The model matches the measurement: has it really identified the material?

A small error can hide an undetermined answer

An AI model receives a test signal and returns two material properties. It reconstructs the signal almost perfectly: did it find the right properties? Not necessarily. If different combinations produce the same observation, the problem precedes the algorithm: data cannot distinguish them. We construct and solve a small inverse problem, then change the measurement design. Reducing residual, making numerical output unique and identifying a physical cause are different achievements.

Abstract. A dimensionless linear model uses synthetic true parameters(0.8,0.2). Measuring their sum admits infinitely many solutions; minimum-norm least squares returns about(0.5,0.5), negligible residual but parameter error near0.4243. A quadratic penalty selects a unique solution without adding information. A nearly redundant second measurement amplifies a0.001 perturbation; a more distinct measurement reduces the shift. These are executed calculations, not material measurements or trained-model results.

From prediction to parameter reconstruction

The forward problem chooses parameters and predicts a signal; the inverse problem observes a signal and reconstructs parameters. Letx=(x₁,x₂) be normalized contributions andy a normalized reading. Modely=x₁+x₂ is deliberately simple, not a validated spectroscopy law; contributions are not asserted chemical fractions. Picture two additive effects. All quantities are dimensionless after assumed normalization; real tests require original scales and units.

We setx*=(0.8,0.2), hencey=1. The star marks truth known only in our synthetic experiment, used to measure reconstruction error. Real inverse problems usually lack that truth. Evaluating only reproduction ofy may miss wrong parameters. A learned model rewarded solely for signal reconstruction faces the same issue: a good signal objective does not guarantee unique identification of hidden effects.

Familyx(t) preserves the sum: increase one contribution and decrease the other equally. Vector(1,−1) is in A’s null space, a direction invisible to the sensor. Nonnegative contributions still permit−0.5≤t≤0.5, so nonnegativity is not uniqueness. Pairs(0.5,0.5),(0.8,0.2),(0.9,0.1) all yield1. This is structural non-identifiability, not solver nonconvergence.

Residual and parameter error answer different questions

Residual||Ax−y||₂ measures signal mismatch; parameter error||x−x*||₂ measures distance from synthetic truth. For(0.5,0.5), mathematical residual is zero but parameter error is√[(-0.3)²+0.3²]≈0.424264. No contradiction: they measure different objects. Numerical residual is about2.22×10⁻¹⁶ from rounding, not physical measurement precision.

Minimum-norm solving selects the symmetric pair because||x(t)||²=0.5+2t² is smallest att=0. It prefers small norm, not knowing synthetic truth(0.8,0.2). A network could learn a different preference from training distribution; no network is trained or evaluated here. The logical point is that output can reflect method preference where measurements cannot distinguish alternatives.

Making a solution unique does not create information

Add a quadratic penalty, the Tikhonov form used here. One term asks for data compatibility, another penalizes large parameters; nonnegativeλ controls tradeoff. Normalization makes terms comparable here; changing original units without rescaling changes preference. λ is neither a physical constant nor a probability, but a method choice requiring documentation and validation.

Subtract conditions:2λ(x₁−x₂)=0, so positiveλ forces equality. Substitution gives(2+λ)x₁=1. Atλ=0.1 both are≈0.476190, predicted sum≈0.952381 and residual≈0.047619. The solution is unique and numerically stable here, but parameter error≈0.425598: truth was not better identified. The penalty selected and shrank a point, not observed the missing direction.

Use||x−x₀||² to expose preference. References(0.5,0.5) and(0.9,0.1) already fit the observation; atλ=0.1 each makes both terms zero and is uniquely selected. Both residuals are negligible, yet parameters differ substantially. This establishes no correct prior, but refutes “unique solution plus good fit implies identified physical properties”.

A second measurement can be almost the first

Addy₂=x₁+(1+ε)x₂. Nonzeroε makesy₂−y₁=εx₂, identifying both parameters in exact arithmetic. But smallε makes the second sensor almost repeat the first; distinguishing information lives in a tiny difference. Clean readings arey₁=1,y₂=1+0.2ε. Add deterministicδ=0.001 only to the second reading, without inventing a noise distribution.

