📊 statistics

Deriving the Variance-Minimizing Design for Standard Addition via c-Optimality

This paper synthesizes existing analytical literature and applies c-optimality theory to demonstrate that for linear responses with non-decreasing measurement errors, the variance-minimizing design for standard addition is a two-point configuration with an optimal allocation ratio that depends on the specific experimental setting, rendering weighted regression unnecessary in this case.

Original authors: Gerhard Gössler, Vera Hofer, Walter Goessler

Published 2026-06-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gerhard Gössler, Vera Hofer, Walter Goessler

Original paper licensed under CC BY 4.0 (http://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

Imagine you are a detective trying to find the exact amount of a hidden substance (let's call it the "mystery ingredient") in a soup. You can't taste the soup directly to know the amount, so you use a trick called Standard Addition.

Here's how the trick works: You take a cup of your soup and add a known amount of the mystery ingredient. Then you take another cup and add a different known amount. You measure the "flavor intensity" (the signal) of each cup. By seeing how the signal changes as you add more ingredient, you can mathematically work backward to figure out how much was in the soup to begin with.

This paper is essentially a guide on how to set up this experiment to get the most accurate answer possible with the least amount of wasted effort.

Here is the breakdown of their findings using simple analogies:

1. The "Two-Point" Rule: Less is More

The authors discovered that the best way to run this experiment is to use only two specific cups (two data points), no matter how many total measurements you plan to take.

Cup A: The original soup with nothing added (the baseline).
Cup B: The soup with one specific amount of the mystery ingredient added.

The Analogy: Imagine you are trying to guess the height of a tree. You could measure it from 10 different distances, but the authors found that measuring it from just two specific spots (one right at the base, and one at a specific distance away) gives you the most precise answer. Adding more measurement spots (like 3, 4, or 5 cups) actually makes the result worse or at best, no better, unless you do some very complex math corrections.

2. The "Sweet Spot" for the Second Cup

Where should you put that second cup (how much ingredient should you add)?

If the "noise" (measurement error) stays the same or grows slowly: You should add as much as you are allowed to. Think of this like stretching a rubber band as far as it can go without breaking; the wider the gap between your two points, the easier it is to calculate the slope accurately.
If the "noise" grows very fast (like a runaway train): You shouldn't add the maximum amount. Instead, there is a "sweet spot" somewhere in the middle. If you go too far, the noise becomes so loud that it drowns out the signal, making your calculation less accurate.

3. The "Split" of Your Effort (50/50 vs. Unbalanced)

You have a limited number of measurements (say, 12 total). How do you split them between Cup A (no addition) and Cup B (with addition)?

The Old Way: Many people assume you should split them evenly (6 for Cup A, 6 for Cup B).
The Paper's Finding: This is often wrong. The best split depends on how "noisy" your measurements are.
- If the noise is constant, a 50/50 split is usually fine.
- If the noise changes as you add more ingredient, you might need to take many more measurements of the baseline (Cup A) and fewer of the spiked one, or vice versa. It's like balancing a scale; if one side is wobbly, you need to weigh it more times to get a stable reading.

4. The "Weighted" Problem

When you have more than two cups (multi-point designs), you usually have to use a complex mathematical tool called "weighted regression" to fix the fact that some measurements are noisier than others. It's like trying to average a group of people's voices where some are whispering and others are shouting; you have to give the whisperers more "weight" in the final average.

The Good News: If you stick to the Two-Point Design (Cup A and Cup B), you don't need this complex weighting. The math works out perfectly on its own. This saves you from having to guess the right weights, which is often a source of error.

5. The "Blind Spot" (Bias)

The paper also checked if this method introduces a "bias" (a systematic error where you are consistently too high or too low).

They found that while the "Two-Point" design is the best for precision (getting the same result every time), it doesn't always minimize bias perfectly. However, the difference is usually small enough that the precision benefits outweigh the tiny bias risks.

Summary of the "Golden Rules" from the Paper:

Stick to two points: Measure the original sample and one spiked sample. Don't bother with 3, 4, or 5 different spiked levels.
Don't assume 50/50: Unless you know your errors are perfectly constant, don't just split your measurements evenly. Calculate the best split based on how the error behaves.
Go wide (mostly): If your errors don't explode as you add more ingredient, push your second measurement to the maximum limit allowed.
Skip the complex math: Because you are only using two points, you don't need to worry about the complicated "weighted regression" math that is required for multi-point designs.

The Bottom Line: To get the most precise answer for the mystery ingredient, stop trying to gather data from everywhere. Focus your energy on two specific, well-chosen points, and you will get a better result with less headache.

