--- title: 'Clipped Linear Lottery: Smooth Selection' url: https://www.emergentmind.com/topics/clipped-linear-lottery type: topic --- # Clipped Linear Lottery: Smooth Selection The **Clipped Linear Lottery (CLL)** is a randomized selection mechanism for settings in which \(n\) candidates are evaluated from numerical reviews and exactly \(k\) must be selected. It was introduced as the central mechanism in “Smooth Partial Lotteries for Stable Randomized Selection” [2605.20069]. The mechanism assigns marginal selection probabilities by applying a common affine transformation to candidate utilities and clipping the result to \([0,1]\), thereby creating three regions: automatic rejection, automatic acceptance, and a linear lottery region. Its defining purpose is to make partial lotteries **smooth**: small changes in review scores should induce only bounded changes in all candidates’ selection probabilities. ## 1. Selection setting and motivation The CLL is formulated for **randomized selection** in domains such as scientific funding, admissions, hiring, and peer review. There are \(n\) candidates, exactly \(k\) must be selected, and candidate \(i\) receives \(r\) numerical reviews on a known scale. After normalization to \([0,1]\), the data is a matrix \[ X \in [0,1]^{n \times r},\quad X_{i,j} \text{ is reviewer } j\text{'s score for candidate } i. \] A utility vector \(u(X)\in\mathbb{R}^n\) is then computed, where \(u_i(X)\) is the estimated quality of selecting candidate \(i\). A canonical example is the mean score, \[ u_i(X) = \frac{1}{r}\sum_{j=1}^r X_{i,j}. \] A randomized selection rule maps the review matrix to a distribution over size-\(k\) subsets, \[ D : [0,1]^{n\times r} \to \Delta(\mathcal{K}),\quad \mathcal{K} = \{S\subseteq [n] : |S| = k\}, \] and induces marginal selection probabilities \[ p_i(X) = \Pr_{S \sim D(X)}[i \in S],\quad p(X)\in[0,1]^n. \] These formal objects define the environment in which the CLL operates [2605.20069]. The mechanism is motivated by a defect of deterministic and threshold-based selection. Deterministic top-\(k\) selection is brittle near the cutoff. Many institutions therefore use **partial lotteries** that automatically accept candidates above a high threshold, automatically reject those below a low threshold, and randomize uniformly among a middle group. The paper argues that such threshold-based designs are **inherently unstable**: an arbitrarily small perturbation in one review can move a candidate across a tier boundary, changing their selection probability from \(0\) to \(1/3\) or \(1\), while simultaneously causing large changes for others. Prior work cited there shows that such mechanisms can exhibit “maximum instability,” and the empirical analysis reports that on ICLR, NeurIPS, and Swiss National Science Foundation data, a one-point change in one review can change a paper’s selection probability by more than \(0.3\) under existing lotteries [2605.20069]. This motivates the paper’s central design principle: **smoothness**. Rather than using sharp tier boundaries, the mechanism should ensure that small perturbations in scores produce only proportionally small perturbations in marginal selection probabilities. The CLL is proposed as a simple and interpretable mechanism satisfying that requirement. ## 2. Formal definition of the mechanism The CLL takes as input a review matrix \(X\), a utility function \(u\) with Lipschitz constant \(D_u\), a target smoothness parameter \(L>0\), and the budget \(k\). It first computes utilities \(u_i(X)\), then sets a common slope \[ w \leftarrow \frac{L}{2D_u}, \] scales utilities, \[ z_i \leftarrow w \cdot u_i, \] and then finds an intercept \(b\in\mathbb{R}\) such that the clipped values satisfy the budget constraint \[ \sum_{i=1}^n \operatorname{clip}_{[0,1]}(z_i + b) = k, \] where \[ \operatorname{clip}_{[0,1]}(x) = \min\{1,\max\{0,x\}\}. \] The marginal probabilities are then \[ p_i(X) = \operatorname{clip}_{[0,1]}(z_i+b), \] and a size-\(k\) subset with these marginals can be sampled, for example by systematic sampling [2605.20069]. Equivalently, letting \(u_i=u_i(X)\) and \(w=L/(2D_u)\), the mechanism can be written explicitly as \[ p_i(X) = \begin{cases} 0 & \text{if } w u_i + b \le 0, \\ 1 & \text{if } w u_i + b \ge 1, \\ w u_i + b & \text{if } 0 < w u_i + b < 1. \end{cases} \] This representation makes the three-region structure explicit. Candidates in the lower region are always rejected, candidates in the upper region are always accepted, and candidates in the middle region participate in a lottery with probabilities that increase linearly in utility. The same rule can be parameterized through implicit utility thresholds \[ \ell = \frac{-b}{w}, \qquad u = \frac{1-b}{w}, \] yielding \[ p_i(X) = \begin{cases} 0 & \text{if } u_i \le \ell,\\[3pt] 1 & \text{if } u_i \ge u,\\[3pt] w u_i + b & \text{if } \ell < u_i < u. \end{cases} \] The thresholds \(\ell\) and \(u\) are **not fixed exogenously**; they are determined by the intercept \(b\) needed to enforce \(\sum_i p_i = k\). This is a central structural feature of the CLL: it preserves the intuitive “three-tier” logic of partial lotteries while replacing discrete tier boundaries with a continuous linear transition [2605.20069]. ## 3. Smoothness as a design principle The paper formalizes smoothness as a Lipschitz condition on the marginal probability map \(p\). The distance between review matrices is measured in the \(\ell_{1,1}\) norm, \[ \|X - X'\|_{1,1} = \sum_{i=1}^n \sum_{j=1}^r |X_{i,j} - X'_{i,j}|. \] A randomized selection rule is **\(L\)-smooth** if \[ \|p(X) - p(X')\|_1 \le L \,\|X - X'\|_{1,1} \quad\text{for all } X,X' \in [0,1]^{n\times r}. \] If only one review changes by \(\delta\), then \(\|X-X'\|_{1,1}=|\delta|\), so the total change in marginal probabilities across all candidates is at most \(L|\delta|\). The paper interprets this as a guarantee that no single review can disproportionately swing the outcome [2605.20069]. The utility function is itself required to be smooth. A utility map \(u:[0,1]^{n\times r}\to\mathbb{R}^n\) is \(D_u\)-Lipschitz if \[ \|u(X) - u(X')\|_1 \le D_u\,\|X - X'\|_{1,1}\quad\forall X,X'. \] The mean-score utility is \(D_u=1/r\)-Lipschitz, while the min, max, and median score per candidate are all \(1\)-Lipschitz. The analysis notes that if utilities were arbitrarily non-smooth, no smooth mechanism could track them without unbounded regret [2605.20069]. The key theorem states that the CLL with slope \[ w=\frac{L}{2D_u} \] is exactly \(L\)-smooth: \[ \|p(X) - p(X')\|_1 \le L \|X - X'\|_{1,1}\quad \forall X,X'. \] The paper also states that this bound is essentially tight: there are instances where the true Lipschitz constant is arbitrarily close to \(L\). The proof intuition isolates two sources of variation. First, the utilities change by at most \(D_u\|X-X'\|_{1,1}\). Second, because the intercept \(b\) must shift to preserve the budget constraint, that common shift can double the total \(\ell_1\) change. This yields an upper bound of \[ 2wD_u\|X-X'\|_{1,1}, \] and substituting \(w=L/(2D_u)\) gives the advertised smoothness guarantee [2605.20069]. The result has an immediate operational interpretation. A one-point change in one normalized review can change the sum of absolute marginal-probability changes across all candidates by at most \(L\). The mechanism therefore excludes discontinuous probability jumps at tier boundaries and instead forces gradual changes in every candidate’s marginal selection probability. ## 4. Utility, regret, and near-optimality The CLL is evaluated relative to the utility vector \(u(X)\). For a given review matrix, the optimal deterministic benchmark is the total utility of the top-\(k\) candidates, \[ \text{OPT}(X) = \max_{S \in \mathcal{K}} \sum_{i\in S} u_i(X). \] The expected utility under a randomized rule with marginals \(p(X)\) is \[ p(X)^\top u(X) = \sum_{i=1}^n p_i(X)u_i(X), \] and the regret is \[ R(D;X)=\text{OPT}(X)-p(X)^\top u(X). \] The paper studies worst-case regret over all review matrices [2605.20069]. A geometric characterization is central to the analysis. Let \[ \mathcal{C}_{n,k} = \{p \in [0,1]^n : \sum_{i=1}^n p_i = k\} \] be the capped simplex. For each \(X\), the CLL