Atε=0.01, cleany₂=1.002 becomes1.003. Reconstruction givesx₂=0.3,x₁=0.7 rather than0.2,0.8. Measurement change0.001 moves each parameter0.1. Residual remains negligible because both perturbed equations can be fit exactly, hiding amplification. Parameter error is≈0.141421. These are normalized dimensionless numbers, not instrument resolutions.

Atε=0.001 the same perturbation gives≈(−0.2,1.2). Interpreting contributions as nonnegative makes this physically inadmissible, diagnosing amplification by unconstrained solving. We did not run positivity-constrained optimization or assign it performance. Constraints may prevent negative signs without proving recovery of(0.8,0.2). Even physically bounded points can differ from truth, as the sum-one family showed.

Change experimental design toε=1, soy₂=x₁+2x₂. The same0.001 perturbation yields≈(0.799,0.201), shifting each parameter0.001, a hundredfold smaller than atε=0.01. Neither noise nor solver improved; the measurement distinguishes contributions more strongly. Experimental design concerns observable directions, not simply number of readings, under comparable model and scaling.

What singular values tell us

Singular value decomposition, SVD, describes how A stretches orthogonal parameter directions. Small singular values mean small signal changes; inversion can amplify perturbations. Atε=0.01, computed singular values≈2.0050125 and0.0049875 give2-norm condition number≈402.0075. Atε=1 it is≈6.8541. This is not a predicted error percentage, but worst-case relative sensitivity under specified norms and scaling.

With zero prior, quadratic regularization replaces inversion factor1/σ byσ/(σ²+λ) in each singular direction. Ifσ²≪λ it attenuates that direction, reducing noise amplification but possibly suppressing real parameter differences. This trades variability against bias rather than obtaining missing information freely. In the one-measurement problem the exactly null direction remains unobserved; prior preference determines its treatment.

Executed results show both sides. Atε=0.001, unregularized error≈1.4142 falls to≈0.4245 withλ=0.01, toward symmetric estimates. Atε=1, where measurements already separate parameters well, the sameλ raises error from≈0.001414 to0.038199. Neither method universally wins. Applications need selection protocol, noise model and independent evidence; these calculated cases only expose the mechanism.

Reproducing calculations without confusing conventions with data

The program uses NumPy2.5.3 andlstsq withrcond=None. Tikhonov augments A withsqrt(λ)I and y withsqrt(λ)x₀, avoiding explicit inversion ofAᵀA. Primary documentation specifies minimum norm when minimizers are multiple and singular-value cutoff behavior. Numerical rank depends on threshold and precision; it is not equivalent to practical identification with noisy data.

Residual is explicitly computed asnorm(A@x−y). An empty solver residuals array is not proof of zero error; documentation specifies emptiness for certain shapes or ranks. The script saves truth, matrices, clean and perturbed observations, singular values and estimates. Noise is deterministic, so no seed applies. Platform rounding differences do not alter the algebraic counterexample, though versions and tolerances remain reproducibility data.

Repeating the first measurement adds proportional rows and does not remove null direction(1,−1). Independent noise could improve sum estimation without separating contributions. This is algebraic reasoning, not an executed repetition experiment. Larger models may incorporate extra knowledge but do not retroactively make a blind measurement informative. Identify added information’s origin: distinct data, prior, physical constraints or training examples.

The opening answer is no: matching a measurement alone does not identify material properties. Non-identifiability admits equivalent explanations; near-redundancy magnifies perturbations; regularization stabilizes through preference, not causal certification. A discriminating measurement may matter more than a new penalty. This exercise verifies that distinction without attributing material tests, proprietary models, discoveries or unperformed experiments to EL-AI.

NumPy primary reference — numpy.linalg.lstsq: minimum-norm solution, residuals and singular-value cutoff; accessed16October2026.

import numpy as np
from experiment import solve, TRUE
for eps in [.001,.01,1.]:
    A=np.array([[1.,1.],[1.,1.+eps]])
    y=A@TRUE+np.array([0.,.001])
    print(eps,solve(A,y)['x'],solve(A,y)['residual_norm'])

Code, data, and instructions · JSON. Educational calculations executed with Python 3.14.0; figures with Matplotlib 3.11.2. AI-assisted analysis, without claiming peer review or human review. Original illustrative ImageGen cover: it does not document EL-AI people, premises, or installations. Sources accessed 16 October 2026.