Block Rotation Strategy

• Block Rotation Strategy is a design principle that partitions structured data into blocks and applies independent rotations to achieve localized transformations while preserving global objectives.
• It is employed in various domains such as adversarial attacks, compression, post-training quantization, factor analysis, and cryptographic key management to balance local and global effects.
• The strategy adapts through variations like randomized rotations, block-diagonal transforms, and blockwise mean rotations, enabling trade-offs between computational efficiency and performance improvements.

Searching arXiv for relevant papers on block rotation strategies across adversarial attacks, compression, and quantization. Searching arXiv for the specific papers cited in the provided source block. Block rotation strategy denotes a family of methods in which an input, feature vector, tensor, loading matrix, or operational resource pool is partitioned into blocks and then transformed by rotations or rotation-like refresh operations at block granularity rather than globally. In contemporary arXiv literature, the phrase spans several distinct lineages: geometric block rotations of image patches for adversarial transferability (Wang et al., 2023), per-block alignment before the discrete cosine transform in compression (Guerreiro et al., 2014), block-diagonal orthogonal transforms for post-training quantization of LLMs (Sanjeet et al., 29 Jan 2026, Choi et al., 2 May 2025, Shao et al., 6 Nov 2025, Lin et al., 20 Apr 2026, Xu et al., 19 May 2026, Gu et al., 27 Jan 2026), and randomized rotated block quantization on the sphere (Ann et al., 19 May 2026). Related but non-isomorphic uses also appear in exploratory factor analysis, where blockwise mean cross-loadings define an oblique target rotation (Beauducel et al., 2023), and in security systems, where rotation refers to key or endpoint refresh under block or pool constraints (Chen et al., 25 Dec 2025, Maiti, 8 Jun 2026). This breadth suggests that block rotation is best understood not as a single algorithm but as a recurring design principle for imposing locality on a transformation while preserving a global objective.

1. Scope and recurrent construction

A recurring pattern across the literature is to first partition a structured object into blocks, then apply either independent blockwise rotations, a shared block-diagonal orthogonal transform, or a rotation schedule constrained by the block structure of the downstream system. The stated objectives vary: corrupting or diversifying attention heatmaps in transfer attacks, aligning edges to DCT axes, diffusing activation outliers before quantization, stabilizing factor correlations in small samples, or regulating exposure under key-rotation and moving-target defense models.

Area	Representative block object	Stated purpose
Adversarial attacks	$n \times n$ image patches	Disrupt attention heatmaps and improve transferability
Compression	$8 \times 8$ image blocks	Align dominant edges with DCT axes
PTQ and MXFP4	channel groups or block-diagonal transforms	Suppress outliers, reduce dynamic range, improve PPL/accuracy
Factor analysis	blocks of salient loadings	Reduce sampling-error bias
Cryptography and operational security	byte blocks, file batches, domain pools	Control diffusion, key reuse, or availability

In the most algebraic formulations, the transform is block diagonal. MixQuant partitions a $d$-dimensional activation into $n=d/b$ blocks of size $b$ and applies an independent normalized Hadamard $R\in\mathbb{R}^{b\times b}$ in each block (Sanjeet et al., 29 Jan 2026). BRQ uses

$$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$

with $R_b^\top R_b=I_g$, so that mixing remains confined to each MXFP4 group (Shao et al., 6 Nov 2025). DuQuant++ adopts

$R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$

with a shared $32\times 32$ orthogonal block to match the microscaling group size (Lin et al., 20 Apr 2026). LoPRo likewise uses a block-diagonal Walsh–Hadamard structure after a permutation, while preserving a leading identity block for the most salient columns (Gu et al., 27 Jan 2026). These constructions formalize the same locality constraint: full mixing is replaced by structured mixing inside blocks.

