# Modibo : Zoë Thoughts

## Assorted links
- "Let juries review facebook ads" -- arguing for random juries to https://www.theatlantic.com/ideas/archive/2019/11/let-juries-review-facebook-ads/601996/
- Drexl Kleiner -- on the efficiency of majority voting: https://www.dropbox.com/s/jq4al78bzusblff/Drexl-Kleiner%20-%20Why%20Voting%201117.pdf?dl=0

We show that whenever an anonymous decisionrule conditions on preference intensities, money is lost, which introduces a trade-off for the agents: they can choose a “good” decision rule, but then they will lose money.
- Bierbrauer Hellwig - https://www.dropbox.com/s/2kjjruy7ytxdt5v/Bierbrauer%20Hellwig%202016%20-%20robustly_coalition_proof_incentive_mechanisms_for_public_good_provision_are_voting_mechanisms_and_vice_versa.pdf?dl=0
- A not very thorough lit review of some recent voting papers I made a few months ago
https://www.dropbox.com/s/33hk6d9hqzgpk0g/recent%20voting%20lit%20common%20values.pdf?dl=0
- One other paper underappreciation of general equilibrium in voting: https://www.dropbox.com/s/xieeibfe2m5ilpk/Dal%20Bo%20Dal%20Bo%20and%20Eyster%202017%20-%20Demand%20for%20bad%20policy.pdf?dl=0

## Sortition / random sample voting

- could read and discuss crazy David Chaum whitepaper and other papers on here https://rsvoting.org/
- check out Ariel Procaccia papers on sortition

Random sample VCG vs. full VCG
- less susceptible to collusion?
- how sensitive to distribution are bounds on efficiency

> [name=mcamara] key sensitivities are to (a) an upper bound on utility functions and (b) the "complexity"/"dimension" of the social choice problem. If (a) and (b) are favorable, then bounds are uniform across all distributions.
## "Optimal matching funds"
- is there some unified way of thinking about "Matching Funds" as a way of combining centralized and decentralized funding?
- how does this depend on network structure/shape of communities?
- notes: https://www.dropbox.com/s/ib5qay1nwlknnnl/hitzig_matching_funds_march2020.pdf?dl=0

Network $G = V, E$, |V|

## Is there another way to do robustness?

Something that's not min max?
* minimax regret (favored by Chuck Manski, some use in micro theory but not as common as minmax)
* approximation ratio (favored by computer scientists); applied to prior-independent and prior-free mechanism design.
* probably approximately correct (PAC) in statistics (like minmax regret, but usually settles for optimal rates of convergence)
* replacing prior knowledge with data (favored by me, where possible; see e.g. literature on sample complexity of auction design; see e.g. literature that applies online learning to auction design)
* "preparing for the worst but hoping for the best: robust (Bayesian) persuasion"; it's kind of a refinement of maxmin; not sure this approach generalizes all that well
* Tilman Boergers has an approach to robust mechanism design that focuses on characterizing the set of admissible (i.e. not weakly dominated) mechanisms as opposed to maxmin. Also has work showing that some maxmin mechanisms derived in previous literature are weakly dominated.
* Maybe there's some room for incorporating human judgement? As in, let people "fill in the blanks" where the theory is unable to say much. I imagine the sort of relationship between diagnostic algorithms and physicians; we still make use of physician's expert knowledge even though it's not easy to distill in a formal way.

## Mechanisms with unknown/learned network structure

> [name=mcamara] Zoë mentioned local public goods, as well as papers in CS that are concerned with learning network structure.
>
> As an example, take Ben Golub's paper on targeting interventions in networks. IIRC the paper identifies some properties of the network that affect what the optimal intervention looks like.
>
> There is a question of whether these properties are aggregate properties that can be learned from a reasonable number of samples, or if you really need to know details of the entire network structure.
>
> Given that question, it's conceivable that the kinds of mechanisms or interventions that you would design differ qualitatively between case (a) where you know the network and case (b) where you have sample access to the network. Then there's the other consideration that certain network properties may be hard to learn (i.e. require too many samples) when the network is very large.
>
> In any case, I'd be curious to know more about how social networks affect the public goods problem, etc. The only related problem I've thought about before is regulation of financial networks.

Network structure determines patterns of altruistic preferences
* Identification: if we know the network structure, can we separate altruistic preferences from correlated preferences within each group? Under what conditions? Related to collusion detection -- is there a literature on that?
* What is my "personal" or "nonaltruistic" preference? Is it identified by some kind of choice data? How my preferences change as the consequences to another agent change?
* This question of disentangling correlation from altruism seems like it should be relevant beyond mechanism design.
* Look into empirical/experimental literature on altruism.
* Is altruism a response to the inefficiency of e.g. majority voting? Can we convince people to be less altruistic if our mechanism is efficient?
* One can also think about robustness to altruism, e.g. how performance changes as people become more altruistic. Related to adversarial corruptions approach in machine learning.
* What aspects of this problem are most interesting to economists? Related to Zoë's comment on finding the sweet spot between Ben's Interventions paper and Mohammad's paper on seeding.
* We discussed the difference between normalized and not-normalized altruistic preferences, and I agree that the latter makes more sense. But suppose that it's true that, if utility is normalized and altruism is symmetric ($i$ cares about $j$ as much as $j$ cares about $i$), then maximizing altruistic preferences is efficient with respect to (normalized) welfare. Normalizing utility seems similar to one-person-one-vote. This leads to the following question: although 1P1V is clearly inefficient when preferences are not altruistic, is it more robust to altruism than other mechanisms?
* See "Other-regarding Preferences and Redistributive Politics" at https://sites.google.com/site/juliensenn/