Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Matrix B of HiPPO-LegS #11

Open
kenkenpa2126 opened this issue Feb 7, 2024 · 1 comment
Open

Matrix B of HiPPO-LegS #11

kenkenpa2126 opened this issue Feb 7, 2024 · 1 comment

Comments

@kenkenpa2126
Copy link

I have a question concerning the derivation found in Appendix D.1.2 of the LSSL paper, specifically equation (18), which is given as:

$$ \begin{align*} \dot{x}_n (t) &= u(t) p_n(t,t) \omega (t, t) + \int^t_0 u(Y) (\frac{\partial}{\partial t} p_n (t, Y)) \omega (t, Y) dY + \int^t_0 u(Y) p_n (t, Y) (\frac{\partial}{\partial t} \omega (t, Y) ) dY \end{align*} $$

(Correctly, $u(Y)$ in the third term of the equation, which was denoted as $f(Y)$ in the paper. I thought this is a typo, considering that the paper introduces $h(t, Y) =u(Y) p_n(t,Y) \omega (t, Y)$.)

The first term is interpreted as follows to derive the HiPPO-LegS ODE:

$$ u(t) p_n(t,t) \omega (t, t) = (2n+1)^{\frac{1}{2}} t^{-1} u(t) $$

However this term was omitted in the HiPPO paper's Appendix D.3 derivation of the HiPPO-LegS, I think that including this term modifies the coefficient $B$ from $B = (2n+1)^{\frac{1}{2}}$ to $B = 2(2n+1)^{\frac{1}{2}}$.

Could you let me know about this discrepancy between these derivations?

@albertfgu
Copy link
Contributor

I don't remember exactly details of the LSSL derivation. However, I am sure that the HIPPO formulas are correct. Note that this repo is deprecated: I would check out the repo github.com/state-spaces/s4 instead. There you can find visualizations under notebooks/ of the HIPPO methods showing that their formulas are correct.

Note that the LSSL paper is also "deprecated": its derivations are generalized and simplified by followups, including the S4 paper (https://arxiv.org/abs/2111.00396) and HTTYH paper (https://arxiv.org/abs/2206.12037). The latter in particular should have more correct derivations of the HIPPO-LegS ODE.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants