Dual-LS: Dual Constructions & Applications

• Dual-LS is a multifaceted framework that integrates dual constructions in least squares calculus, notably using dual (co)excisive methods in Goodwillie theory to analyze topological invariants.
• It extends to signal processing and communications by employing adaptive dual least squares algorithms for self-interference cancellation and compressed sensing in massive MIMO systems.
• Dual-LS also underpins scalable machine learning and data systems, optimizing dual-objective tasks via tensor network methods, dual-latent models, and even aiding astrophysical observations.

Dual-LS refers to a constellation of mathematical and algorithmic concepts relating to “dual” constructions involving least squares (LS), the Lusternik–Schnirelmann (LS) category/cocategory, and dual-layer or dual-latent integration. The term has been employed in multiple advanced contexts: Goodwillie/LS calculus in algebraic topology, digital self-interference cancellation in MIMO systems using LS with adaptive order, large-scale dual-form LS-SVM training with tensor networks, dual-objective LS-driven indexes for data systems, dual-latent fusion in generative image models, and “dual” LS sequences in discrepancy theory. This article surveys the most rigorous mathematical and engineering usages of the term, focusing on foundational principles, algorithms, and applications.

1. Dual Constructions in LS Calculus and Goodwillie Theory

A central instance of “dual-LS” arises in the homotopy-theoretic context of Goodwillie calculus, where LS refers to the Lusternik–Schnirelmann cocategory or category, and the “dual” calculus mirrors the standard “excisive” calculus by considering dual (co-excisive) approximations.

For a reduced homotopy endofunctor $F$, the $n$-excisive approximation $P_nF$ is defined by iterating the construction $$T_nF(X) = \operatorname{holim}_{U\in\mathcal{P}_0([n])} F(U\star X)$$, which can be decomposed as $T_nF \simeq R_n F L_n$, where $L_n$ is a diagrammatic left adjoint and $R_n$ is a right adjoint involving homotopy limits. For the identity functor, $T_nI = R_nL_n$ defines a monad whose algebraic structure determines the spaces of symmetric LS cocategory $n$.

Dualizing this, a dual calculus is established with $T^n F = L^n F R^n$ and co-excisive approximations $n$0, capturing dual properties related to LS category. Spaces of LS cocategory $n$1 are retracts of $n$2, while dual constructions characterize spaces of LS category $n$3 via homotopy sections from $n$4 (Eldred, 2012).

The dual calculus also encodes dual topological invariants: while excisive towers force vanishing of iterated Whitehead products (controlling nilpotence), co-excisive towers ensure vanishing of cup products above a certain length. This duality threads through the arithmetic of classical and newer nilpotency concepts in homotopy theory.

2. Dual-LS Algorithms in Signal Processing and Communications

In signal processing, “dual-LS” most often designates dual usage or cascades of least-squares (LS) algorithms, or LS with dual adaptation properties.

One example is the use of LS algorithms with adaptive order for digitally-assisted photonic analog self-interference cancellation (SIC) in in-band full-duplex MIMO systems. Here, the LS algorithm estimates the multipath SI (self-interference) channel digitally, reconstructing a reference signal adapted in order (impulse response length) as wireless channel conditions vary. The adaptive order, $n$5, is updated according to error feedback via formulas such as

$n$6

$n$7

ensuring robust channel tracking and complexity control. The reconstructed reference is injected into a dual-parallel Mach–Zehnder modulator for analog cancellation (Han et al., 2022). Experimentally, SIC depths up to $n$8 dB (for 10 GHz, 0.1 Gbaud signals) are reported, and the LS algorithm dynamically converges to the theoretically optimal order under different multipath scenarios.

3. Dual-LS and Large-Scale Support Vector Machines

In kernel methods, the dual form of LS-SVMs is computationally attractive but prohibitive at scale due to the size of the kernel matrix. Tensor Network Kalman Filtering (TNKF) offers a “dual-LS” scheme for solving the dual problem efficiently by combining row-wise Bayesian updates with tensor train (TT) representations.

The dual problem

$n$9

is solved recursively via a state-space model: $P_nF$0 with all variables stored and updated in TT format to enable scaling to $P_nF$1 data points. TNKF outperforms Nyström and fixed-size LS-SVM methods, particularly when the kernel spectrum decays slowly, and provides model confidence via the TT-represented covariance matrix (Lucassen et al., 2021).

