--- title: 'Tight Mechanisms: Precise Performance Bounds' url: https://www.emergentmind.com/topics/tight-mechanisms type: topic --- # Tight Mechanisms: Precise Performance Bounds Searching arXiv for recent and relevant papers on “tight mechanisms” across mechanism design and adjacent technical uses of “tight” mechanism terminology. Tight mechanisms are mechanisms or mechanism classes for which the relevant performance statement is exact: an approximation ratio matches a lower bound, an impossibility frontier is met, an “if and only if” characterization closes the feasible design space, or a numerical accountant returns certified privacy bounds whose residual gap is only discretization or truncation error. Across the literature, the phrase is used in several related but non-identical senses, most prominently in approximation-based mechanism design, auditability theory, and differential privacy [1902.04635] [1811.00763] [2305.09314] [2006.07134]. ## 1. Core meanings of tightness The term “tight” is not tied to a single model. In budget-feasible procurement, it means matching lower bounds against the relevant benchmark. In posted pricing and auction design, it means identifying the exact supremum of a revenue ratio. In facility location, it means determining the exact worst-case approximation ratio of a strategyproof mechanism. In auditability, it means exact characterization of the smallest coalition needed to expose any deviation. In differential privacy, it means exact privacy parameters or rigorously certified lower and upper bounds for the true privacy loss. | Domain | Tight quantity | Representative result | |---|---|---| | Budget-feasible procurement | Approximation ratio | Randomized ratio \(2\); deterministic ratio \(3\) against \(\textsf{Fractional Knapsack}\) [1902.04635] | | Bayesian pricing | Exact revenue gap | \(OPT/AP = C \approx 2.6202\) and \(BUPP/UP = C\) [1811.00763] | | Multi-unit auctions | Asymptotic revenue gap | Within-group \(1+\Theta(1/\sqrt{k})\); cross-group \(\Theta(\log k)\) [2102.07454] | | Euclidean facility location | Exact approximation ratio | \(\sqrt{2}\) for \(1\le p\le 2\), \(2^{1-\frac{1}{p}}\) for \(p\ge 2\) for coordinate-wise median [2606.08621] | | Auditability | Auditability index | \(\#IA=2\), \(\#DA=N\) [2305.09314] | | Differential privacy | Exact privacy characterization | Exact iff conditions and tight zCDP formulas [1505.07254] [2510.25746] | This suggests that “tight mechanism” is best understood as a closure notion: the mechanism attains the best provable bound available for the model and benchmark under study, or the paper proves that no better mechanism of the relevant type can exist. ## 2. Budget-feasible procurement with additive valuations In "Optimal Budget-Feasible Mechanisms for Additive Valuations" [1902.04635], the setting is a reverse auction with \(n\) agents, private costs \(c_i\), public buyer values \(v_i\), additive valuation \(v(A)=\sum_{i\in A}v_i\), and a hard budget \(B\). The offline benchmark is knapsack, and the stronger fractional benchmark is \[ fopt = \max_{x_i\in[0,1]} \sum_i x_i v_i \quad \text{s.t. } \sum_i x_i c_i \le B. \] The paper’s main randomized mechanism is a two-stage composition. A pruning step removes items with low value-per-cost ratio \(v_i/c_i\), producing a threshold \(r\) and surviving set \(S(r)\) such that \[ c_i \le \frac{v_i}{r} \le B \quad \forall i\in S(r), \] and \[ fopt \le v(S(r)) + r(B-c(S(r))) < 2v(S(r)). \] The second stage uses randomized posted prices for the highest-value item \(i^*\in S(r)\) and the remaining set \(T=S(r)\setminus\{i^*\}\). The analysis hinges on the identity \[ rB = 2q_{i^*}v_{i^*} + 2q_T v(T), \] together with lower bounds on the expected contribution of \(i^*\) and each \(i\in T\), yielding \[ fopt \le 2\cdot alg. \] The resulting mechanism is a \(2\)-approximation against \(\textsf{Fractional Knapsack}\), hence also against the integral knapsack benchmark, and the ratio