Orthogonal Subspace Wake-up (OSW)

• Orthogonal Subspace Wake-up (OSW) is a method that isolates and exploits orthogonal subspaces to support distinct functions in both signal processing and continual learning.
• It utilizes DFT/I-DFT transformations and an alternating minimization algorithm to design wake-up sequences with flat energy distribution and minimal leakage.
• In continual learning, OSW projects new gradients onto the orthogonal complement of historical subspaces, preserving fragile legacy knowledge while enabling new updates.

Orthogonal Subspace Wake-up (OSW) refers to the principled identification and exploitation of orthogonal subspaces to allow the coexistence of distinct functional capabilities within composite systems—either in signal processing (wake-up radios and OFDM), or in continual learning for LLMs. In both domains, OSW provides mathematically grounded safety guarantees against destructive interference: it isolates critical subspaces for legacy operations and rigorously constrains new updates or coexisting signals to their orthogonal complement. This approach overcomes fundamental limitations of conventional multitask or replay-based methods, especially under challenging conditions such as Rayleigh fading channels or fragile knowledge structures.

1. OSW in Frequency-Domain Multiplexing for Wake-up Radios

In wireless communications, OSW realizes strict orthogonality between low-power Wake-up Radio (WUR) signals and high-throughput OFDM data. The approach reserves a contiguous block of $L<N$ subcarriers for wake-up signaling: these tones constitute a dedicated “wake-up subspace” orthogonal to the OFDM QAM data, which occupies the $N-L$ complementary tones. The transmitter constructs frequency-domain OOK sequences $s^{(0)}$, $s^{(1)} \in \mathbb{C}^L$ (for logical ‘0’, ‘1’), supported only on the reserved tones, and maps them into the total spectrum $S^{(i)}$ (zero outside the wake-up subspace). The combined spectrum is $S_{\text{tot}} = d + S^{(i)}$, ensuring complete orthogonality: the OFDM receiver remains unaffected by the wake-up energy, enabling simultaneous, interference-free signaling (Sahin et al., 2018).

2. Optimization of Wake-up Sequences and Leakage Control

The OSW sequence design process aims for three objectives: (a) flat ON-interval energy (constant $|x[n]|^2 = E_s$), (b) minimal OFF-interval leakage ($|x[n]|^2 \approx 0$), and (c) uniform frequency power profile ($|s_k|^2$) to maximize channel diversity. The optimization is formulated as:

$$J(s, \phi) = \|F^H M s - \sqrt{E_s} \phi\|_2 + \alpha \|s\|_\infty$$

subject to (i) forced zero on DC tone, (ii) time-domain leakage bounded by $\delta$ in the OFF interval, and (iii) unimodular phase constraint $|\phi_n|=1$ in ON. The cost term $\alpha$ regularizes spectral peakiness. Constraints maintain practical orthogonality and robustness in fading channels.

3. Sequence Construction via Alternating Minimization (Extended CAN)

The solution employs an alternating minimization procedure (extended Cyclic Algorithm-New, or SCAN). In each iteration, the algorithm alternates: (A) phase update (closed-form over ON interval), and (B) sequence update (convex quadratic program for $s$, subject to leakage and spectral constraints). Both steps guarantee non-increasing cost $J$ and local convergence. The method supports Manchester coding by designing time-reversed or conjugate-mirrored pairs $s^{(0)}, s^{(1)}$. Sequence renormalization enforces constant energy.

4. OSW in Continual Learning for LLMs

In continual learning, OSW addresses the instability-plasticity trade-off: Experience Replay (ER) mixes gradients from old and new tasks, consolidating robust features but catastrophically degrading fragile, structured knowledge (e.g., code generation, formal logic). OSW’s innovation is a two-phase “subspace wake-up”: first, anchor examples from old tasks probe the model for $T_{\text{wake}}$ steps, accumulating LoRA gradients $g_i$ and constructing the historical subspace $S_{\text{hist}}$ by singular value decomposition. Second, new-task gradients are projected onto the orthogonal complement:

$$g_{\text{safe}} = (I - U_{\text{hist}} U_{\text{hist}}^\top) g_{\text{raw}}$$

Only this projected gradient updates the parameters, mathematically guaranteeing zero first-order interference with legacy tasks. All optimization proceeds in the low-rank LoRA space, maintaining tractability for billion-parameter models (Meng, 26 Jan 2026).

5. Theoretical Guarantees and Empirical Evidence

OSW’s safety guarantee states: if $U_{\text{hist}}$ exactly spans all directions decreasing past-task loss, then future updates orthogonal to $S_{\text{hist}}$ leave old capability unchanged to first order. Empirical evaluation demonstrates OSW’s superiority in preserving code accuracy (Py150, Acc $=10.57\%$ after four tasks), versus ER which degrades code structure ($8.37\%$). OSW maintains high plasticity for new reasoning tasks (ScienceQA Acc $=73.0\%$), yet does not consolidate robust NLP tasks as strongly as ER (see Table below).

Task	Seq	ER	OSW
C-Stance Acc	63.0	76.0	61.0
MeetingBank ROUGE-L	29.68	33.83	26.99
Py150 Acc	10.50	8.37	10.57
ScienceQA Acc	73.0	67.0	73.0

OSW uniquely preserves fragile code capability, confirming the tradeoff between broad consolidation (ER) and structural safety (OSW).

6. Implementation and Practical Insights

For wake-up radios, OSW requires only a single DFT/I-DFT block at the transmitter; the receiver operates with a simple energy detector and low-pass sampling. Manchester-encoding exploits time-reversal and shifting of a base sequence. For continual learning, OSW mandates a brief wake-up phase with a small anchor set, and tuning of wake-up length ($T_{\text{wake}}$) and subspace rank ($r$) to balance stability with plasticity. The approach admits scalable deployment in low-dim LoRA space and provides a mechanism for “do no harm” to structured capabilities during sequential domain adaptation.

7. Trade-offs, Extensions, and Generalization

A shorter ON-period in signaling yields peaky-time energy (higher leakage unless $\alpha$ increases); smoother time-domain waveforms require longer ON or nonzero $\delta$. Increasing $\alpha$ flattens the spectral profile, enhancing fading robustness but slightly raising OFF-leakage. OSW extends to multiple simultaneous wake-up streams (partitioning orthogonal subspaces) and to multiplexing >2 OOK symbols per OFDM symbol through frequency–time shift properties. In continual learning, OSW is especially advantageous where structured domains coexist with unstructured ones. A plausible implication is that as model architectures scale or tasks diversify, orthogonal subspace approaches like OSW will become increasingly central to ensuring reliable knowledge retention and robust multitask signal coexistence.