--- title: Restriction-Sensitive Choice (RSC) Theory url: https://www.emergentmind.com/topics/restriction-sensitive-choice-rsc type: topic --- # Restriction-Sensitive Choice (RSC) Theory Restriction-Sensitive Choice (RSC) most specifically denotes a formal choice-theoretic model of the forbidden fruit effect: restricting access to some opportunities may steer desire toward their substitutes, so that removal of an option can induce a reversal toward another option of the same subjective “type” rather than leave the remaining ranking unchanged. In adjacent literatures, the same underlying idea of sensitivity to restrictions reappears in menu-dependent shortlisting, restricted codomains of feasible choice outputs, privacy-preserving elicitation, and preference-domain restrictions, while the acronym “RSC” is also used independently for unrelated constructions in machine learning and analysis [2509.11673]. ## 1. Formal choice-theoretic RSC and the forbidden fruit effect In the choice-theoretic sense, RSC is defined on a finite set of options \(X\), with \(\mathcal X = 2^X \setminus \emptyset\), and a choice function \(c:\mathcal X\to X\) such that \(c(A)\in A\) for every menu \(A\). Its basic behavioral signature is a choice reversal caused by removal of an unchosen option. The key revealed relation is \[ x \mathbf{R}^c y \quad\text{if there exists } z \text{ such that } z=c\{x,y,z\} \text{ and } x=c\{x,z\}. \] Thus \(x\mathbf{R}^c y\) means that when \(y\) is present in the triple \(\{x,y,z\}\), the chosen option is \(z\), but when \(y\) is removed and only \(\{x,z\}\) remains, choice switches to \(x\). The paper interprets this as \(x\) “reacting” to the absence of \(y\) [2509.11673]. The model then builds a subjective similarity relation from such reactions. If \(x\) reacts to the absence of \(y\), then \(x\) and \(y\) are interpreted as similar; more generally, similarity is the transitive closure of reaction links. This yields equivalence classes of “types,” the subjective categories within which the forbidden fruit effect can operate. The core representation is two-stage. A choice function \(c\) is an RSC if there exist a partition \(\mathcal T\) of \(X\) into types, a linear order \(\succsim_1\) (“welfare preference”), and a linear order \(\succsim_2\) (“reaction preference”), such that \[ c(A)=\max(d(A),\succsim_2), \qquad d(A)=\bigcup_{T\in\mathcal T}\max(T\cap A,\succsim_1). \] First, within each type \(T\), only the best available option according to \(\succsim_1\) survives. Second, across the surviving type-representatives, the final choice is the \(\succsim_2\)-maximal option. The mismatch between \(\succsim_1\) and \(\succsim_2\) is what generates restriction-induced reversals. The main characterization theorem states: \[ c \text{ satisfies Exp, NRS, and IR } \iff c \text{ is an RSC.} \] Here **Expansion (Exp)** requires that if \(x=c(A)=c(B)\), then \(x=c(A\cup B)\); **No-Reaction Similarity (NRS)** imposes transitivity-like consistency among pairwise choices within the same subjective type; and **Independent Reaction (IR)** requires that the strength of reaction depends only on the availability of similar options, not on the identity of dissimilar outside options. Proposition \(\ref{RSC_property}\) further identifies exactly when a reaction is observed: \(x\mathbf{R}^c y\) iff there exist a type \(T\) and an outside option \(z\notin T\) such that \(x,y\in T\), \(y\succ_1 x\), and \(x\succ_2 z \succ_2 y\) [2509.11673]. ## 2. Identification, single-peaked reaction, welfare, and freedom A central contribution of the RSC model is identifiability from observed reversals. The equivalence classes of the revealed similarity relation are observable from \(\mathbf{R}^c\), and these classes can be taken as the types: \[ \mathcal T = X_\approx. \] Proposition \(\ref{proposition:types}\) states that if \(c\) is an RSC, then there exists a rationalizing RS-structure with exactly these types, and any rationalizing structure must place similar options in the same type. The types are therefore revealed by reaction data rather than imposed ex ante [2509.11673]. The paper also studies **single-peaked RSC**. For each type \(T\), there exists a threshold \(x_T^\star\) such that: \[ \succsim_2 = \succsim_1 \text{ on } \left[x_T^\star,\max(T,\succsim_1)\right]^{\succsim_1}_T, \] and \(\succsim_2\) is single-peaked with respect to \(\succsim_1\) on \[ \left[\min(T,\succsim_1),x_T^\star\right]^{\succsim_1}_T. \] This means that no restriction motive operates above the threshold, while below the threshold reaction rises and then eventually weakens. The corresponding behavioral postulate is **Single-Peaked Reaction (SPR)**, and the theorem is: \[ c \text{ satisfies SPR } \iff c \text{ is single-peaked RSC.