⚡ electrical engineering

Subspace Overlap Predicts Residual Sensing Feasibility in OFDM ISAC Systems

This paper introduces a subspace overlap metric to predict the feasibility of exploiting MMSE equalization residuals for sensing in OFDM ISAC systems, demonstrating that a proposed K–R approach significantly outperforms baselines when the overlap between residual and target subspaces is sufficiently low.

Original authors: Ramakrishna Pasupuleti

Published 2026-07-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ramakrishna Pasupuleti

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Turning "Garbage" into Gold

Imagine you are trying to listen to a friend (the Target) whispering in a very loud room where a band is playing (the Communication Signal).

In current technology, if you want to hear your friend, you usually need a second, separate microphone just for them. But this paper asks a bold question: Can we just listen to the "leftover" noise after we've tried to understand the music, and find our friend there?

Usually, when a computer tries to clean up a noisy signal to understand the music, it throws away the "mistakes" or "residuals" (the parts it couldn't figure out). This paper says: Don't throw that away! That leftover noise actually contains a hidden map of where your friend is standing.

The Problem: The "Ghost" in the Machine

When the computer tries to clean the signal, it doesn't do a perfect job. It leaves behind "artifacts"—smudges or echoes that look like noise.

The Old Way: Engineers thought these smudges were just random junk.
The New Discovery: The author found that these smudges aren't random. They are structured, like a low-resolution shadow. Sometimes, this shadow hides your friend; sometimes, it reveals them.

The Secret Ingredient: "Subspace Overlap" (The Venn Diagram)

The paper introduces a new way to measure if we can find the friend in the noise. They call it Subspace Overlap (let's call it the "Shadow Match").

Imagine two transparent sheets of paper:

Sheet A: The pattern of the "smudges" left by the computer's cleaning process.
Sheet B: The pattern of your friend's whisper.

Low Overlap (Good): If the patterns on the sheets are totally different (like a circle and a square), they don't mix. The computer can easily separate the friend from the smudges. The paper says if this overlap is very small (below 0.05), we can successfully find the target.
High Overlap (Bad): If the patterns are almost identical (like two circles on top of each other), the friend gets lost in the smudges. If the overlap is too high (above 0.10), the method fails.

Key Insight: It doesn't matter how "simple" the smudge looks; it matters how much it looks like the friend.

The Solution: The "K–R" Receiver

The author built a new tool called the K–R Receiver. Think of it as a smart filter that works in two steps:

Step 1 (The Cleanup): It uses a standard method (MMSE) to remove the loud music. This is like turning down the volume of the band by 20 decibels. Suddenly, the whisper is much clearer relative to the music.
Step 2 (The Magic Filter): It looks at the leftover noise. It uses a mathematical trick (SVD/PCA) to find the "smudges" and subtracts them out, leaving only the friend's whisper.

The Blind Trick:
Usually, to do this, you need to know exactly what the friend's whisper looks like beforehand (an "oracle"). The author created a Blind Algorithm that figures this out automatically. It looks for a sudden drop in the "loudness" of the smudges (an "eigengap") and decides exactly how many layers of noise to peel away. This works without needing to know the target in advance.

The Results: Does it Work?

The author tested this on 12 different types of environments (from open fields to crowded cities, using standard 3GPP models).

The Winner: The new K–R method beat every other method, including a "perfect" method that knows everything about the channel.
Why? Because the first step (turning down the music) was so effective. By removing 20 dB of the loud communication signal, the "friend" became much easier to see in the leftovers.
The Score: In the best cases, this method improved the ability to detect the target by about 2 to 4.5 decibels.
Success Rate: With the new "Blind" algorithm, the method worked in 11 out of 12 environments.

Summary in One Sentence

This paper proves that the "trash" left over after cleaning a communication signal isn't useless; by measuring how much the "trash" overlaps with the target, we can mathematically separate them to detect objects (sensing) without needing extra hardware, effectively turning a standard 6G radio into a radar.

Technical Summary: Subspace Overlap Predicts Residual Sensing Feasibility in OFDM ISAC Systems

Problem Statement
Integrated Sensing and Communication (ISAC) is a foundational capability for 6G, yet current receiver architectures typically rely on parallel sensing chains that require dedicated resources. This paper investigates whether the Minimum Mean Square Error (MMSE) equalization residual—conventionally discarded as noise in communication receivers—can be exploited as a sensing resource. The core challenge is determining whether the artifacts remaining after channel estimation and equalization contain exploitable target structure or if they are merely incoherent noise. Furthermore, the paper addresses the lack of a quantitative metric to predict when such residual sensing is feasible and proposes a method to adaptively remove artifacts without prior knowledge of the target subspace.