2. Vision, compression, and tensor rotation

In adversarial machine learning, block rotation appears as a randomized input transformation. “Boosting Adversarial Transferability by Block Shuffle and Rotation” observes that existing input transformation based attacks result in different attention heatmaps on various models, and that breaking the intrinsic relation of the image can disrupt the attention heatmap of the original image (Wang et al., 2023). The method partitions an image $8 \times 8$0 into $8 \times 8$1 contiguous patches $8 \times 8$2, applies a random shuffle $8 \times 8$3, then rotates each patch independently by $8 \times 8$4, producing

$8 \times 8$5

This transform is inserted into an MI-FGSM loop by averaging the gradients from $8 \times 8$6 independently transformed images before momentum and projection updates. The reported defaults are $8 \times 8$7, $8 \times 8$8, $8 \times 8$9, $d$0, $d$1, and $d$2. On Inc-v3$d$3Inc-v4, the ablation gives MI-FGSM $d$4, BS $d$5, BR $d$6, and BSR $d$7. Averaged over six held-out models, the paper reports approximately $d$8 success for MI-FGSM, $d$9 for shuffle only, $n=d/b$0 for rotation only, and $n=d/b$1 for full BSR, isolating a large contribution from the rotation step.

In transform coding, “Maximizing compression efficiency through block rotation” treats block rotation as geometric pre-alignment for separable DCT bases (Guerreiro et al., 2014). Because horizontal and vertical edges concentrate energy in a few low-frequency coefficients, while slanted edges scatter energy across many coefficients, each $n=d/b$2 block is rotated by an angle $n=d/b$3 chosen to minimize reconstruction MSE after retaining the $n=d/b$4 largest DCT coefficients. The paper gives both an exhaustive-search criterion and a fast gradient-histogram angle estimator. Two variants are distinguished. Variant A preserves constant sampling rate by embedding the rotated rhombus into a square of up to $n=d/b$5, at the cost of variable-size DCTs and higher complexity. Variant B preserves constant $n=d/b$6 block size, reducing complexity but losing up to $n=d/b$7 of spatial sampling when $n=d/b$8. On the Lenna image, the method raises PSNR by up to $n=d/b$9 dB at $b$0–$b$1 retained coefficients, still yields $b$2 dB at $b$3, and loses its advantage above $b$4 coefficients.

A higher-order tensor variant appears in quaternion video processing. The $b$5-block circulant operator $b$6 is diagonalized by the mode-3 DFT, enabling a slice-wise quaternion polar decomposition in the Fourier domain and reconstruction of a unitary tensor $b$7 under the QT-product (Zhang et al., 12 Feb 2026). Applied to video rotation, this strategy yields a tensor of coherent per-frame rotations while preserving color-vector norms. On a 300-frame test, QT-Polar attains $b$8, $b$9, and $R\in\mathbb{R}^{b\times b}$0, outperforming direct multiplication and naïve per-frame methods in temporal consistency.

3. Blockwise orthogonal rotation in post-training quantization

The most extensive contemporary use of block rotation strategy is in PTQ for LLMs. Here the central problem is outlier suppression under hardware and format constraints. “MixQuant” gives a non-asymptotic analysis of block Hadamard rotations, showing that for a block partition $R\in\mathbb{R}^{b\times b}$1 of size $R\in\mathbb{R}^{b\times b}$2,

$R\in\mathbb{R}^{b\times b}$3

where $R\in\mathbb{R}^{b\times b}$4 (Sanjeet et al., 29 Jan 2026). The bound is controlled by the block with the largest $R\in\mathbb{R}^{b\times b}$5 mass, which motivates permutations before rotation. MixQuant therefore calibrates a permutation $R\in\mathbb{R}^{b\times b}$6 by a greedy mass diffusion algorithm that balances expected blockwise $R\in\mathbb{R}^{b\times b}$7 norms across a calibration set, and then exploits permutation-equivariant regions of transformer subgraphs to fold the permutation into weights offline. On Llama3 1B INT4 with block size $R\in\mathbb{R}^{b\times b}$8, the no-permutation baseline gives PPL $R\in\mathbb{R}^{b\times b}$9, MixQuant gives $$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$0, and the full-vector rotation reference gives $$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$1, so MixQuant recovers about $$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$2 of the benefit of full-vector rotation at $$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$3.

