Inter-Block Complexity in Algorithmic Systems

• Inter-block complexity is the study of how computational cost, convergence rate, and algorithmic difficulty scale with block variable decomposition and update interactions.
• It underpins methods in block-coordinate descent, signal processing, and Boolean function theory, influencing metrics like iteration complexity and block sensitivity.
• Research highlights include advances in convergence bounds, improved dependencies on block count, and practical reductions in search complexity for applications like video coding.

Inter-block complexity captures how computational cost, convergence rate, or algorithmic hardness scale with respect to the decomposition of variables, updating schemes, or interactions across blocks in block-structured problems and systems. Its analysis is central in block-coordinate optimization, block-partitioned algorithms in signal processing, coding, Boolean network dynamics, and block-structured complexity in Boolean function theory. The concept encompasses both quantitative measures of algorithmic resource consumption per iteration (e.g., number of block operations, per-iteration cost, stationarity residual rate) and qualitative complexity classification (e.g., computational hardness of dynamic questions under block-parallel updates).

1. Block-Partitioned Structures and Inter-Block Algorithmics

Block decomposition is the foundational principle guiding both the deployment and the analysis of inter-block complexity. In composite optimization, variables are split into m blocks, and the system’s objective function is decomposed accordingly, e.g., in the m-block nonsmooth convex composite problem: $$\min_{x_1,\ldots,x_m} F(x_1,\ldots,x_m) = g(x_1,\ldots,x_m) + \sum_{i=1}^m h_i(x_i)$$ where $g$ is block-Lipschitz smooth and each $h_i$ is convex (possibly nonsmooth or indicator). The qualitative shape and interaction among blocks—through smoothness, constraints, or coupling—determine the achievable complexity rates and the nature of update dependencies (Hong et al., 2013).

In dynamical systems, such as Boolean automata networks with block-parallel update modes, inter-block complexity reflects the impact of repeated or staggered block updates, often inflating the temporal and computational depth relative to traditional block-sequential paradigms (Perrot et al., 2024).

2. Inter-Block Complexity in Block-Coordinate Descent and MM Frameworks

Block-coordinate descent (BCD) and block majorization-minimization (BMM) algorithms epitomize algorithmic settings where inter-block complexity is directly quantifiable via iteration complexity bounds. Central results under the Block Successive Upper-bound Minimization (BSUM) and related frameworks include the following:

• General multi-block convex problems admit global sublinear convergence rates $\mathcal{O}(1/r)$, where r is the iteration (or cycle) index. The leading constants in these bounds explicitly depend on the number of blocks $m$, block coupling terms (e.g., Lipschitz constants $G_i$ capturing cross-block curvature), and the strength of majorizing surrogates (strong convexity constants $\gamma_i$ of the per-block upper bounds) (Hong et al., 2013, Li et al., 2023).
• Exact block-coordinate minimization (BCM) achieves the same $\mathcal{O}(1/r)$ rate without requiring per-block strong convexity. For two-block Gauss-Seidel setups, Nesterov-type momentum yields an accelerated $\mathcal{O}(1/r^2)$ rate (Hong et al., 2013).
• Recent analyses show that the iteration complexity dependence on the number of blocks $K$ can be improved from the classical $g$0 to within a $g$1 factor of the global gradient-descent rate, even for cyclic block updates, thus removing the direct linear penalty for block multiplicity (Sun et al., 2015).
• In the randomized block-coordinate descent context, expected-value complexity to $g$2-accuracy is typically $g$3 for n blocks, and high-probability bounds are similarly sharp, with acceleration techniques recovering the optimal $g$4 decay (Lu et al., 2013).
• In block-majorization-minimization over products of Riemannian manifolds, the overall iteration complexity $g$5 demonstrates a linear scaling in the number of blocks $g$6, with constant factors depending on surrogate regularization and global geometric parameters (Li et al., 2023).

Tabulated dependencies from several frameworks are shown below:

Algorithmic Setting	Inter-Block Complexity Scaling	Reference
General BSUM/BCPG	$g$7	(Hong et al., 2013)
Cyclic BCD (quadratic, convex)	$g$8	(Sun et al., 2015)
Randomized Block-Coordinate Descent	$g$9	(Lu et al., 2013)
Riemannian Block Majorization-Minimization	$h_i$0	(Li et al., 2023)

3. Inter-Block Complexity in Constrained and Nonconvex Multi-Block Methods

For nonconvex, linearly constrained, or composite settings, inter-block complexity is impacted not only by block count but also by parallel versus sequential block update structures:

• Jacobi-type non-Euclidean ADMM for multi-block linearly constrained nonconvex programs achieves nonasymptotic $h_i$1 iteration complexity (in squared stationarity residual) for $h_i$2 blocks with parallel Bregman-proximal updates, provided block-specific regularization constants remain positive after accounting for all cross-block penalty terms, the penalty parameter $h_i$3, and the dual relaxation $h_i$4 (Melo et al., 2017).
• In all such frameworks, block-coupling appears through block-specific constants in potential function and descend bounds; ensuring $h_i$5 for each block (where $h_i$6 is the effective strong convexity minus coupling-induced decrements) is necessary for rate guarantees (Melo et al., 2017).

