BoundedME Algorithm Overview

• BoundedME algorithm is a dual-concept approach that applies bounded-support noise for differential privacy and boundedness conversion in mixed linear arithmetic.
• In the differential privacy context, BoundedME adds independent bounded noise to each query, guaranteeing absolute error bounds and up to 30% improved empirical accuracy over Gaussian methods.
• For mixed linear arithmetic, BoundedME preprocesses constraints using a double-bounded reduction and Mixed-Echelon-Hermite transformation to ensure decidable, finite termination.

The term “BoundedME algorithm” designates two unrelated but prominent algorithms in distinct research contexts. This article documents both: (1) the “Bounded-Noise Mechanism” (BoundedME) for differential privacy, developed by Dagan & Kur (Dagan et al., 2020), and (2) the “BoundedME” preprocessing pipeline for boundedness conversion in mixed linear arithmetic, as formulated by Niemetz et al. (Bromberger, 2018). While they share the same shorthand and a focus on boundedness, their technical aims and fields are entirely separate.

1. BoundedME for Differential Privacy

The Bounded-Noise Mechanism ("BoundedME") is an asymptotically optimal $(\epsilon,\delta)$-differentially private method for answering $k$ (potentially adaptive) $\Delta$-sensitive queries. Distinguishing features include the addition of independent, bounded-support noise to each query, yielding absolute error guarantees and practical advantages in empirical performance over the classical Gaussian mechanism (Dagan et al., 2020).

1.1. Formal Setup and Parameters

BoundedME for DP takes the following inputs:

• $x \in X^n$: dataset of size $n$.
• $\epsilon \in (0,1]$, $\delta \in (0,1)$: privacy parameters.
• $k \in \mathbb N$: number of queries (possibly adaptively chosen).
• Each query $q_i : X^n \to \mathbb R$ is $\Delta$-sensitive: for neighboring $k$0, $k$1.
• “Shape” function $k$2 (e.g., $k$3, $k$4).
• The scale $k$5, where $k$6 depends on $k$7.

1.2. Bounded Noise Distribution

The base (unscaled) density over $k$8 is

$k$9

The scaled, bounded-support noise over $\Delta$0:

$\Delta$1

Concrete choices for $\Delta$2 include polynomially and doubly-exponentially diverging options. Noise is sampled i.i.d. for each query, guaranteeing that every output lies in a finite interval $\Delta$3.

1.3. Algorithmic Description

$\Delta$4

1.4. Theoretical Guarantees

• Upper Bound (Theorem): For all $\Delta$5 meeting the required analytical properties and for $\Delta$6 above a regime-dependent threshold $\Delta$7, BoundedME is $\Delta$8-DP for $\Delta$9 $x \in X^n$0-sensitive queries.
• Absolute Error Bound: With probability one,

$x \in X^n$1

and this is a deterministic statement due to the bounded noise.

• Optimality: No bounded, i.i.d., unimodal noise mechanism can achieve strictly smaller bounds in the relevant $x \in X^n$2 regime.

1.5. Proof Techniques and Lemmatic Structure

The analysis reduces DP to a high-dimensional “density-ratio” tail event and controls its probability via Taylor expansions and careful bounding of sums involving derivatives of $x \in X^n$3. Sub-Gaussianity emerges for the leading fluctuation term via a truncation+MGF argument. The second-derivative term is bounded via growth-control of $x \in X^n$4. These steps establish both privacy and absolute error guarantees.

1.6. Complexity and Parameter Impact

Sampling noise dominates per-query cost; practical rejection sampling applies. $x \in X^n$5 depends on $x \in X^n$6's analytical properties. By tuning the shape parameter ($x \in X^n$7) in $x \in X^n$8, one may trade off $x \in X^n$9 versus $n$0 and overall error at given privacy.

1.7. Empirical Comparison

Empirical results show that, for parameters such as $n$1, $n$2, and $n$3, BoundedME yields absolute error (measured as the 95%-quantile max error) up to 30% lower than the Gaussian mechanism in both worst-case and high-probability tails. Publicly available code supports numerical evaluation (Dagan et al., 2020).

2. BoundedME for Boundedness in Mixed Linear Arithmetic

An independently named BoundedME algorithm addresses decidability and termination in mixed-satisfiability for systems of linear mixed arithmetic—that is, constraints of the form $n$4 with both rational and integer variables (Bromberger, 2018).

