Strategic Resource Selection with Homophilic Agents

My strongest case is a narrow one: the paper has a very natural population-continuization of its impact-blind equilibrium result. I would anchor it on Theorem 5, proved in this paper, which states that Algorithm 1 computes an IBE in runtime \(O(m+k\log k)\). I would not anchor the NP-hardness result: allowing mass to split may destroy the discrete obstruction, so claiming hardness transfer there would be premature.

The scenario is large-scale strategic school choice. There are many households, but only moderately many complete behavioural types. A type \(t\) consists of its colour \(\chi(t)\in\{R,B\}\) and its complete accessibility set \(S_t\subseteq Q\) of schools. Two households have the same type exactly when they have the same homophilic preference and the same accessible schools. The society is a rational distribution \(\mu=(\mu_t)_{t\in\mathcal T}\), where \(\mu_t\) is the fraction of households of type \(t\), \(\sum_t\mu_t=1\), and \(n\gg|\mathcal T|\). This is a genuine high-multiplicity regime: millions of households are represented by a small number of repeated access-and-colour profiles.

The continuous problem is:

\[ \textsc{Mass-IBE} \]

An instance consists of \(Q\), the type set \(\mathcal T\), the access sets \(S_t\), rational masses \(\mu_t\), and the paper’s threshold \(\vartheta\in[0,1]\). A state is a mass assignment \(x=(x_{tq})\) satisfying

\[ x_{tq}\ge 0,\qquad x_{tq}=0\ \text{if }q\notin S_t,\qquad \sum_{q\in S_t}x_{tq}=\mu_t. \]

Let

\[ \ell_q=\sum_t x_{tq},\qquad r_q=\sum_{\chi(t)=R}x_{tq},\qquad b_q=\sum_{\chi(t)=B}x_{tq}. \]

For an occupied resource,

\[ \widehat u_t(q,x)= \begin{cases} \min\{r_q/\ell_q,\vartheta\},&\chi(t)=R,\[2mm] \min\{b_q/\ell_q,\vartheta\},&\chi(t)=B. \end{cases} \]

For an empty accessible resource, set \(\widehat u_t(q,x)=\vartheta\), reflecting the paper’s convention that moving alone to an empty resource gives maximum utility. The task is to output a feasible \(x\) such that whenever \(x_{tq}>0\),

\[ \widehat u_t(q',x)\le \widehat u_t(q,x) \qquad\text{for every }q'\in S_t. \]

Thus no positive-mass type has a profitable infinitesimal move. The action is a mass flow from types to resources; the objective is to find any mass state satisfying the impact-blind equilibrium predicate. The corresponding welfare is naturally

\[ W(x)=\sum_{t,q}x_{tq}\widehat u_t(q,x), \]

although welfare optimization is not needed for this anchor.

This is recognisably the paper’s problem, not a softened substitute. The accessibility graph remains central, the utility is still the fraction of same-colour users capped at \(\vartheta\), and the equilibrium notion is specifically the paper’s impact-blind one. Indeed, the paper itself motivates impact-blindness by large systems in which agents observe type fractions rather than exact counts, and says that Bayesian equilibria converge to IBE as the population grows. In the nonatomic limit, an individual deviation has zero effect on aggregate fractions, so the paper’s informational approximation becomes the natural equilibrium semantics.

I expect \(\textsc{Mass-IBE}\) to be tractable. Algorithm 1 should lift by replacing cardinalities with masses: repeatedly assign every blue type whose remaining access set is a singleton; choose the resource maximizing the ratio of remaining accessible red mass to forced blue mass; assign all remaining red mass incident to that resource; then remove the resource and the assigned mass. The monotonicity argument behind Lemma 3 uses only that the numerator decreases and the denominator increases, so it appears to survive for rational masses. This suggests a type-level algorithm with \(O(|E_{\mathcal T}|+|Q|\log |Q|)\) arithmetic operations, and polynomial bit complexity for rational input. The formal weighted proof would itself be a worthwhile result.

The expected classification is therefore Class A, at least for the intended finite-type representation. It also generates several questions: whether best or worst Mass-IBE can be optimized efficiently; whether the paper’s Price of Anarchy bounds survive with mass states; how well a rational mass equilibrium can be rounded to \(n\) named households; and whether an impact-aware continuum analogue should use finite-mass block deviations, since ordinary infinitesimal deviations make impact-awareness collapse into impact-blindness.

The weakest point is that the positive result may look like a weighted reimplementation of Algorithm 1 rather than a new continuous-optimization phenomenon. If \(|\mathcal T|\) approaches \(n\), the population compression disappears. I would concede that this mirror does not establish a new theorem about every aspect of the paper, nor does it justify a continuous version of its NP-hardness result. But that is not fatal: the authors’ own large-system interpretation makes IBE the correct limit object, and Theorem 5 gives a named, proved computational result whose algorithm has a credible type-mass formulation. The mirror covers Theorem 5 emphatically, while leaving the impact-aware, welfare-hardness, and price-of-anarchy results for separate investigations.