4. Inter-Block Search and Complexity in Signal Processing and Coding

In video coding and signal processing algorithms, block-structured search determines inter-block computational cost:

• Block matching for motion estimation (e.g., in inter-frame video codecs) has traditional complexity per frame scaling as $h_i$7, where $h_i$8 is the per-block search radius. Sensor-aided techniques can exploit inertial motion estimates and per-block depth maps to collapse the search region per block from $h_i$9 positions to $\mathcal{O}(1/r)$0, with $\mathcal{O}(1/r)$1. This achieves a complexity reduction factor on the order of $\mathcal{O}(1/r)$2, empirically ≈2.5×, with negligible PSNR degradation (Khoury et al., 2020).

Method	Search Area per Block	Complexity Scaling	Image Quality Loss
Full search	$\mathcal{O}(1/r)$3	$\mathcal{O}(1/r)$4	None
Sensor-aided w/ depth	$\mathcal{O}(1/r)$5 ($\mathcal{O}(1/r)$6)	$\mathcal{O}(1/r)$7	$\mathcal{O}(1/r)$80.1dB PSNR

• Partition map-based block coding (e.g., VVC inter coding with quadtree plus nested multi-type tree splits) recasts the recursive search over exponentially many block partitionings, replacing most with fast neural predictions. Here, complexity scales linearly with the number of pixels for network inference and subexponentially for the few blocks falling back to full RDO search. Dual-threshold pruning and partition-adaptive warping reduce encoding time by over 50% at a bit-rate increase of ≈2% (Feng et al., 25 Apr 2025).

5. Block Sensitivity and Inter-Block Complexity in Boolean Function Theory

In Boolean function complexity, inter-block phenomena manifest in several key measures:

• Block sensitivity $\mathcal{O}(1/r)$9 generalizes classical sensitivity by counting the number of disjoint input blocks whose simultaneous flipping changes the function value, making it a fundamental inter-block complexity measure. Ambainis et al. provide the bound: $m$0 where $m$1 is the sensitivity, and refine previous $m$2 bounds to strictly better exponential scaling (Ambainis et al., 2013). For constant min sensitivity, $m$3 (linear relation).

Quantities	Definition	Upper Bound
Sensitivity $m$4	Max bits s.t. flip alters output	–
Block sensitivity $m$5	Max disjoint blocks s.t. flip alters	$m$6

• The precise relation between sensitivity and block sensitivity is a major open problem, with the exponential gap remaining. However, limitations stem from inter-block combinatorics, as block-disjoint dependencies condition attainable upper bounds.

6. Computational Complexity in Block-Parallel Dynamical Systems

Boolean automata networks with block-parallel update modes illustrate how certain block update schemes can dramatically escalate problem hardness:

• Evaluating reachability, preimage, and fixed point questions becomes PSPACE-complete under block-parallel updates, in contrast to P or $m$7 complexity in block-sequential settings. The super-polynomial increase in “internal time” through repeated block updates (with period length $m$8 up to $m$9 for $G_i$0 nodes) enables simulation of exponentially deep computations in a single high-level step (Perrot et al., 2024).
• Despite this, purely local properties (e.g., bijectivity of the update map) remain only coNP-complete, as they reduce to per-substep behavior, unlike reachability-type questions that involve the global block-interaction structure.

7. Practical Implications and Design Guidelines

Understanding inter-block complexity elucidates trade-offs in algorithm and system design:

• Algorithm designers should partition variables to minimize inter-block Lipschitz constants; use upper bounds or surrogates tight enough for fast per-block progress yet simple enough for tractable subproblems; and choose update and regularization strategies that balance cheap iterations (random/greedy or parallel update) against global progress rates.
• Signal processing systems (e.g., video encoders) can exploit side information and decimation techniques to localize inter-block search, preserve fidelity, and sharply lower arithmetic complexity.
• In theoretical regimes (Boolean complexity, dynamical systems), block interaction models critically affect upper/lower bounds and complexity class placement, revealing nontrivial transitions in computational hardness and iteration complexity that cannot be inferred from single-block or sequential paradigms.

The inter-block complexity landscape remains active, encompassing both practical and structural directions: optimizing trade-offs in large-scale algorithmic deployments, analyzing block-induced hardness transitions, and refining theoretical bounds through sharper block-aware combinatorial and convex-analytic arguments.