“Grouped Sequency-arranged Rotation” modifies the internal ordering of Walsh–Hadamard blocks rather than the permutation of coordinates (Choi et al., 2 May 2025). The rotation matrix is block diagonal,

$$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$4

where $$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$5 sorts rows by sequency, clustering similar oscillation patterns. The stated purpose is to reduce per-block dynamic range and hence the quantization MSE bound. In the paper’s WikiText-2 results for QuaRot, W2A16 gives PPL $$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$6 for GH, $$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$7 for GW, $$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$8 for LH, and $$R_{\mathrm{block}}=\mathrm{diag}(R_1,R_2,\dots,R_B),$$9 for GSR, with corresponding zero-shot accuracies $R_b^\top R_b=I_g$0, $R_b^\top R_b=I_g$1, $R_b^\top R_b=I_g$2, and $R_b^\top R_b=I_g$3. Under W2A4, GSR yields PPL $R_b^\top R_b=I_g$4 and zero-shot $R_b^\top R_b=I_g$5. The same paper states that replacing $R_b^\top R_b=I_g$6 with GSR in SpinQuant and OSTQuant yields $R_b^\top R_b=I_g$7–$R_b^\top R_b=I_g$8 points lower PPL and $R_b^\top R_b=I_g$9–$R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$0 points higher accuracy.

A separate line addresses MXFP4, where global rotation becomes problematic because the format uses per-block power-of-two scaling. “Block Rotation is All You Need for MXFP4 Quantization” argues that global orthogonal rotation redistributes outlier energy into many intermediate-sized values, which then incur large block-scale quantization error under MXFP4’s coarse PoT scale (Shao et al., 6 Nov 2025). BRQ replaces the global transform by a block-diagonal orthonormal rotation aligned to the hardware group size $R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$1. For dimension $R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$2, the paper contrasts $R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$3 operations for global rotation with $R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$4 for block-wise rotation, a $R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$5 speedup, and reports a $R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$6 lower latency overhead in prefill compared to global Hadamard rotation. On LLaMA-3 8B W4A4, MXFP4 RTN gives PPL $R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$7 and average zero-shot $R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$8, QuaRot$R=\mathrm{BlockDiag}(R_b,R_b,\dots,R_b),$9 gives $32\times 32$0 and $32\times 32$1, and BRQ gives $32\times 32$2 and $32\times 32$3. On LLaMA-3.2 1B, the corresponding values are $32\times 32$4, $32\times 32$5, and $32\times 32$6.

“DuQuant++” adopts the same MXFP4 group size $32\times 32$7 but makes the rotation outlier-aware rather than randomized (Lin et al., 20 Apr 2026). After the SmoothQuant rescaling $32\times 32$8, it inserts a single block-diagonal orthogonal matrix $32\times 32$9, with one shared $8 \times 8$00 block learned by a greedy sequence of Givens-like rotations on calibration activations. Because each MXFP4 group has its own scaling factor, the cross-block variance issue that required two rotations and a zigzag permutation in the original DuQuant is removed. The result is one $8 \times 8$01 matmul instead of two plus a permutation, described as roughly halving the online rotation cost. On LLaMA-3-8B W4A4, MR-GPTQ gives average zero-shot $8 \times 8$02 and PPL $8 \times 8$03, DuQuant++ gives $8 \times 8$04 and $8 \times 8$05, and DuQuant+++GPTQ gives $8 \times 8$06 and $8 \times 8$07, compared with the FP16 baseline $8 \times 8$08 and $8 \times 8$09.

TORQ generalizes the MXFP4 problem into two structural imbalances: extreme inter-block variance imbalance and intra-block codebook utilization imbalance (Xu et al., 19 May 2026). Its first rotation $8 \times 8$10 uses the Schur–Horn theorem to flatten the diagonal of $8 \times 8$11, while its second rotation $8 \times 8$12 maximizes codebook entropy inside each MXFP4 block. On Qwen3-32B, this reduces WikiText perplexity from $8 \times 8$13 for GPTQ or $8 \times 8$14 for QuaRot down to $8 \times 8$15, compared with $8 \times 8$16 for BF16, and increases average zero-shot accuracy from $8 \times 8$17 to $8 \times 8$18, compared with $8 \times 8$19 for BF16. The inverse transforms are fused into the next linear layer’s weights, so the online inverse cost is zero.