2.1. Problem Setting

Consider $n$5 for $n$6 with the first $n$7 variables rational and the remainder integer. Unbounded directions generally render naïve branch-and-bound procedures non-terminating.

2.2. Double-Bounded Reduction

A direction in $n$8 is “bounded” if, for its associated row $n$9, there exist $\epsilon \in (0,1]$0 with $\epsilon \in (0,1]$1 holding for all feasible $\epsilon \in (0,1]$2; equivalently, bounding directions $\epsilon \in (0,1]$3 are those with $\epsilon \in (0,1]$4. An equality-basis computation enables partitioning of $\epsilon \in (0,1]$5 into:

• $\epsilon \in (0,1]$6: unbounded constraints,
• $\epsilon \in (0,1]$7: “double-bounded” system.

Feasibility is preserved: $\epsilon \in (0,1]$8 is satisfiable iff the system formed by these two blocks is.

2.3. Mixed-Echelon-Hermite Transformation

The double-bounded system $\epsilon \in (0,1]$9 is converted to Mixed-Echelon-Hermite Normal Form (MEHNF): a block-lower-triangular structure where the rational-variable block is reduced to column-echelon, and the integer-variable block is in Hermite normal form. The transformation is achieved by a specific block-mixed invertible transformation $\delta \in (0,1)$0, computable in polynomial time (Bareiss elimination of rationals, column zeroing, Kannan–Bachem Hermite algorithm).

2.4. High-Level Algorithm

Algorithm BoundedME(A ∈ ℚ^{m×n}, b ∈ ℚ^m)
Input: A x ≤ b, first n₁ vars rational, last n₂ integer
Output: “SAT” with solution x* or “UNSAT” with Farkas certificate

1. E := EqualitiesBasis(A, 0)
   (A_u, b_u, D, u) := Split(A, b, E)

2. if RationalUnsat(D, u) then return UNSAT
   if AbsolutelyUnbounded(A_u) then return SAT, x* := CubeTest(A_u)

3. (H, V) := MEHTransform(D)

4. (res, y*) := SolveBounded(H, u)
   if res == UNSAT then return UNSAT
   x* := V y*
   return SAT, x*

2.5. Correctness and Certificates

Mixed-satisfiability and Farkas certificates lift through the transformation: for solutions $\delta \in (0,1)$1 of $\delta \in (0,1)$2, $\delta \in (0,1)$3 solves the original system. For unsatisfiability, Farkas multipliers propagate backward (via Corollary 3.8 of (Bromberger, 2018)).

2.6. Computational Complexity

All preprocessing (equality-basis, splitting, MEH transformation) is polynomial-time. Final branch-and-bound on the triangularized bounded system is finitely terminating (though, in the worst case, exponential in the integer variables). Empirically, overheads are negligible and the reduced system is often small.

2.7. Empirical Results

BoundedME’s implementation in SPASS-IQ v0.2 was tested on partially-unbounded mixed-arithmetic benchmarks (e.g., SlackedQFLIA, RandomUnbd, and mixed-type variants). In direct comparison to leading SMT solvers (CVC4, MathSAT5, SMTInterpol, Yices, Z3), BoundedME solved more cases and was typically faster. Most transformations proved lightweight, resulting in compact and tractable final search.

3. Comparative Summary Table

Context	Purpose	Key Steps / Properties
Differential Privacy	Asymptotically optimal DP answering of $\delta \in (0,1)$4 queries	Adds i.i.d., bounded-support noise $\delta \in (0,1)$5; absolute error bound
Mixed Arithmetic	Guarantee termination in bounded search space	Double-bounded reduction, MEH normalization, lifts certs and solns

4. Distinct Terminology and Research Impact

The homonymy of BoundedME arises from distinct lines of work: Dagan & Kur’s mechanism for bounded-noise DP, and Niemetz et al.'s pipeline for preprocessing mixed arithmetic systems. Each addresses a core bottleneck—error in DP, non-termination in arithmetic search—by exploiting “boundedness” in a fundamental sense. Both lines report substantial empirical improvements and foundational optimality results. Their theoretical analyses, spanning advanced privacy measures and algebraic transformation, have directly influenced developments in their respective domains (Dagan et al., 2020, Bromberger, 2018).