2D BP: NFDH for small items

Dependencies:

  1. Geometric packing problems (BP, KS, SP) and rvol
  2. NFDH algorithm

Let $I$ be a 2D geometric bin packing instance, where items have height at most $\eps_H$ and width at most $\eps_W$. Let bins have height $H$ and width $W$.

The number of bins used by NFDH to pack $I$ is less than \[ \ceil{\frac{a(I)}{(W-\eps_W)(H-\eps_H)}}. \]

Moreover, for each bin except the last, the total area of items in the bin is more than $(W-\eps_W)(H-\eps_H)$. Consequently, if $a(I) \le (W-\eps_W)(H-\eps_H)$, then NFDH packs $I$ into a single bin.

Proof

Using NFDH for bin packing is equivalent to using NFDH for packing items into shelves and then packing the shelves into bins using the Next-Fit algorithm.

Let there be $m$ bins in NFDH's output. Consider the $q\Th$ bin, where $q < m$. Let there be $p$ shelves in this bin. Let $H_i$ be the height of the first item in the $i\Th$ shelf in this bin. Let $W_i$ be the sum of widths of items in the $i\Th$ shelf in this bin. Let $A_i$ be the total area of items in the $i\Th$ shelf in this bin. Let $H_{p+1}$ be the height of the first shelf in the $(q+1)\Th$ bin.

Note that $W_i > W - \eps_W$ and $\sum_{i=1}^{p+1} H_i > 1$ and $A_i \ge W_iH_{i+1}$. Hence, the total area of items in the $q\Th$ bin is \[ \sum_{i=1}^p A_i > \sum_{i=1}^p W_iH_{i+1} \gt (W-\eps_W)\sum_{i=1}^p H_{i+1} > (W-\eps_W)(H-\eps_H). \]

Let $B_i$ be the total area of items in the $i\Th$ bin. For $i < m$, we have $B_i > (W-\eps_W)(H-\eps_H)$. Therefore, \begin{align} & a(I) > \sum_{i=1}^{m-1} B_i > (m-1)(W-\eps_W)(H-\eps_H) \\ &\implies m < \frac{a(I)}{(W-\eps_W)(H-\eps_H)} + 1. \end{align} Since $m \in \mathbb{Z}$, we get \[ m \le \ceil{\frac{a(I)}{(W-\eps_W)(H-\eps_H)}}. \]

Dependency for:

  1. PTAS for 2D geometric density-restricted knapsack

Info:

Transitive dependencies:

  1. Bin packing and Knapsack
  2. Geometric packing problems (BP, KS, SP) and rvol
  3. NFDH algorithm