In many scientific and engineering domains, we are forced to observe the world through a "blind" lens. We might measure a signal, but we don't know the exact sensitivity of the sensor that captured it. We might observe the output of a system, but the input that triggered it remains a mystery.
Mathematically, these scenarios often boil down to a product: $y = \alpha \cdot \beta$. If we only see the result $y$, we face a fundamental scaling ambiguity. Is $y$ large because $\alpha$ is large, or because $\beta$ is large? This indeterminacy is a recurring obstacle across fields as diverse as wireless communications, medical imaging, and ecological modeling.
The Geometry of Indeterminacy
The core problem, as framed by Alireza Aghasi, Ali Ahmed, and colleagues in A convex program for bilinear inversion of sparse vectors, is that bilinear interactions are inherently non-identifiable. If you scale one vector up and the other down by the same factor, the observed product remains identical.
This isn't just a theoretical nuisance; it has physical consequences. In ultrasound imaging, Christian Wachinger and Nassir Navab note in Estimation of acoustic impedance from multiple ultrasound images that one cannot determine absolute tissue impedance without a known reference medium to "anchor" the scale. Similarly, in neuroscience, Samuel F. Feng demonstrates that a decrease in a reward signal is mathematically indistinguishable from an increase in internal neural noise (The dynamics of explore–exploit decisions reveal a signal-to-noise mechanism for random exploration).
Lifting into Higher Dimensions
To solve these problems, researchers often use a technique called lifting. Instead of trying to find two vectors $x$ and $w$ that satisfy $y = x w^T$, they transform the problem into finding a single matrix $X = x w^T$.
Henrik Ohlsson and Roy Dong show in Blind Identification via Lifting that this "lifting" turns a non-convex, ambiguous search into a convex optimization problem. While this doesn't magically reveal the absolute scale, it allows for the recovery of the structure of the signals up to a single multiplicative constant. Roy Dong further applies this to systems with piecewise constant inputs, showing that even if we don't know the "volume" of the input, we can perfectly reconstruct its timing and relative changes (Blind Identification of ARX Models with Piecewise Constant Inputs).
Breaking the Symmetry with Constraints
If the math itself is ambiguous, we must introduce "side information" to break the symmetry. This is the logic behind blind calibration.
In Self-Calibration and Biconvex Compressive Sensing, Shuyang Ling and Thomas Strohmer argue that while we cannot distinguish between a scaled signal and an inversely scaled sensor gain, we can often assume the sensor gains belong to a specific subspace. By constraining the possible ways the sensor can fail, we narrow the set of solutions.
A similar approach is taken in ecology and epidemiology by Shahana Ibrahim and Xiao Fu. In Under-Counted Tensor Completion with Neural Incorporation of Attributes, they address the "under-counting" problem: is a low count due to a low population (latent intensity) or poor detection (probability)? By using neural networks to process environmental attributes, they provide the first theoretical proof that these two factors can be decoupled, provided the attributes provide enough "diversity" to the observations.
Message Passing and Statistical Inference
When the systems become too large for standard convex optimization, researchers turn to iterative statistical methods. Christophe Schülke, Florent Krzakala, and Lenka Zdeborová introduced Cal-AMP, a message-passing algorithm that treats signal and gain as joint variables (Blind Calibration in Compressed Sensing using Message Passing Algorithms).
The intuition here is that if we observe multiple different signals through the same uncalibrated sensor, the commonalities across observations eventually reveal the sensor's bias. As Chongwen Huang and colleagues demonstrate in the context of Reconfigurable Intelligent Surfaces (RIS), tensor decompositions can decouple these "cascaded" channels by exploiting the multi-dimensional structure of the data (Channel Estimation for RIS-Empowered Multi-User MISO Wireless Communications).
Why It Matters
Without resolving scaling ambiguity, our models remain "proportional" rather than "absolute." As Di Wang and Jinhui Xu point out, an estimator that is only proportional to the true parameter cannot make precise predictions (Estimating Stochastic Linear Combination of Non-Linear Regressions). Whether we are counting molecules in a cell (Andreas Ruttor and Manfred Opper, Approximate parameter inference in a stochastic reaction-diffusion model) or separating sound sources in a room (Hiroshi Sawada et al., Estimating the number of sources using independent component analysis), the goal is the same: to move past relative shapes and recover the true, physical scale of the world.