} \] In minimal single-peaked representations, the threshold is identifiable: \[ x^\star_T = \min\big(\{x\in T\mid \nexists y,\ x\mathbf{R}^c y\},\ \succsim_1\big). \] The paper also identifies the most reaction-inducing options through the set of options that are never themselves “reacted to.” This yields a minimal RS-structure from revealed reaction patterns. The normative analysis distinguishes welfare from observed choice. Welfare is represented by \(\succsim_1\), not by \(\succsim_2\), and the paper therefore defines a revealed welfare-improving relation \(\gg^c\). It also develops a freedom measure. For each type \(T\), the set of freedom-satisfying options is \[ F^T \equiv \{x\in T : x\succ_1 x_T^\star\}. \] A menu satisfies freedom \(T\) if it contains at least one option in \(F^T\), and its freedom value is \[ n(A)=\text{number of types }T\text{ such that }F^T\cap A\neq\emptyset. \] Theorem \(\ref{menu_pref}\) states: \[ \succsim \text{ satisfies R-Dominance and R-Composition} \iff A\succsim B \iff n(A)\ge n(B). \] Freedom is thus measured by the number of represented freedom-types [2509.11673]. The model is applied to two further phenomena. In the belief backfire application, there exists a threshold \(p^\star < 1/2\) such that for priors \(p\in[p^\star,1/2]\), removing a moderate \(R\)-biased source makes the decision maker prefer the extreme source \(\sigma^{RR}\) to the moderate source \(\sigma^L\), and then choose action \(r\) after observing the signal from \(\sigma^{RR}\). In the integration-policy application, the restricted menu is \[ K_g = \{(t,d)\in [0,1]^2 : t + d^\beta g \le 1\}, \] the parent maximizes \[ t + P(d)V(g), \qquad P(d)=d+(1-d)q, \] and the steady-state minority share is \[ q^\star(g^m)=\frac{V(g^m)/g^m}{V(1)+V(g^m)/g^m}. \] Since \(V(g)/g\) rises with \(g\) when \(g\ge \hat g\), stronger repression increases the steady-state size of the minority [2509.11673]. ## 3. Related menu-dependent and globally constrained models A closely related two-stage model is **Choice by Rejection (CBR)**. In that model, \[ C(S)=\max\bigl(S \setminus \min(S,R),\,P\bigr), \] where \(R\) is the first-stage rationale and \(P\) is the second-stage rationale. The rejected set is defined by \[ \min(S,R)=\{x \in S : \exists z \in S \text{ s.t. } zRx \text{ and } \nexists z' \in S \text{ s.t. } xRz'\}. \] Here \(R\) is transitive and possibly incomplete, while \(P\) is complete. The shortlist is therefore not the set of \(R\)-maximal elements, as in the Rational Shortlist Method, but the set of options that are not minimal under \(R\). The model is explicitly restriction-sensitive because shortlisting is menu-dependent and the final choice is taken over a restricted survivor set [2108.07424]. CBR is analyzed through reversals under set inclusion. With pairwise choice defined by \[ x \succ_c y \quad \Longleftrightarrow \quad C(\{x,y\})=x, \] the model admits weak reversals, strong reversals, and double reversals such as \[ C(\{x,y\})=x,\quad C(\{x,y,z\})=y,\quad C(\{x,y,z,w\})=x. \] Its main behavioral characterization is: > A choice function \(C\) is CBR representable iff it satisfies (A1)–(A4). The weakened WARP-like discipline is **R-WARP**, and, together with **WCC\(^*\)**, it implies **R-WARP\(^*\)**, which allows at most two reversals along an inclusion chain. Identification is also partial and structured: in a minimal CBR representation \((R^*,P^*)\), \[ R^*=R^c,\qquad P^c \subseteq P^*. \] Thus the first rationale is pinned down exactly by the transitive closure of revealed reversals, while the second rationale is only partially identified [2108.07424]. A different but related restriction-sensitive construction appears in sequential choice under uncertainty. There, dynamic decision problems are represented by finite decision trees, and abandoning separability or consequentialism invalidates Bellman’s principle. Under **resolute choice**, successive Selves cooperate to implement an initially chosen strategy rather than behave as adversaries in a sophisticated backward-induction game. The paper’s minimal cooperative version is **justifiable choice**: each Self chooses only among decisions that belong to at least one undominated strategy. Operationally, the original tree \(T\) is