Editing Open Problems:84
Warning: You are not logged in. Your IP address will be publicly visible if you make any edits. If you log in or create an account, your edits will be attributed to your username, along with other benefits.
The edit can be undone.
Please check the comparison below to verify that this is what you want to do, and then save the changes below to finish undoing the edit.
Latest revision | Your text | ||
Line 2: | Line 2: | ||
|source=focs17 | |source=focs17 | ||
|who=Alon Orlitsky | |who=Alon Orlitsky | ||
+ | |title=Efficient Profile Maximum Likelihood Computation | ||
}} | }} | ||
β | Given a sequence of samples $\mathbf{s}=(s_1,\dots,s_n)\in\mathbb{N}^n$, the ''sequence maximum likelihood'' (SML) estimator is the probability distribution over $\mathbb{N}$ | + | Given a sequence of samples $\mathbf{s}=(s_1,\dots,s_n)\in\mathbb{N}^n$, the ''sequence maximum likelihood'' (SML) estimator is the probability distribution over $\mathbb{N}$ which maximizes the probability of this sequence, i.e. |
$$ | $$ | ||
p^{\rm SML} \stackrel{\rm def}{=} \arg\!\!\!\max_{p\in\Delta(\mathbb{N})} p^{\otimes n}(\mathbf{s}) | p^{\rm SML} \stackrel{\rm def}{=} \arg\!\!\!\max_{p\in\Delta(\mathbb{N})} p^{\otimes n}(\mathbf{s}) | ||
Line 19: | Line 20: | ||
p^{\rm PML} \stackrel{\rm def}{=} \arg\!\!\!\max_{p\in\Delta(\mathbb{N})} \sum_{\mathbf{s'}: \Phi(\mathbf{s'})=\Phi(\mathbf{s})} p^{\otimes n}(\mathbf{s'})\,. | p^{\rm PML} \stackrel{\rm def}{=} \arg\!\!\!\max_{p\in\Delta(\mathbb{N})} \sum_{\mathbf{s'}: \Phi(\mathbf{s'})=\Phi(\mathbf{s})} p^{\otimes n}(\mathbf{s'})\,. | ||
$$ | $$ | ||
β | The PML is particularly well-suited to dealing with symmetric properties and functionals of distributions (i.e., those invariant | + | The PML is particularly well-suited to dealing with symmetric properties and functionals of distributions (i.e., those invariant by relabeling of the domain), as shown in {{cite|AcharyaDOS-17}}. In particular, in the sublinear sample regime, it provably outperforms the SML. However, from a computational point of view, it is unclear whether one can compute it efficiently. |
β | Is there a polynomial (or even strongly subexponential) time algorithm to compute | + | Is there a polynomial- (or even strongly subexponential)-time algorithm to compute the PML? To (multiplicatively) approximate it? |