4. Dual-LS in Compressed Sensing for Massive MIMO

In dual-wideband (frequency-and spatial-wideband) massive MIMO-OFDM at sub-THz, compressed channel estimation can be structured as a dual-stage (dual-LS) process: first, an LS estimation using prior beam support exploits temporal correlation; second, compressed sensing (CS; L1-minimization) is applied to the LS residual to detect newly emerging sparse components. In the MMV-LS-CS framework:

• LS step: $P_nF$2
• CS step: $P_nF$3

Two algorithms—TS (Two-Stage) and M-FISTA (joint MMV FISTA)—implement this, with channel refinement and hierarchical codebook search enhancing path estimation resolution and efficiency. This dual-stage approach markedly improves NMSE and spectral efficiency relative to one-shot or unstructured estimators (Chou et al., 2022).

5. Dual-LS in Quasi-Monte Carlo and Uniform Distribution

LS-sequences, a family of low-discrepancy sequences introduced by Carbone, are constructed via interval splitting governed by integer parameters $P_nF$4. While LS sequences achieve low discrepancy in one dimension when $P_nF$5 (making them suitable for Quasi–Monte Carlo integration), naively combining two 1D LS sequences coordinatewise does not always result in uniform distribution in higher dimensions. Specifically, if the interval-splitting base parameters $P_nF$6 and $P_nF$7 for dimensions 1 and 2 satisfy

$P_nF$8

or $P_nF$9, the resulting two-dimensional sequence is not even dense in $$T_nF(X) = \operatorname{holim}_{U\in\mathcal{P}_0([n])} F(U\star X)$$0 (Aistleitner et al., 2012). The “dual” caution here is that multidimensional pairing of LS‐sequences is highly sensitive to arithmetic constraints, closely paralleling the coprimality requirements for Halton sequences.

6. Dual-Objective LS in Data Systems and Generative Models

Dual-LS also designates architectures that use LS (or related) mechanisms to optimize two objectives simultaneously.

In DobLIX, a dual-objective learned index for LSM trees, a piecewise linear or regression model is simultaneously trained to optimize both index lookup error (via LS fit) and data block access cost, with block partitioning determined by both prediction error bound $$T_nF(X) = \operatorname{holim}_{U\in\mathcal{P}_0([n])} F(U\star X)$$1 and block size $$T_nF(X) = \operatorname{holim}_{U\in\mathcal{P}_0([n])} F(U\star X)$$2. The index is dynamically tuned during operation via a Q-learning RL agent, ensuring optimal performance across varying workloads (Heidari et al., 7 Feb 2025).

For high-fidelity image synthesis, the DLSF framework (“Dual-Layer Synergistic Fusion”) employs a dual-latent architecture: a base latent encodes global image structure and a refined latent encodes local details. These latents are fused via adaptive global fusion (AGF: softmax-weighted combination after concatenation and convolution) or dynamic spatial fusion (DSF: spatially adaptive sigmoid-weighted mixing), ensuring joint preservation of semantics and textures. Channel and spatial attention for fusion are mathematically formalized, and empirical evaluation demonstrates clear improvements over single-latent (SDXL) models (Chen et al., 16 Jul 2025).

7. Dual-LS in Astrophysical Observations

A further use-case is in astrophysics, where “dual quasars” refer to physically distinct binary quasars (not gravitational lens images) that have nearly identical redshifts and projected separations of $$T_nF(X) = \operatorname{holim}_{U\in\mathcal{P}_0([n])} F(U\star X)$$3–$$T_nF(X) = \operatorname{holim}_{U\in\mathcal{P}_0([n])} F(U\star X)$$4 kpc, determined via angular separation and the angular diameter distance $$T_nF(X) = \operatorname{holim}_{U\in\mathcal{P}_0([n])} F(U\star X)$$5. Spectroscopic analysis involving LS-based extraction and matching disentangles dual quasars from lensed systems, with key discriminants being velocity offsets $$T_nF(X) = \operatorname{holim}_{U\in\mathcal{P}_0([n])} F(U\star X)$$6 and subtle differences in spectral features (He et al., 15 Jan 2025).

Conclusion

“Dual-LS” is a multi-faceted term with rigorous definitions and impactful applications across topology, statistical learning, signal processing, communications, discrepancy theory, key-value data systems, generative models, and observational astrophysics. Whether referring to dual constructions in algebraic calculus, the dual application or adaptation of LS-based algorithms, or architectures optimizing dual objectives using least-squares paradigms, the unifying thread is the systematic integration of duality principles—category/cocategory, primal/dual tasks, or global/local structures—with LS-based mathematical frameworks. Each context demonstrates that such dualization yields new invariants, more expressive models, or enhanced empirical performance, and often reveals deeper structural relationships within the underlying mathematics or system architecture.