is tight because there is already a matching lower bound of \(2\) for randomized budget-feasible mechanisms even against the weaker integral benchmark [1902.04635]. The same paper gives a deterministic mechanism with a \(3\)-approximation against the fractional optimum. It reuses pruning but replaces randomized prices with an adaptive posted-price rule that distinguishes the cases \(v_{i^*}\le \frac12 v(T)\), \(v_{i^*}\ge 2v(T)\), and \(\frac12 v(T)1. \] For hierarchical exchange, maximally auditable mechanisms are exactly vice ownership mechanisms; for sequential dictatorships, maximally auditable ones are exactly vice dictatorships; and mechanisms with \(\#\varphi=1\) are exactly those with a sequential clinching implementation [2305.09314]. In this line of work, tightness means exact structural characterization of the maximally auditable class. "Tight Samurai Accountant" [2509.24673] uses the word differently but no less precisely. In the Border–Sobel auditing/plunder model, a mechanism \(m=(a,r_P,r_{\emptyset})\) specifies an advance-payment-dependent audit probability and refunds. The paper defines the deviation-loss function \[ \lambda_m(x)=\inf_{y\in[\underline{x},x]} \Bigl[a(y)x+(1-a(y))(y-r_{\emptyset}(y))\Bigr], \] and says that \(m^\ast\) is tighter than \(m\) if \((\Pi_m,\lambda_m)\le (\Pi_{m^\ast},\lambda_{m^\ast})\). The main theorem states that if \(m\) is tight or efficient, then \[ \lambda_m\in\Lambda,\qquad R_m=\lambda_m,\qquad a=\max\{\alpha_{\lambda_m},\beta_{\lambda_m}\}, \] where \(\Lambda\) is the set of weakly increasing, weakly concave \(\lambda\) with \(\lambda(\underline{x})=\underline{x}\) and \(\lambda\le \operatorname{id}\). The paper further proves \[ \text{tight} \Rightarrow \text{efficient}, \] so tightness refines efficiency without loss of optimality [2509.24673]. Here the tight object is not an approximation ratio but a dominance order over profit and incentive-threat profiles. ## 6. Exact differential privacy mechanisms In finite categorical data, "Differentially Private Response Mechanisms on Categorical Data" [1505.07254] studies the discrete exponential mechanism \[ \mathbb{P}(X_d = d') = C\, e^{-k h(d,d')}, \qquad C = (1 + m e^{-k})^{-n}. \] A central contribution is a minimal verification principle: for neighboring databases \(d\sim d'\), it is enough to check privacy on the sufficient set \[ S_{d,d'} = \{\, d^* \in D^n : \mathbb{P}(X_d = d^*) > \mathbb{P}(X_{d'} = d^*) \,\}. \] The main privacy theorem is exact: \[ e^k \le e^\varepsilon + \frac{m\delta}{1-\delta} \] if and only if the mechanism is \((\varepsilon,\delta)\)-differentially private. The paper also derives a tight lower bound on maximal expected Hamming error and shows that the optimal product sanitisation mechanism uses \[ p = \frac{1-\delta}{e^\varepsilon + m}. \] This is an instance where privacy characterization, verification burden, and utility lower bound all line up exactly [1505.07254]. "Exact zCDP Characterizations for Fundamental Differentially Private Mechanisms" [2510.25746] gives analogous exactness results for zCDP. For the \(\epsilon\)-DP Laplace mechanism, the tight zCDP parameter is \[ \rho_{\mathrm{Lap}}=\epsilon+e^{-\epsilon}-1. \] For discrete Laplace, the paper gives a sensitivity-dependent exact formula that interpolates between binary randomized response and continuous Laplace. For \(k\)-randomized response with \(2\le k\le 6\), \[ \rho_{kRR} = \epsilon\cdot\frac{e^\epsilon-1}{e^\epsilon+k-1}, \] and for RAPPOR, \[ \rho_{\mathrm{RAPPOR}}=\epsilon\tanh(\epsilon/4). \] The paper also proves the exact worst-case zCDP guarantee for \(\eta\)-bounded range mechanisms: \[ \rho_{\mathrm{BR}}= \frac{\eta}{e^\eta-1} + \log\!