output is the Euclidean projection of the scaled utility vector \(w u(X)\) onto \(\mathcal{C}_{n,k}\): \[ p(X) = \arg\min_{p \in \mathcal{C}_{n,k}} \frac{1}{2}\|p-wu(X)\|_2^2. \] Thus, among all feasible marginal vectors with total mass \(k\), the CLL chooses the one closest in \(\ell_2\) to the scaled utilities [2605.20069]. The resulting worst-case regret bound is \[ \max_X R(D;X) \le \frac{k(1-k/n)D_u}{2L}. \] For mean-score utility, where \(D_u=1/r\), \[ \max_X R(D;X) \le \frac{k(1-k/n)}{2Lr}. \] The paper interprets this as a smoothness–utility tradeoff. Larger \(L\) permits steeper slopes and smaller regret; more reviews per candidate reduce \(D_u\) and therefore reduce the worst-case regret at fixed \(L\). The factor \(k(1-k/n)\) is also identified with the regret of uniform selection, which is perfectly smooth but ignores the scores [2605.20069]. The analysis does not stop at an upper bound. For mean utility, every \(L\)-smooth selection rule satisfies a lower bound of the form \[ \max_X R(D;X)\ge \begin{cases} \displaystyle \frac{k(1-k/n)}{2 L r} \Big(1 - \frac{1}{2}\frac{1-k/n}{L r}\Big)^2, & \text{if } L \ge \frac{1-k/n}{r},\\[8pt] \displaystyle \frac{k(1-k/n)}{2r}, & \text{if } L < \frac{1-k/n}{r}. \end{cases} \] In the regime \[ L \ge \frac{1-k/n}{r}, \] the lower bound differs from the CLL upper bound only by a factor of \((1-k/n)\). The paper therefore concludes that, up to this small constant factor, the CLL is **minimax optimal** among \(L\)-smooth selection rules. It also states that analogous near-optimality results extend to other utility functions, including median, min, and max [2605.20069]. ## 5. Relation to fairness, privacy, softmax, and other meanings of “lottery” A recurring source of confusion is that the CLL is neither a reformulation of **Individual Fairness** nor a privacy mechanism, even though it has stability guarantees. Individual Fairness requires that similar candidates on a **fixed input** receive similar probabilities: \[ |p_i(X)-p_j(X)| \le \alpha |u_i(X)-u_j(X)|. \] Smoothness, by contrast, constrains how the **same candidate’s** probability changes across different inputs \(X\) and \(X'\). The paper states that the two notions are formally incomparable. A constant rule that always selects the first \(k\) candidates is \(0\)-smooth but can violate Individual Fairness, while an Individually Fair rule can be arbitrarily non-smooth across datasets. At the same time, prior work cited there shows that on a fixed instance, the clipped-linear rule with slope \(w=\alpha\) is the exact instance-optimal solution to maximizing utility under Individual Fairness constraints [2605.20069]. The same section distinguishes smoothness from **Differential Privacy**. Standard \(\varepsilon\)-DP is defined over neighboring inputs differing in one entry: \[ \Pr[D(X)=S] \le e^\varepsilon \Pr[D(X')=S]. \] The paper proves that standard DP does **not** imply smoothness: for every \(\varepsilon>0\), there exists an \(\varepsilon\)-DP selection rule that is not \(L\)-smooth for any finite \(L\). Conversely, smoothness does not imply DP, because one can keep marginals fixed while changing the joint support dramatically. When privacy is reformulated as **metric DP** with respect to \(\|X-X'\|_{1,1}\), \[ \Pr[D(X)=S] \le \exp(\varepsilon \|X-X'\|_{1,1}) \Pr[D(X')=S], \] the paper shows that this does imply smooth marginals with \[ L \le \varepsilon k, \] and that the dependence on \(k\) is tight for small distances. However, metric-DP-based design yields a much weaker regret guarantee, of order \[ \Omega\left(\frac{k^2 D_u \log(n/k)}{L}\right), \] which is worse than the CLL’s \(\Theta\!