replaced by a reduced tree \(T_0\) or \(T_1\) spanned by undominated or justifiable strategies, and backward induction is then carried out within that restricted structure. This yields a choice procedure in which local feasibility depends on global undominatedness, not only on subtree-local information [1301.7388]. The paper formalizes the benchmark **subjective expected utility** \[ U(d) = \sum_{i=1}^n P(A_i)\,u(c_i) \qquad (1) \] and the non-expected-utility criterion **Choquet Expected Utility** \[ V(d) = u(c_1) + \sum_{i=2}^{n} \Pi(A_i)\,[u(c_i)-u(c_{i-1})] \qquad (2). \] Because CEU can violate dynamic consistency, ordinary dynamic programming may produce dominated strategies. The proposed implementation instead generates undominated strategies through families of positive weight systems, constructs a reduced subtree \(T_1\), and then rolls it back recursively. A limited-cooperation version replaces full commitment by \(\varepsilon_0\)-acceptability, pruning substrategies that later Selves will not accept [1301.7388]. ## 4. Restricted choice structures and linear orders on sets Another formalization of restriction-sensitive choice arises when the very codomain of feasible outputs is restricted. A **restricted choice structure** is \[ \mathbb{S}=(A,\mathcal{D},\mathcal{R}), \] where \(A\) is the set of alternatives, \(\mathcal{D}\subseteq \mathcal{P}(A)\) is the domain of admissible inputs, and \(\mathcal{R}\subseteq \mathcal{D}\) is the set of realizable choices. A choice function for \(\mathbb S\) is a function \(C:\mathcal D\to\mathcal R\) such that whenever there exists a realizable choice \(E\in\mathcal R\) with \(E\subseteq S\), the function returns such an \(E\); if no such \(E\) exists, the function returns a distinguished fallback value \(K\in\mathcal R\) [2506.03315]. The paper’s key point is that in such settings, representation by an order on alternatives is insufficient. One instead orders sets of alternatives. Let \(\ll\) be a \(\mathcal D\)-smooth linear order on \(\mathcal R\), where \(\mathcal D\)-smoothness means that every relevant input \(S\in\mathcal D\) with some realizable subset has at least one \(\ll\)-minimal such subset. Let \(K\)-minimality mean \[ \min(\ll)=\{K\}. \] Then the canonical linear choice operator is \[ \triangledown_{\ll}^{K}(S)= \begin{cases} E & \text{if } {S}\!\ll=\{E\},\\[2mm] K & \text{otherwise.} \end{cases} \tag{\(\star\)} \] If there is a unique \(\ll\)-minimal realizable subset of \(S\), that subset is chosen; otherwise the fallback \(K\) is returned [2506.03315]. The existence theorem states that, assuming the Axiom of Choice, every restricted choice structure admits a \(K\)-minimal linear choice function. The union-closed characterization is: > Let \(\mathbb{S}=(A,\mathcal{D},\mathcal{R})\) be union-closed and let \(K\in\mathcal{R}\). A function \(\triangledown:\mathcal{D}\to\mathcal{R}\) is a \(K\)-minimal linear choice function for \(\mathbb{S}\) iff it satisfies \((SS0a)\)–\((SS6)\). For the general non-union-closed case, \((SS5)\) and \((SS6)\) are replaced by \((SS5E)\) and \((SS6E)\). The axioms include: \[ \tag{SS0a} \text{If } \mathcal{R}\cap \mathcal{P}(S)\neq\emptyset,\text{ then }\triangledown(S)\subseteq S, \] \[ \tag{SS1} \text{If } \triangledown(S)\not\subseteq S,\text{ then }\triangledown(S)=K, \] \[ \tag{SS2} \text{If } K\subseteq S,\text{ then }\triangledown(S)=K, \] together with reciprocity, monotonicity, cycle-prevention, and persistence conditions [2506.03315]. This framework is applied to **theory change** and **abstract argumentation**. In theory change, a choice-based change operator \(K\ovee S = C_K(S)\) is linear when each \(C_K\) is \(K\)-minimal linear, yielding the axioms \((LCR1)\)–\((LCR6)\). In argumentation, a semantics is linear if each \(\Pi_F\) is a \(K_F\)-minimal linear choice function, yielding the axioms \((LCA1)\)–\((LCA6)\) [2506.03315]. ## 5. Preference-domain restrictions and privacy-preserving elicitation Restriction sensitivity also appears when the difficulty and behavior of collective choice depend on structural restrictions on preferences. In computational social choice, restricted domains such as single-peaked, single-crossing, and Euclidean preferences alter both axiomatic and algorithmic properties. For a profile \(P=(v_1,\dots,v_n)\), a domain restriction is a set of admissible profiles; many such restrictions are hereditary, meaning that every subprofile obtained by deleting voters and/or alternatives remains in the domain [2205.09092]. For **single-peaked preferences**, alternatives lie on an axis \(\lhd\), and each vote has one peak and declines away from it. The survey gives the equivalence: \[ \text{single-peaked on }\lhd \iff \text{no valleys on }\lhd \iff \text{upper contour sets are intervals of }\lhd. \] For odd numbers of voters, the majority relation is transitive and the median voter’s top choice is the unique Condorcet winner. Recognition is possible in \(O(mn)\), with a reduction to the consecutive-ones property. For **single-crossing preferences**, the voter order rather than the alternative order is constrained; recognition can be done in \[ O(nm\log m). \] For **\(1\)-Euclidean preferences**, the survey emphasizes the strict inclusion \[ 1\text{-Euclidean} \subsetneq \text{single-peaked} \cap \text{single-crossing}, \] while for fixed \(d\ge 2\), \(d\)-Euclidean recognition is \(\exists\mathbb R\)-complete. Under these domain restrictions, hard winner-determination problems such as Dodgson, Young, Kemeny, and Chamberlin–Courant often become polynomial-time solvable [2205.09092]. A different operational form of restriction-sensitive choice appears in privacy-preserving polling for sensitive multiple-choice questions. The paper proposes two methods in which the respondent’s answer is not the exact choice, but a subset containing the true one. The first is the **Pair method**: the respondent states the true choice together with one other choice picked uniformly at random from the remaining \(N-1\) parties, and reports the unordered pair. The response probabilities are \[ u_{ij} = \frac{1}{N-1}(p_i+p_j), \qquad i\neq j, \] and the unbiased estimator is \[ \hat p_i = \frac{N-1}{N-2}\sum_j \hat u_{ij} - \frac{1}{N-2}. \] Its variance is \[ Var[\hat p_i] = \frac{1}{n}\left(\frac{1+(N-3)p_i}{N-2}-p_i^2\right). \] The second is the **List method**: the respondent is shown a list of parties and answers whether the preferred party is on the list. Both methods are modeled as \[ u = Ap, \] with identification guaranteed if \(A\) has full column rank \(N\), and both use the estimator \[ \hat p = \frac{1}{n}(A'A)^{-1}A'X. \] The privacy–efficiency comparison is explicit: the Pair method is more informative and less private, while the List method is more private and less informative [1803.10568]. ## 6. Acronymal and analytical uses beyond choice theory Outside choice theory, the acronym **RSC** is used for unrelated constructions. In spiking neural networks, **RSC** stands for **Randomized Smoothing Coding**. The paper links Poisson coding, randomized smoothing, and certified adversarial robustness through the smoothed classifier \[ g(x)=\arg\max_{c\in\mathcal{Y}} \mathbb{P}(f(x+\epsilon)=c), \qquad \epsilon \sim \mathcal{N}(0,\sigma^2 I). \] It defines **RSC-I** by \[ \widetilde{x}=x+\epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2 I), \] and **RSC-II** by \[ \widetilde{x}_i = x_i+\epsilon_i,\qquad \epsilon_i \sim \mathcal{N}(0,\sigma^2 I), \quad i=1,\dots,|T|. \] Its main theoretical point is that Poisson coding has attack-dependent covariance, while RSC has fixed covariance. On ImageNet in the white-box setting, the reported accuracies are: ANN clean \(67.00\), FGSM \(0.66\), PGD \(0.00\); Direct SNN clean \(64.40\), FGSM \(4.56\), PGD \(0.00\); Poisson SNN clean \(52.29\), FGSM \(15.73\), PGD \(4.70\); RSC-SNN clean \(53.79\), FGSM \(25.86\), PGD \(7.38\) [2407.20099]. In graph neural networks, **RSC** denotes **Randomized Sparse Computation**. The framework targets sparse matrix–dense matrix multiplications in the backward pass and manages approximation globally, layer by layer and epoch by epoch. Its constrained optimization has the budget form \[ \sum_{l=1}^{L} \sum_{i\in\mathrm{Top}_{k_l}} \#nnz_i d_l \le C \sum_{l=1}^{L} |\mathcal{E}| d_l, \] uses a greedy solver with \[ \alpha = 0.02 |\mathcal{V}|, \] caches sampled sparse matrices across nearby iterations, and switches back to exact sparse operations in the final phase using **80% approximate / 20% exact**. The reported speedups are up to \(\mathbf{11.6\times}\) for a single sparse operation and up to \(\mathbf{1.6\times}\) end-to-end wall-clock time speedup, with an end-to-end accuracy drop of about \(\approx 0.3\%\) [2210.10737]. In harmonic analysis, the phrase “restriction-sensitive” is used in a different technical sense. One paper proves that a Fourier restriction estimate \[ \|\widehat f\|_{L^q(S,\sigma)} \le C_{\mathrm{restr}}\|f\|_{L^p(\mathbb R^d)} \] with the strict exponent gap \(p