do one of three things. First, it could just unroll  all the way to completion, if  is the sort of history that's compatible with . Second, it could partially unroll  and deviate to Nirvana early, even if  is compatible with more, because of how  is defined. Finally,  could deviate from  and then  would have to respond with Nirvana. And of course, once you're in a Nirvana state, you're in there for good. Hm, you have total uncertainty over histories  compatible with  and , any of them at all could be made by . So, this projection of the infrakernel unrolling  is just . We get

And then we use what  does to functions, to get

Which is exactly the form we need. So, all three translation procedures produce identical belief functions.

T7.4 The only thing which remains to wrap this result up is guaranteeing that, if  is 1 if the history contains Nirvana, (or infinity for the  type signature), and  otherwise, then .

We rephrased  as 

We can notice something interesting. If  is the sort of history that  is capable of making, ie , then all of the  are going to end up in Nirvana except  itself, because all the  are like "follow , any deviations of  from  go to Nirvana, and also Nirvana can show up early". In this case, since any history ending with Nirvana is maximum utility, and otherwise  is copied, we have . However, if  is incompatible with , then no matter what, a history  which follows  is going to hit Nirvana and be assigned maximum utility. So, we can rewrite

as

Which can be written as an update, yielding

And then, we remember that regardless of  and , for pseudocausal belief functions, which  is, we have that , so the update assigns higher expectation values than  does. Also, . Thus, the infinimum is attained by , and we get

and we're done.

 

Theorem 8: Acausal to Pseudocausal Translation: The following diagram commutes when the bottom belief function  is acausal. The upwards arrows are the pseudocausal translation procedures, and the top arrows are the usual morphisms between the type signatures. The new belief function  will be pseudocausal.

T8.0 First, preliminaries. The two translation procedures are, to go from an infradistribution over policy-tagged destinies , to an infradistribution over destinies , we just project. 

And, to go from an acausal belief function  to a pseudocausal one , we do the translation

We assume that  and  commute, and try to show that these translation procedures result in a  and  which are pseudocausal and commute with each other. So, our proof goals are that

and

Once we've done that, it's easy to get that  is pseudocausal. This occurs because the projection of any infradistribution is an infradistribution, so we can conclude that  is an infradistribution over . Also, back in the Pseudocausal Commutative Square theorem, we were able to derive that any infradistribution  over , when translated to a , makes a pseudocausal belief function, no special properties were used in that except being an infradistribution.

T8.1 Let's address the first one, trying to show that

We'll take the left side, plug in some function, and keep unpacking and rearranging until we get the right side with some function plugged in. First up is applying definitions.

by how  was defined. Now, we start unpacking. First, the projection.

Then unpack the semidirect product and , for

Then unpack the update, to get

Then pack up the 

And pack up the semidirect product

And since  for the original acausal belief function, we can rewrite this as

And pack this up in a projection.

And then, since we defined , we can pack this up as

And then pack up the update

So, that's one direction done.

T8.2 Now for the second result, showing that

Again, we'll take the more complicated form, and write it as the first one, for arbitrary functions .bWe begin with

and then unpack what  is to get

Undo the projection

Unpack the semidirect product and .

Remove the projection

Unpack the semidirect product and  again.

Let's fold the two infs into one inf, this doesn't change anything, just makes things a bit more concise.

and unpack the update.

Now, we can note something interesting here. If  is selected to be equal to , that inner function is just , because  is only supported over histories compatible with . If  is unequal to , that inner function is  with 1's for some inputs, possibly. So, our choice of  is optimal for minimizing when its equal to , so we can rewrite as

and write this as 

And pack up as a semidirect product

and then because , we substitute to get

and then write as a projection.

Now, this is just , yielding

and we're done.

 

Proposition 2: The two formulations of x-pseudocausality are equivalent for crisp acausal belief functions.

One formulation of x-pseudocausality is that

 where 

The second formulation is,

This proof will be slightly informal, it can be fully formalized without too many conceptual difficulties. We need to establish some basic probability theory results to work on this. The proof outline is we establish/recap four probability-theory/inframeasure-theory results, and then show the two implication directions of the iff statement.

P2.1 Our first claim is that, given any two probability distributions  and , you can shatter them into measures , and  such that  and , and . What this decomposition is basically doing is finding the largest chunk of measure that  and  agree on (that's ), and then subtracting that out from  and  yields you your  and . These measures have no overlap, because if there was overlap, you could put that overlap into the  chunk of measure. And, the mass of the "unique to " piece must equal the mass of the "unique to " piece because they must add up to probability distributions, and this quantity is exactly the total variation distance of .

P2.2 Our second claim is that 

Why is this? Well, there's an alternate formulation of total variation distance where

As intuition for this, we can pick a function U which is 1 on the support of , and 0 elsewhere. In such a case, we'd get

and then remember that  and  have disjoint supports, so we get

And any other function in the  range wouldn't do as well at attaining different expectation values on  and . So that's where the identity comes from. If we try to do the optimal separation but instead restrict  to be in , we'd get

and any other utility function in  wouldn't do as well at separating the two.

P2.3 Our third claim is that the function which maps  to , if  goes outside of , and otherwise maps  to , corresponds to taking the set of sa-measures that is  (the set form of ), and taking the upper completion of it under all signed-measure, b pairs where . Call this an x-sa measure. Ie, if , then one unit of +b is big enough to cancel out 3 units of negative-measure in the signed measure. Clearly, if , and  has  (is an x-sa measure) then we have

So, adding in the cone of all x-sa pairs like that doesn't actually affect the expectation values of any function bounded in  at all. However, for any function which goes over the x upper-bound, you can consider a x-sa measure which concentrates all its large amount of negative measure on a point where the function exceeds the upper bound, and exactly compensates for it with the +b term. Adding this on can make the function, evaluated at said point, have an unboundedly negative expectation value, justifying our choice of  for the expectation value.