\left(\frac{k(1-k/n)D_u}{L}\right)\) scaling by an extra factor of approximately \(k\log(n/k)\) [2605.20069]. The paper also compares the CLL to **top-\(k\) softmax**, equivalently the Exponential Mechanism or Gumbel-top-\(k\) sampling. With temperature \(T>0\), \[ \sigma_T(u)_i=\frac{\exp(u_i/T)}{\sum_j \exp(u_j/T)}, \] and sampling proceeds without replacement. If \(u\) is \(D_u\)-Lipschitz, the top-\(k\) softmax rule induces \(L\)-smooth marginals with \[ L \le \frac{2D_u}{T} \] and has worst-case regret \[ \max_X R(D;X)\le kT\log n. \] Setting \(T=2D_u/L\) to match a target smoothness level gives \[ \max_X R(D;X)\le \frac{2kD_u\log n}{L}. \] The paper therefore concludes that softmax is worse than the CLL by a factor of \(\Theta(\log n)\) in worst-case regret at the same smoothness level. It also highlights a structural distinction: softmax has full support, whereas the CLL is sparse, automatically accepting or rejecting some candidates and randomizing only in the middle region [2605.20069]. A separate terminological clarification concerns earlier work on **lotteries** in computational mechanism design. In “Pricing Randomized Allocations” [0904.2400], a lottery is a pair \((\phi,p)\) consisting of a probability vector over items and a price, and buyer utility is linear in \(\phi\). That literature studies randomized allocations for revenue extraction rather than score-based subset selection. The shared word “lottery” refers to randomization over outcomes in both cases, but the formal objects, objectives, and constraints differ [0904.2400]. ## 6. Empirical behavior, operational tuning, and limitations The empirical evaluation in the CLL paper uses four datasets: ICLR 2025 with \(n=3{,}710\) and at least \(3\) reviews per paper on a \(1\)–\(10\) scale; NeurIPS 2024 with \(n=4{,}034\) and at least \(3\) reviews per paper on a \(1\)–\(10\) scale; Swiss National Science Foundation data with \(n=353\) and at least \(5\) reviews per proposal on a \(1\)–\(6\) scale; and a Synthetic Beta dataset with \(n=200\), \(r=5\), and reviews drawn i.i.d. from \(\text{Beta}(\alpha,\alpha)\) on \([0,1]\) and discretized into \(10\) levels. The utility is always mean normalized score, and the experiments consider acceptance rates \(k/n\in\{10\%,33\%,50\%\}\) [2605.20069]. Against existing interval-based lotteries such as MERIT and the Swiss NSF lottery procedure, the paper measures local smoothness under worst one-review, one-tick perturbations. On ICLR and NeurIPS, a one-point change in one review can change a candidate’s selection probability by more than \(0.3\) under MERIT or Swiss NSF. The reported local norm ratios are in the range of approximately \(5\)–\(25\), far above the target \(L=1/r\) scale. In plots of regret versus smoothness, MERIT and Swiss NSF lie high and to the right, while the CLL can achieve much smaller \(L\) for the same or lower regret [2605.20069]. The comparison with softmax is similarly sharp. Sweeping the smoothness parameter \(L\), the experiments measure normalized regret \(R(D;X)/k\) for both mechanisms. Across all datasets and all acceptance rates, the CLL’s regret–smoothness curve lies **strictly below** softmax’s: \[ R_{\text{CLL}}(L) < R_{\text{softmax}}(L). \] The paper interprets this as empirical confirmation of the theoretical logarithmic gap in regret between the two mechanisms. The smoothness bounds themselves are also reported to be tight: for synthetic hard utility profiles with \(n=1000\) and \(k\in\{100,333,500\}\), empirical worst-case smoothness is within \(99\%\) of the theoretical bound for the CLL and within \(95\%\) of the corresponding bound for softmax [2605.20069]. Operationally, the principal tuning parameter is the global smoothness level \(L\). Once \(L\) is chosen, the slope is set to \[ w=\frac{L}{2D_u}. \] Larger \(L\) yields a steeper slope, a smaller lottery region, and behavior closer to deterministic top-\(k\) selection; smaller \(L\) yields a flatter slope, a larger lottery region, and greater stability. The intercept \(b\) is then determined as the unique solution to \[ \sum_{i=1}^n \operatorname{clip}_{[0,1]}(wu_i+b)=k. \] This automatically generates a lower cutoff \(\ell=-b/w\), an upper cutoff \(u=(1-b)/w\), and a data-dependent lottery band \([\ell,u]\) [2605.20069]. The paper also proves two monotonicity properties useful for deployment. First, the mechanism is monotone in budget: fixing utilities and \(L\), if the budget increases from \(k\) to \(k'>k\), then \[ p_i^{(k')}(X) \ge p_i^{(k)}(X)\quad \forall i. \] Second, letting \[ \mathcal{R}(w)=\{i:0