LoPRo relocates blockwise rotation to the residual matrix after low-rank approximation (Gu et al., 27 Jan 2026). After computing $8 \times 8$20 and residual $8 \times 8$21, it sorts columns by the ratio $8 \times 8$22, leaves the first $8 \times 8$23 columns untouched, and applies Walsh–Hadamard blocks of size $8 \times 8$24 to the remaining columns. The rationale is explicit: preserve the quantization accuracy of the most salient column blocks while rotating columns of similar importance. On LLaMA2-7B at 2-bit, the paper reports PPL $8 \times 8$25 for GPTQ, QuIP#, LoPRo, and LoPRo$8 \times 8$26, with zero-shot accuracy $8 \times 8$27. It also reports up to a $8 \times 8$28 speedup and states that Mixtral-8x7B quantization completes within $8 \times 8$29 hours while reducing perplexity by $8 \times 8$30 and improving accuracy by $8 \times 8$31.

“Block-Sphere Vector Quantization” shifts the focus from outlier smoothing to geometry preservation after a Haar random rotation $8 \times 8$32 (Ann et al., 19 May 2026). The rotated vector $8 \times 8$33 is partitioned into contiguous blocks of length $8 \times 8$34, and each block is quantized against a codebook on the unit $8 \times 8$35-ball derived from the spherical marginal $8 \times 8$36. The paper proves improvements over EDEN, RabitQ, and TurboQuant for both reconstruction MSE and expected inner-product distortion. For $8 \times 8$37, it gives

$8 \times 8$38

for $8 \times 8$39, and for $8 \times 8$40,

$8 \times 8$41

In KV-cache quantization for Llama-3.1-8B, the needle-in-a-haystack score is $8 \times 8$42 for Block-sphere, compared with $8 \times 8$43 for EDEN, $8 \times 8$44 for RaBitQ, and $8 \times 8$45 for TurboQuant.

Taken together, these PTQ papers support a common interpretation: block rotation is a locality-constrained orthogonalization strategy whose effectiveness depends on how well block boundaries match the hardware scaling rule, the outlier geometry, and the quantizer’s distortion criterion.

4. Blockwise mean rotation in exploratory factor analysis

A distinct statistical use appears in oblique target rotation for small samples. “Robust oblique Target-rotation for small samples” argues that minimizing single cross-loadings can make target-rotated solutions highly sensitive to sampling error, and therefore replaces single cross-loadings by blockwise mean cross-loadings over salient-loading blocks $8 \times 8$46 in an independent clusters model (Beauducel et al., 2023). Starting from an unrotated loading matrix $8 \times 8$47, the method performs an initial orthogonal Procrustes-type alignment to the target, forms $8 \times 8$48, and computes the weighted block-mean matrix

$8 \times 8$49

It then solves the oblique least-squares problem

$8 \times 8$50

with $8 \times 8$51, normalizes $8 \times 8$52 to $8 \times 8$53, and applies $8 \times 8$54 to the full unrotated loading matrix.

The simulation study spans $8 \times 8$55 factors, $8 \times 8$56 items per factor, $8 \times 8$57, $8 \times 8$58, and $8 \times 8$59 replications per condition. The main reported result is that with small $8 \times 8$60, moderate $8 \times 8$61, and large $8 \times 8$62, mean oblique target rotation greatly reduces the negative bias in estimated inter-factor correlations. The example $8 \times 8$63, $8 \times 8$64, $8 \times 8$65, $8 \times 8$66, $8 \times 8$67 gives mean $8 \times 8$68 for conventional OT and $8 \times 8$69 for OMT. In an empirical example based on IPIP Big-Five markers, with $8 \times 8$70 respondents and $8 \times 8$71 disjoint subsamples of size $8 \times 8$72, mean RMS loadings are OT $8 \times 8$73 $8 \times 8$74 versus OMT $8 \times 8$75 $8 \times 8$76, while mean RMS correlations are OT $8 \times 8$77 $8 \times 8$78 versus OMT $8 \times 8$79 $8 \times 8$80. In this setting, the “rotation” is not geometric rotation of data coordinates but an oblique factor transformation estimated from blockwise aggregates.

