--- title: 'Minimum Bins Inequality: Theory & Applications' url: https://www.emergentmind.com/topics/minimum-bins-inequality-mbi type: topic --- # Minimum Bins Inequality: Theory & Applications Searching arXiv for recent papers on “Minimum Bins Inequality” and closely related formulations. Search query: "Minimum Bins Inequality Bin Packing Problem with Setups arXiv" The **Minimum Bins Inequality (MBI)** denotes a lower-bounding principle about the minimum number of bins required by a constrained allocation or occupancy process, but the term is used in more than one arXiv-adjacent context. In the most formal and explicit sense, it is a valid inequality for the **Bin Packing Problem with Setups (BPPS)** that strengthens a compact ILP by imposing a global lower bound on the number of opened bins [2509.10075]. In a second, closely related usage, the phrase refers to the question of whether a most balanced initial allocation minimizes the expected final residue in a stochastic ball-removal process; that question is answered affirmatively for the uniform-selection model by a coupling argument [2602.12523]. Related balls-and-bins, occupancy, hashing, and binomial-minimum papers do not always use the name MBI, but they study quantitatively similar threshold, occupancy, or minimum-load phenomena [2205.14494], [1810.13187], [2502.18611]. ## 1. Named inequality in BPPS In the BPPS literature, the MBI is introduced as a **global strengthening** of the natural compact formulation for the **Bin Packing Problem with Setups (BPPS)** [2509.10075]. In this model one is given a set of items \[ \mathcal I=\{1,2,\dots,n\}, \] identical bins of capacity \[ d\in \mathbb Z_{>0}, \] item weights \(w_i\), a partition of the items into classes \[ \mathcal P=\{\mathcal I_1,\mathcal I_2,\dots,\mathcal I_m\}, \] setup weights \(s_c\), setup costs \(f_c\), and a bin cost \(r\). If at least one item of class \(c\) is packed in a bin, then class \(c\) is active in that bin, and the bin incurs a one-time capacity reduction of \(s_c\) and a one-time cost of \(f_c\) for that class-bin pair [2509.10075]. The natural compact ILP uses binary variables \(x_{ib}\) for item assignment, \(y_{cb}\) for class activation, and \(z_b\) for bin usage. Its capacity constraints are \[ \sum_{i \in \mathcal I} w_i \, x_{ib} + \sum_{c \in \mathcal C} s_c \, y_{cb} \leq d \, z_b \qquad \forall b\in\mathcal B. \] The MBI is designed to strengthen the **global implication** of these per-bin capacity constraints on the total number of bins [2509.10075]. The exact inequality is \[ \sum_{b \in \mathcal B} z_{b} \;\geq\; \left\lceil\frac{\sum_{i \in \mathcal I} w_i \;+\; \sum_{c \in \mathcal C} \gamma_c \, s_c}{d} \right\rceil, \] where \[ \gamma_c = \left\lceil \frac{\sum_{i \in \mathcal I_c} w_i}{d-s_c}\right\rceil \qquad \forall c\in\mathcal C. \] The right-hand side is denoted \[ \underline{k} = \left\lceil\frac{\sum_{i \in \mathcal I} w_i + \sum_{c \in \mathcal C} \gamma_c s_c}{d}\right\rceil. \] The paper presents the MBI as a **single inequality** for a BPPS instance, not as a family indexed by subsets or bins, and it is described as a compact, closed-form valid inequality computable in linear time [2509.10075]. ## 2. Derivation, validity, and relation to MCIs The motivation for the MBI is the weakness of the LP relaxation of the natural formulation. The LP relaxation \(\mathcal F_N^{LP}\) is obtained by replacing the binary restrictions with \[ 0 \le x_{ib} \le 1,\qquad 0 \le y_{cb} \le 1,\qquad 0 \le z_b \le 1, \] and its optimal value is denoted \(\zeta(\mathcal F_N)\). The paper proves that an optimal LP solution is the completely symmetric fractional point \[ x_{ib}=\frac1k,\qquad y_{cb}=\frac1k,\qquad z_b=\frac{\sum_{i\in\mathcal I}w_i+\sum_{c\in\mathcal C}s_c}{kd}, \] with objective value \[ \zeta(\mathcal F_N) = \frac{r}{d}\left(\sum_{i\in\mathcal I} w_i+\sum_{c\in\mathcal