\left(\frac{e^\eta-1}{\eta}\right) -1. \] An important nuance is that the KL-endpoint rule is not universal: for sufficiently large \(k\), the best zCDP bound for \(k\)-randomized response is not attained at \(\alpha\to 1\) [2510.25746]. ## 7. Tight accounting and empirical auditing in differential privacy A further body of work treats tightness as exact or near-exact privacy accounting under composition. "Tight Differential Privacy for Discrete-Valued Mechanisms and for the Subsampled Gaussian Mechanism Using FFT" [2006.07134] works with privacy loss distributions (PLDs) and the exact formula \[ \delta_{X/Y}(\varepsilon) = \delta_{X/Y}(\infty) + \int_\varepsilon^\infty (1-e^{-s})\,\omega_{X/Y}(s)\,ds. \] For discrete one-dimensional outputs, the paper places the PLD on a grid, forms left- and right-rounded distributions, and computes repeated convolutions by FFT, obtaining rigorous lower and upper bounds on the true \((\varepsilon,\delta)\) guarantee. The analysis gives strict bounds for the subsampled Gaussian mechanism, exact or near-exact accounting for the exponential mechanism on counting queries, and reports that in the binomial setting the accountant allows up to \(75\%\) reduction in noise variance at equal privacy compared to existing bounds [2006.07134]. "Tight Accounting in the Shuffle Model of Differential Privacy" [2106.00477] adapts the PLD/Fourier-accountant framework to shuffled mechanisms. For general \(\varepsilon\)-LDP shufflers it uses Feldman et al.’s dominating pair construction; for shuffled \(k\)-randomized response it extracts exact privacy-loss random variables from the Balle et al. analysis. The paper then reduces PLD construction from \(\mathcal O(n^2)\) to \(\mathcal O(n)\) terms, up to a logarithmic tolerance factor, by truncating binomial tails with Hoeffding’s inequality, with no noticeable change in the resulting \(\delta(\varepsilon)\)-upper bounds. It also shows that prior analytic composition bounds are substantially looser, especially for repeated shuffle rounds [2106.00477]. "Tight Group-Level DP Guarantees for DP-SGD with Sampling via Mixture of Gaussians Mechanisms" [2401.10294] avoids black-box conversion from example-level to group-level privacy by modeling a group change directly as a composition of mixture-of-Gaussians mechanisms. Under Poisson sampling, the per-step scalar sensitivity is dominated by \(\operatorname{Binom}(k,q)\); under fixed batch-size sampling, it is dominated by \[ 2\cdot \operatorname{Hypergeom}(B,n+k,k). \] The paper states that the resulting group-level guarantees are tight up to discretization error in implementation, assuming every intermediate iterate is released, and gives explicit one-dimensional constructions where the dominating pair is exact [2401.10294]. Finally, "Tight Auditing of Differentially Private Machine Learning" [2302.07956] studies empirical privacy lower bounds rather than formal accountants. In the white-box setting with access to all model updates, the audit uses \(f\)-DP and GDP rather than only final \((\varepsilon,\delta)\) guarantees. Because GDP composes tightly via \[ \mu = \sqrt{\mu_1^2 + \mu_2^2 + \cdots}, \] the method reduces the audit from thousands or millions of training runs to essentially two runs per observation step. The paper reports that around \(1{,}000\) observations can suffice for tight GDP-based estimates in the white-box setting, compared with \(100\) million or more observations for comparable tightness under direct \((\varepsilon,\delta)\)-based auditing, and uses the method to surface implementation bugs such as clipping after gradient averaging and biased noise sampling [2302.07956]. Taken together, these privacy papers show that tightness can mean exact mechanism-specific privacy parameters, certified composition theorems, or empirical lower bounds that nearly meet provable upper bounds. A plausible implication is that, in privacy as in mechanism design, the decisive technical issue is often not merely constructing a mechanism but identifying the correct benchmark and proving that no hidden slack remains once composition, sampling, or numerical approximation is accounted for. Source: https://www.emergentmind.com/topics/tight-mechanisms