5. Cryptographic and key-management meanings

In cryptography, block rotation strategy can refer either to literal bit rotations inside a block cipher or to temporal key rotation intervals for multi-block encryption workloads. “A Block Cipher using Rotation and Logical XOR Operations” defines an 8-round cipher on 64-character blocks, viewed as an $8 \times 8$81 matrix of 8-bit words, where each round applies bytewise circular rotation followed by nearest-neighbour XOR diffusion (Kumar et al., 2012). The round-key matrix $8 \times 8$82 specifies how many bit positions each byte is rotated. Session keys evolve block by block through

$8 \times 8$83

and round subkeys are derived by column shifts. The paper reports a key space of $8 \times 8$84, per-block cost of $8 \times 8$85 RORs plus $8 \times 8$86 XORs, throughput of approximately $8 \times 8$87 Mb/s, and about $8 \times 8$88 per block on a $8 \times 8$89 GHz core.

“Security Boundaries of Quantum Key Reuse” treats block rotation strategy as key refresh under concrete security bounds for CTR, CBC, and ECBC-MAC when QKD keys are combined with classical block ciphers (Chen et al., 25 Dec 2025). The central quantity is the maximum number of files $8 \times 8$90 that can be safely encrypted under one key subject to an adversary-advantage target $8 \times 8$91. For CTR,

$8 \times 8$92

so $8 \times 8$93 is the largest value satisfying

$8 \times 8$94

The paper shows that if the key is rotated uniformly every $8 \times 8$95 files, the increase in security strength satisfies

$8 \times 8$96

For SM4-CTR with $8 \times 8$97, $8 \times 8$98, $8 \times 8$99, $d$00 blocks, and $d$01, the numerical solution is $d$02 files, or approximately $d$03 GB per key. The paper gives $d$04 bits, $d$05, and $d$06. Here rotation is temporal rather than orthogonal, but the block structure of the workload is still decisive.

6. Endpoint rotation, sustainability frontiers, and cross-domain interpretation

“Block-A-Mole” extends the semantics of rotation further, from algebraic transforms to moving-target endpoint schedules in censorship resistance (Maiti, 8 Jun 2026). The defender rotates cloud endpoints across address-domain space, while the censor discovers and blocklists IPs and domains. The decisive quantity is not raw rotation speed but the domain burn rate

$d$07

the ratio between how quickly the censor burns domains and how quickly the defender introduces fresh ones. The paper models the stock of live unblocked domains as a birth–death process on $d$08, derives the stationary empty-pool probability $d$09, and gives the closed-form availability law

$d$10

Its main impossibility result is that when $d$11, no amount of IP-rotation or endpoint redundancy can drive $d$12; the binding constraint is the domain economy, not the IP refresh rate. The simulator reproduces a sharp phase transition at the sustainability frontier $d$13 under adversary profiles representative of the GFW, Russia’s TSPU, and Iran.

Across these otherwise heterogeneous literatures, block rotation strategy consistently mediates a trade-off between local controllability and global effect. In image attacks, shuffling and rotating blocks corrupts attention while preserving an $d$14-bounded optimization loop (Wang et al., 2023). In compression, rotating each $d$15 block aligns local geometry to a fixed transform basis (Guerreiro et al., 2014). In PTQ, blockwise orthogonalization is repeatedly used to smooth outliers without paying the cost or incurring the incompatibilities of global rotation (Sanjeet et al., 29 Jan 2026, Shao et al., 6 Nov 2025, Lin et al., 20 Apr 2026, Xu et al., 19 May 2026). In factor analysis, block means regularize an oblique target transformation (Beauducel et al., 2023). In key management and censorship resistance, rotation becomes a temporal scheduling variable whose benefit is bounded by explicit adversarial advantage or burn-rate laws (Chen et al., 25 Dec 2025, Maiti, 8 Jun 2026). This suggests that the unifying idea of block rotation is not the specific mechanics of a $d$16-D or orthogonal rotation, but the imposition of structured locality to reshape a difficult global objective.