C} s_c\right) +\sum_{c\in\mathcal C} f_c. \] The paper states that this relaxation can be **arbitrarily weak**, because setup mass can be shared fractionally among bins even though in any integral solution a setup must be paid each time a class appears in a bin [2509.10075]. To repair the classwise undercounting, the paper first introduces the **Minimum Classes Inequalities (MCIs)**: \[ \sum_{b \in \mathcal B} y_{cb} \;\geq\; \gamma_c := \left\lceil \frac{\sum_{i \in \mathcal I_c} w_i}{d - s_c}\right\rceil, \qquad c \in \mathcal C. \] These enforce that class \(c\) must be active in at least \(\gamma_c\) bins, because once class \(c\) is active, only \(d-s_c\) capacity remains for its items. With MCIs, the LP optimal value becomes \[ \zeta(\mathcal F_{NM}) = \frac{r}{d}\left(\sum_{i\in\mathcal I} w_i + \sum_{c\in\mathcal C}\gamma_c s_c\right) +\sum_{c\in\mathcal C}\gamma_c f_c, \] and the paper states that this restores the classical \(1/2\) worst-case guarantee: \[ \frac{\zeta(\mathcal F_{NM})}{\psi} > \frac12. \] However, MCIs remain **classwise** constraints on the \(y\)-variables and do not directly impose a lower bound on the total number of opened bins \(\sum_{b\in\mathcal B} z_b\). That remaining gap is exactly what the MBI fills [2509.10075]. The derivation is direct. Since each item must be packed, all item weights contribute to required capacity. Since MCIs imply that class \(c\) must be activated in at least \(\gamma_c\) bins, setup weight \(s_c\) must be paid at least \(\gamma_c\) times in total. Therefore any feasible integral solution must consume at least \[ \sum_{i\in\mathcal I}w_i + \sum_{c\in\mathcal C}\gamma_c s_c \] units of capacity, and since each used bin contributes at most \(d\), the number of used bins must satisfy the MBI [2509.10075]. The formal proof in the paper derives the inequality algebraically from the relaxation with MCIs. It uses \[ \sum_{b\in\mathcal B}\sum_{c\in\mathcal C} f_c y_{cb} \ge \sum_{c\in\mathcal C} \gamma_c f_c \] and a lower bound on the total objective to obtain \[ r\sum_{b\in\mathcal B} z_b \ge \frac{r}{d}\left(\sum_{i\in\mathcal I} w_i+\sum_{c\in\mathcal C}\gamma_c s_c\right). \] Since \(r>0\) and the \(z_b\) are integer in every feasible solution, the ceiling form follows [2509.10075]. ## 3. Structural role in the BPPS formulation The paper emphasizes that MCIs and the MBI are **complementary** rather than interchangeable [2509.10075]. The MCIs act at the class level, \[ \sum_{b\in\mathcal B} y_{cb}\ge \gamma_c \qquad \forall c\in\mathcal C, \] while the MBI acts at the global bin-count level, \[ \sum_{b\in\mathcal B} z_b \ge \underline{k}. \] The MBI constrains only the aggregate number of bins, whereas the MCIs ensure that the contribution of each class is explicitly enforced. The paper does not claim that one dominates the other, and mathematically they live in different variable spaces [2509.10075]. Conceptually, the MBI is a BPPS analogue of the classical bin-packing weight lower bound \[ \left\lceil\frac{\sum_i w_i}{d}\right\rceil. \] The BPPS adaptation replaces total item weight with \[ \sum_{i\in\mathcal I}w_i+\sum_{c\in\mathcal C}\gamma_c s_c, \] which includes the minimum setup burden implied by the MCIs. This can be read as a **setup-aware global volume bound** [2509.10075]. The paper also derives a paired **upper bound** on the number of bins used in an optimal solution. For each class \(c\), let \(\overline{\beta}_c\) be an upper bound on the minimum number of bins of capacity \(d-s_c\) needed to pack all items of class \(c\). Then \[ \overline{k}=\sum_{c\in\mathcal C}\overline{\beta}_c \] is a valid upper bound on the number of bins used in any optimal BPPS solution. This upper bound is used to replace the trivial choice \(k=n\) in the compact model, and the paper notes that the number of variables and constraints scales linearly with \(k\): variables \((n+m+1)k\), constraints \((n+1)k+n\) [2509.10075]. The addition of the MBI does not improve the asymptotic worst-case guarantee beyond the \(1/2\) already obtained with MCIs. The paper states that there exists a sequence of instances such that \[ \frac{\zeta(\mathcal F_{NMB})}{\psi}\to \frac12 \qquad \text{as } d\to\infty. \] This shows that the MBI is structurally important without changing the established worst-case ratio [2509.10075]. ## 4. Stochastic balancing interpretation of MBI A second usage of the term appears in the stochastic balls-and-bins removal problem studied in "About a Ball Removal Process on Bins" [2602.12523]. There the process starts from an occupancy vector \[ \vec n=(n_1,\dots,n_k)\in \mathbb N^k, \qquad \sum_{i=1}^k n_i=n, \] and repeatedly removes a ball from a **non-empty** bin chosen uniformly at random, stopping as soon as only one non-empty bin remains. If \(X_{\vec n}\) denotes the number of balls in the last non-empty bin at the stopping time \(T(\vec n)\), the objective is \[ f(\vec n):=\mathbb E[X_{\vec n}], \] and the optimization problem is \[ \min\left\{ f(\vec n): \vec n\in\mathbb N^k,\ \sum_{i=1}^k n_i=n\right\}. \] The paper states Will Ma’s conjecture as: the minimizing allocation is one that distributes the balls as evenly as possible, i.e. occupancies differ by at most \(1\) [2602.12523]. The main theorem proves that conjecture. It states that \[ \min\{ f(\vec n): \vec n\in\mathbb N^k,\ \sum_{i\in[k]} n_i=n\} \] is attained only by assignments satisfying \[ |n_i-n_j|\le 1\ \ \forall i,j. \] Equivalently, among all allocations of \(n\) balls into \(k\) bins, the as-balanced-as-possible allocation uniquely minimizes the expected number of balls remaining, up to permutation symmetry [2602.12523]. A key structural point is that the objective function is **not Schur-concave**. The result is instead proved through a local balancing or smoothing inequality: whenever one bin exceeds another by at least \(2\), moving one ball from the larger bin to the smaller bin strictly improves the objective. The relevant statement is \[ f(\vec n) > f(\vec n-e_1+e_i) \quad\text{whenever } n_1\ge n_i+2. \] For \(k=2\), the paper gives the recursion \[ f(a,b)=\frac12 f(a-1,b)+\frac12 f(a,b-1),\qquad a,b\ge 1, \] with boundary values \[ f(a,0)=a,\qquad f(0,b)=b, \] and proves \[ f(n_1,n_2)>f(n_1-1,n_2+1) \qquad\text{whenever } n_1\ge n_2+2. \] The general-\(k\) proof uses a coupling on the common sample space of selected bins, together with a stopping time \(S(\vec n)\) and a partition of sample paths into event classes that are compared one by one [2602.12523]. This usage differs from the BPPS inequality in formal status. In the BPPS paper, MBI is a named linear inequality. In the ball-removal paper, the term refers to a balancing principle or conjecture about the minimization of the expected final residue. This suggests that current usage treats MBI as a broader label for minimum-bin-count or minimum-residue phenomena driven by balancing. ## 5. Related occupancy and minimum-load results Several nearby papers do not use the exact label “Minimum Bins Inequality,” but they establish MBI-type results for occupancy thresholds, occupied-bin counts, or minimum-load surrogates. In "Balls and Bins -- Simple Concentration Bounds," the central quantity is \[ \rho_{m,k}:=\frac{m\,\|p\|_k}{k}, \qquad \|p\|_k := \left(\sum_{i=1}^n p_i^k\right)^{1/k}, \] for a nonuniform occupancy model in which \(m\) balls are thrown independently according to a distribution \(p\). The paper’s core inequality is \[ \Pr[\operatorname{Binom}(m,\|p\|_k)\ge k] \;\le\; \Pr[M\ge k] \;\le\; \binom{m}{k}\,\|p\|_k^k, \] where \(M\) is the maximum load after \(m\) throws. It interprets this through a phase change around \(\rho_{m,k}\approx 1\): if \(\rho_{m,k}\ll 1\), a \(k\)-loaded bin is unlikely; if \(\rho_{m,k}\gg 1\), a \(k\)-loaded bin is likely [2205.14494]. The same paper derives threshold statements for the waiting time \[ T_k := \min\{m:\, M\ge k\} \] and expectation bounds \[ \left(\frac{1}{e}\right)\left(\frac{k}{k+1}\right)\frac{k}{\|p\|_k} \;\le\; \mathbb E[T_k] \;\le\; \frac{k}{\|p\|_k}. \] These are not MBI in name, but they provide a minimum-throw-scale characterization for the first \(k\)-loaded bin [2205.14494]. In "Non-Empty Bins with Simple Tabulation Hashing," the relevant variable is the number of occupied bins \[ |h(X)|=\sum_{j=0}^{n-1} I_j, \qquad I_j=\mathbf 1[\exists x\in X:\ h(x)=j]. \] Under fully random hashing the baseline expectation is \[ \mu_0 = n\left(1-\left(1-\frac1n\right)^m\right). \] For simple tabulation hashing, the paper proves \[ \left|\mathbb E[|h(X)|]-\mu_0\right| \le \frac{m^{2-1/c}}{n} \] and the high-probability lower bound \[ |h(X)| \ge n\left(1-\left(1-\frac1n\right)^{\Omega(m)}\right) \qquad \text{whp}. \] It also gives lower-tail bounds around \(\mu_0\), but with an unavoidable extra term \(m^2/(nt^2)\), so the lower tail is not purely subgaussian in general [1810.13187]. In "Tight Bounds on the Binomial CDF, and the Minimum of i.i.d Binomials, in terms of KL-Divergence," the object is \[ Z=\min_{1\le i\le r} X_i \quad\text{when}\quad X_i \overset{\mathrm{iid}{\sim} \frac1n \mathrm{Bin}(n,p). \] For \(a\le p\), the paper proves \[ \log \mathbb{P}(X \le a) \in -n KL(a\Vert p) \pm \left(4\log(n+1)+\left[\log \frac{p}{1-p}\right]_+\right), \] and then shows that, with explicit finite-sample correction \[ \Delta(\delta,p,n) = \log \frac{1}{\delta/2} +4\log(n+1) +\left[\log\frac{p}{1-p}\right]_+, \] the minimum satisfies \[ KL(Z\Vert p)\in \frac{\log r \pm \Delta(\delta,p,n)}{n} \] with probability \(1-\delta\), except in the regime where the theorem implies \(Z=0\) with high probability [2502.18611]. These papers show that, even when the exact label MBI is absent, a common mathematical pattern recurs: thresholding or bounding a minimal occupancy-related quantity by translating local collision or activation constraints into a global condition. ## 6. Computational significance, limitations, and non-uniform variants The strongest computational evidence for the MBI as a named inequality comes from BPPS. The experiments compare four variants: 1. \(\mathcal F_N\), 2. \(\mathcal F_{NM}\), 3. \(\mathcal F_{NMB}\), 4. \(\mathcal F_{NMBU}\). Over 480 instances, the paper reports: - \(\mathcal F_N\): 162 solved to optimality, average gap \(14.4\%\); - \(\mathcal F_{NM}\): 220 solved, average gap \(6.7\%\); - \(\mathcal F_{NMB}\): 253 solved, average gap \(6.0\%\); - \(\mathcal F_{NMBU}\): 268 solved, average gap \(5.7\%\). On the subset of 286 instances solved optimally by at least one formulation, the overall average integrality gap is reported as: - \(\mathcal F_N\): \(16.8\%\), - \(\mathcal F_{NM}\): \(8.0\%\), - \(\mathcal F_{NMB}\): \(2.2\%\). The paper also notes that \(\zeta(\mathcal F_{NMB})=\zeta(\mathcal F_{NMBU})\), so the LP improvement from \(\mathcal F_{NM}\) to \(\mathcal F_{NMB}\) is attributable specifically to the MBI, whereas the upper bound \(\overline{k}\) reduces model size rather than the LP value [2509.10075]. In the stochastic ball-removal setting, the main limitation concerns non-uniform bin-selection rules. The paper explicitly studies a non-uniform two-bin extension with recursion \[ f_p(a,b)= \begin{cases} a, & b=0,\\ b, & a=0,\\ p\,f_p(a-1,b)+(1-p)\,f_p(a,b-1), & \text{otherwise}, \end{cases} \] and derives \[ f_p(a,1)=a-\frac{p-p^a}{1-p}, \qquad f_p(a,2)=a+2-\frac{2}{1-p}+p^a\left(a+2+\frac{2p}{1-p}\right). \] Using \(n=6\), \(k=2\), and \[ p=\frac56,\qquad 1-p=\frac16, \] the paper shows that the proportional allocation \((5,1)\) is worse than \((4,2)\): \[ f_{5/6}(5,1)=\frac{3125}{1296}\approx 2.41, \qquad f_{5/6}(4,2)=\frac{139}{81}\approx 1.72. \] Thus \[ f_{5/6}(4,2)