P2.4 Time for the fourth claim. To build up to it, the monotone hull of a functional  is as follows: if  is the original functional, then  is the monotone hull defined by 

Our fourth claim is that, in the set-based value, this operation corresponds to restricting the set of x-sa measures that is  to just the a-measures in it. Why? Well, we have

This is because  and  both agree that any function with negative parts gets  value, and both agree on the expectation of any function bounded in , but for functions above 0 which also stray above  somewhere, , while .

Therefore, since  (when considered as a function, we aren't using the infradistribution reversed ordering here), we have that  (higher functions correspond to smaller sets). Now,  cannot contain any x-sa-measures with negative measure anywhere in the measure component. For, if there was such a point in there, we could consider a function  that assigned extremely high value to the region of negative measure, and 0 value elsewhere, and  would attain a negative expectation value on that x-sa measure, which contradicts the nonnegative worst-case expectation value of  due to how  was defined. So, the set  must consist entirely of a-measures. Further, for any function  with range in , it is minimized by an a-measure in , and the same point is present in  (no extra a-measures added), so the  corresponding to that set must assign the exact same expectation value to  as  and  do. Also, any set consisting entirely of a-measures must have its expectation functional be monotone.

So, recapping what we've done so far, we know that . We know that  must consist entirely of a-measures, otherwise we'd get a contradiction with monotonicity. Further,  (the biggest set that  could possibly be) has its expectation functional  agree with  on the expectations of functions in , and it's monotone like  is. Now, if  (strict subset), we'd have the functional  corresponding to the intersection being strictly below  on some function. However  is the minimal monotone function that agrees with  on the expectations of functions in , so this rule out strict subset, and we must have equality. The monotone hull of an expectation functional corresponds to the restriction to the set to a-measures. Further, we can rewrite the definition

as just

And we're done with this.

Putting claims 3 and 4 together, we can now express what the set  is, from our definition of  (interpreted as a set of sa-measures) is just the upper completion of  under x-sa measures, those with  (this is the restriction to ), and then the intersection of that set with the cone of a-measures (monotone hull), and then upper completion under sa-measures (the restriction to ).

P2.5 Time to begin showing one direction of the iff statement. Remember, we're assuming that all the  are induced by a set of probability distributions, they're crisp. We'll start with the easy direction, that

implies

where . (well, technically, this shouldn't just be continuous functions, but lower-semicontinuous functions as well) To begin with, let  be arbitrary with  bounded in . We can start unpacking

Now, this minimal value must be attained by some minimal point in , which is some probability distribution . So, we have

We can, at this point, select the closest probability distribution to  in  according to total variation distance, call that . We do our usual shattering of  into  and , the chunk of measure where  and  overlap, and the chunk of measure that's unique to . Note that the chunk of measure where  and  overlap must be supported on histories compatible with  since , so we can break things down as

And then we can break  down further into  and , for the chunk of measure on histories compatible with  and incompatible, respectively, yielding

Which can be rephrased as

Now it's time for some inequalities, explained afterwards.

So, for the first line, the first two inequalities are obvious. For the second line, the initial inequality is us invoking the total variation distance variant of x-pseudocausality since  is as close as possible to  while also being in , and the equality is reexpressing total variation distance in accordance with our second claim. The third line is just shattering  and  in our way connected with total variation distance, and the fourth line is just canceling. For the fifth line, the first equality is the supremum being attained by a function that's  on , and 0 on , as the two measures are disjoint, and the second equality is . Then, for line 6, the inequality is obvious because , and then we just wrap the inf up as an expectation according to , and apply the definition of .

P2.6 Now for the other direction of the iff-statement, going from

to

Using our set-based characterization of inframeasure ordering, we can accurately rephrase our initial statement as

The former set is "take all the minimal points of , convert all their measure on histories incompatible with  to the +b term, do closed convex hull and upper completion", and the latter set is "take , do upper completion w.r.t. the cone of x-sa measures, restrict to just the a-measures, and then add in the cone of sa-measures again"

Pick a . When we 1-update it on compatibility with , we get . And said point lies in , which means it must be possible to create by taking a minimal point in  (call that one ), and adding an x-sa measure to it. Call the x-sa measure . So, our net result is:

The update of  equals  plus an x-sa measure. Now, breaking up  into its positive and negative parts which are disjoint,  and , we can rewrite things as:

Adding  to both sides and folding it into , and rearranging a bit and adding parentheses, that second equality can be restated as

Now,  is supported entirely over histories compatible with , and  is a chunk of negative measure taken out of , so  must be supported entirely over histories compatible with . Also,  and  are disjoint (Jordan decomposition), and   is supported entirely over histories incompatible with , so the positive measure  and the negative measure  must be disjoint. Since  is made from  by adding on a chunk of positive measure and a chunk of negative measure which are disjoint, we have . At this point, we only have one step left.

For the last step, because  is an x-sa measure, we have . Also, because , our inequality turns into

and this gets us our

result. We're done!

 

Proposition 3: If a utility function  is bounded within  and a belief function  is x-pseudocausal, then translating  to  (pseudocausal translation) fulfills .

Blessedly, this one is very easy. First, let's recap the translation of a belief function (assumed to be x-pseudocausal) from acausal to pseudocausal. It is

So, applying this, we have

And then we recall that x-pseudocausality is  Also, recall that since , and our utility function is bounded in , we get the statement  Additionally, . Therefore, the infinimum is attained by  itself, and we get

and we're done!

New Comment