Question

Assume that the cost function $E(\underline{x})$ takes integer values and let $U(\beta)=\langle E(\underline{x})\rangle_\beta$. Due to the form (13.6) the computation of $U(\beta)$ is essentially equivalent to statistical inference. Assume, furthermore that $\Delta_{\text {max }}= \max \left[E(\underline{x})-E_*\right]$ is bounded by a polynomial in $N$. Show that $$ 0 \leq \frac{\partial U}{\partial T} \leq \frac{1}{T^2} \Delta_{\max }^2|\mathcal{X}|^N e^{-1 / T} $$ where $T=1 / \beta$. Deduce that, by taking $T=\Theta(1 / N)$, one can obtain $\mid U(\beta)- E_* \mid \leq \varepsilon$ for any fixed $\varepsilon>0$.

    Assume that the cost function $E(\underline{x})$ takes integer values and let $U(\beta)=\langle E(\underline{x})\rangle_\beta$. Due to the form (13.6) the computation of $U(\beta)$ is essentially equivalent to statistical inference. Assume, furthermore that $\Delta_{\text {max }}= \max \left[E(\underline{x})-E_*\right]$ is bounded by a polynomial in $N$. Show that

$$
0 \leq \frac{\partial U}{\partial T} \leq \frac{1}{T^2} \Delta_{\max }^2|\mathcal{X}|^N e^{-1 / T}
$$

where $T=1 / \beta$. Deduce that, by taking $T=\Theta(1 / N)$, one can obtain $\mid U(\beta)- E_* \mid \leq \varepsilon$ for any fixed $\varepsilon>0$.
Show more…
Information, Physics, and Computation
Information, Physics, and Computation
Marc MĆ©zard, Andrea… 1st Edition
Chapter 13, Problem 1 ↓
AceChat toggle button
Close icon
Ace pointing down

Please give Ace some feedback

Your feedback will help us improve your experience

Thumb up icon Thumb down icon
Thanks for your feedback!
Profile picture
Assume that the cost function $E(\underline{x})$ takes integer values and let $U(\beta)=\langle E(\underline{x})\rangle_\beta$. Due to the form (13.6) the computation of $U(\beta)$ is essentially equivalent to statistical inference. Assume, furthermore that $\Delta_{\text {max }}= \max \left[E(\underline{x})-E_*\right]$ is bounded by a polynomial in $N$. Show that $$ 0 \leq \frac{\partial U}{\partial T} \leq \frac{1}{T^2} \Delta_{\max }^2|\mathcal{X}|^N e^{-1 / T} $$ where $T=1 / \beta$. Deduce that, by taking $T=\Theta(1 / N)$, one can obtain $\mid U(\beta)- E_* \mid \leq \varepsilon$ for any fixed $\varepsilon>0$.
Close icon
Play audio
Feedback
Powered by NumerAI
*

Labs

-

Want to see this concept in action?

NEW

Explore this concept interactively to see how it behaves as you change inputs.

View Labs

*

Recommended Videos

-
q-x-n-pur-0-1xur-iojqeur-mous-e-eetx-e8tb-a-where-gu-log1-y-ye-yu-with-y-aa-b-hint-write-x-as-a-convex-combination-of-a-and-b-where-the-convex-weighting-parameter-depends-upon-x-exploit-the-56909

(a) Show that for any t > 0 and a ≤ X ≤ b, E(etX) ≤ eg(t(b-a)), where g(u) = log(1 - ̳ + ̳eu) - ̳u, with ̳ = a/(a - b). (Hint: write X as a convex combination of a and b, where the convex weighting parameter depends upon X. Exploit the convexity of the function x → etx and the fact that inequalities are preserved upon taking expectations of both sides, since expectations are integrals.) (b) By using Taylor’s theorem, show that for all u > 0, g(u) ≤ u^2 / 8. Furthermore show that E(etX) ≤ e^(t^2(b-a)^2 / 8).

Need help? Use Ace
Ace is your personal tutor. It breaks down any question with clear steps so you can learn.
Start Using Ace
Ace is your personal tutor for learning
Step-by-step explanations
Instant summaries
Summarize YouTube videos
Understand textbook images or PDFs
Study tools like quizzes and flashcards
Listen to your notes as a podcast
Continue solving this problem
Create a free account to:
  • View full step-by-step solution
  • Ask follow-up questions with Ace AI
  • Save progress and study later
Continue Free
Join the community

18,000,000+

Students on Numerade


Trusted by students at 8,000+ universities

Numerade

Get step-by-step video solution
from top educators

Continue with Clever
or



By creating an account, you agree to the Terms of Service and Privacy Policy
Already have an account? Log In

A free answer
just for you

Watch the video solution with this free unlock.

Numerade

Log in to watch this video
...and 100,000,000 more!


EMAIL

PASSWORD

OR
Continue with Clever