Methodology
The study utilizes an OFDM system model operating at 28 GHz (NR FR2) with 256 subcarriers and 64 symbols. The received signal is modeled as the sum of communication and target channels, where the target is 20–40 dB weaker than the communication channel. The proposed K–R Receiver operates in two stages:

K-Step (Standard MMSE): Performs standard channel estimation (LS + interpolation + LMMSE smoothing) and hard decision equalization to generate a residual signal ( $r_s$ ).
R-Step (Residual Processing): Applies Singular Value Decomposition (SVD) to the residual. An adaptive Principal Component Analysis (PCA) removes the dominant artifact components based on rank selection, followed by a 2D-FFT to generate a Range-Doppler map.

The authors analyze 12 distinct channel models, including seven parametric channels and five 3GPP Clustered Delay Line (CDL) profiles. They introduce a new metric, $\eta_1$ (subspace overlap), defined as the normalized Frobenius norm of the projection between the artifact singular vectors and the target singular vectors. This metric quantifies the geometric separability between the residual artifacts and the target signal.

To address the practical limitation of not knowing the target subspace, the paper proposes Algorithm 1, a blind adaptive rank selection method. This algorithm identifies the optimal rank ( $k^*$ ) for artifact removal by detecting the "eigengap" (the ratio $\sigma_k / \sigma_{k+1}$ ) in the observed residual's singular values, eliminating the need for an oracle.

Key Contributions

Universal Low-Rank Structure: The authors demonstrate that residual artifacts across all 12 tested channels are universally low-rank, with the first singular value capturing 28–45% of the total energy. This is attributed to frequency-domain interpolation acting as a low-pass filter on estimation errors.
Subspace Overlap Metric ( $\eta_1$ ): The paper establishes $\eta_1$ as a statistically significant predictor of sensing feasibility. Regression analysis shows that when $\eta_1 < 0.05$ , rank-1 PCA achieves a positive differential; when $\eta_1 > 0.10$ , separation fails. In a multi-variable model including delay spread, target gain, and SNR, $\eta_1$ contributes to an explanatory power of $R^2 = 0.55$ .
Blind Adaptive Rank Selection: The proposed eigengap-based algorithm successfully selects the optimal rank without target knowledge, extending sensing feasibility to 11 out of 12 channels (rescuing High Doppler, Weak NLOS, and LOS scenarios).
Performance Superiority: The K–R PCA-1 approach outperforms all baselines, including matched filtering with perfect Channel State Information (CSI).

Results

Performance Gain: Under UMi conditions (SNR = 20 dB, target at −10 dB), the K–R PCA-1 method achieves a +2.11 dB differential over the baseline. This significantly outperforms matched filtering with perfect CSI (+0.35 dB) and standard 2D-FFT on the residual (−0.47 dB).
Mechanism of Gain: The improvement is attributed to the MMSE equalizer's ability to pre-cancel approximately 20 dB of communication interference. While the raw signal buries the target 20 dB below the communication channel, the residual presents the target with a near-unity target-to-interference ratio, allowing PCA to isolate it effectively.
Statistical Validation: Results were validated over 500 Monte Carlo trials with $p < 0.001$ . Peak performance reached +2.12 ± 0.16 dB.
Complexity: The implementation is computationally efficient. Using power iteration for rank-1 SVD costs 0.08 ms (93% cheaper than full SVD). The total R-step adds 0.23 ms overhead to the communication baseline, resulting in a total receiver latency of 1.08 ms/frame.
Feasibility Limits: Single-frame Constant False Alarm Rate (CFAR) detection remains unachievable ( $H_1/H_0 \approx 1.0$ ). However, multi-frame accumulation (50 frames) bridges this gap, achieving a detection probability ( $P_d$ ) of 0.98 at a false alarm rate ( $P_{fa}$ ) of 0.01.

Significance
The paper claims to establish the specific subspace conditions under which residual-based sensing is feasible without dedicated hardware. The primary insight is that feasibility depends not on the low-rankness of the artifacts (which is universal), but on the geometric separability ( $\eta_1$ ) between the artifact and target subspaces. By providing a quantitative predictor ( $\eta_1$ ) and a blind adaptive algorithm, the work demonstrates that MMSE residuals can serve as a viable sensing resource in OFDM ISAC systems, provided the subspace overlap is sufficiently low. The findings suggest that the MMSE equalizer's interference cancellation capability is the critical enabler for this dual-use functionality.