Technical Summary: Deriving the Variance-Minimizing Design for Standard Addition via c-Optimality

Problem Statement
The paper addresses the scarcity of accessible literature regarding optimal experimental designs for the standard addition method in analytical chemistry. While mathematical optimality theory exists, its application to standard addition is often obscured by complex literature or scattered across sources. The core problem is determining the experimental design—specifically the number of spiked concentrations ( $n_e$ ), their specific values ( $x_i$ ), and the allocation of a fixed total number of measurements ( $n$ )—that minimizes the variance of the estimated analyte concentration ( $\hat{C}_0$ ). The study considers linear response models subject to measurement errors that may be homoscedastic (constant variance) or heteroscedastic (variance increasing with concentration, e.g., linearly or quadratically).

Methodology
The authors apply the theory of optimal experimental design, specifically c-optimality, to the standard addition problem.

Model Formulation: The relationship between instrumental response ( $Y$ ) and concentration ( $c = C_0 + x$ ) is modeled as linear: $Y = \beta_0 + \beta_1 x + \epsilon(c)$ . The estimator for the unknown concentration is $\hat{C}_0 = \hat{\beta}_0 / \hat{\beta}_1$ .
Variance Approximation: Using error propagation (Taylor series expansion), the variance of $\hat{C}_0$ is approximated as a linear combination of the variances and covariance of the regression parameters ( $\hat{\beta}_0, \hat{\beta}_1$ ). This allows the minimization of $\text{Var}(\hat{C}_0)$ to be framed as a c-optimal design problem.
Weighted Regression: For heteroscedastic cases, the authors utilize Weighted Least Squares (WLS) with weights inversely proportional to the variance function $v(x) = V_0 + (\beta_1(C_0+x))^k$ . This transforms the problem into one with homoscedastic errors, enabling the use of standard optimality tools.
Elfving's Theorem: Instead of solving complex Karush-Kuhn-Tucker (KKT) conditions for every specific variance structure, the authors employ Elfving's Theorem. This geometric approach characterizes the optimal design by finding a supporting line to the "Elfving set" (the convex hull of the regression function vectors $\pm f(x)$ ).
Simulation and Validation: The theoretical derivations are validated through simulations ( $10^4$ iterations) comparing optimal designs against common multi-point designs (equidistant concentrations) and simple 50:50 two-point designs across various variance behaviors ( $k=0, 1, 2, 3, 4, 5$ ) and concentration ranges ( $r$ ).

Key Contributions and Results

Two-Point Optimality: The primary theoretical result is that, for non-decreasing measurement error variances, the variance-optimal design is always a two-point design ( $n_e = 2$ ). One point is invariably the unspiked sample ( $x_1 = 0$ ), and the second is a single spiked concentration ( $x_2 = x^*$ ). This holds regardless of the specific variance behavior (homoscedastic or heteroscedastic).
Optimal Spiked Concentration ( $x^*$ ):
- If the variance grows at most quadratically ( $k \le 2$ ), the optimal spiked concentration $x^*$ is the maximum permissible concentration ( $r$ ).
- If the variance grows faster than quadratically ( $k > 2$ ), the optimal $x^*$ may be strictly less than $r$ . In these cases, extending the range beyond a specific bound ( $x^*_{max}$ ) degrades precision.
Optimal Allocation: The optimal distribution of measurements between the unspiked and spiked samples ( $n_1$ and $n_2$ ) is generally not 50:50. It depends on the variance structure and parameters. However, for the practically important case of quadratic variance growth ( $k=2$ ) with small blank variance, the optimal allocation approaches 50:50.
Weighted Regression Necessity: A significant practical finding is that weighted regression is unnecessary for two-point designs to achieve optimality. While multi-point designs require weighting to handle heteroscedasticity effectively, the two-point design achieves optimal results without it, avoiding the difficulties of weight determination.
Performance Comparison:
- Optimal two-point designs consistently outperform multi-point designs (even weighted ones) in terms of precision.
- A simple 50:50 two-point design (using $x_2=r$ ) performs significantly better than common multi-point designs (e.g., 4 equidistant points) across most settings, particularly when $k \le 2$ .
- For $k > 2$ , blindly increasing the range $r$ in multi-point or 50:50 designs can lead to catastrophic precision loss if $r$ exceeds $x^*_{max}$ .

Significance and Claims
The paper claims to bridge the gap between abstract optimality theory and practical analytical chemistry by providing a rigorous yet accessible derivation of optimal standard addition designs.

Practical Utility: The authors emphasize that while the exact optimal design requires knowledge of unknown parameters ( $\beta_0, \beta_1, k, V_0$ ), the structural insight that a two-point design is superior is robust.
Heuristic Recommendation: In the absence of exact parameter knowledge, the paper suggests that a 50:50 two-point design is a robust, near-optimal strategy that significantly outperforms traditional multi-point approaches, provided the variance does not increase extremely rapidly ( $k \le 2$ ).
Bias vs. Precision: The study notes that while the design minimizing variance is the primary focus, the bias of the estimator is generally moderate and follows similar trends to the variance, though the variance-optimal design is not strictly identical to the bias-optimal design.
Iterative Strategy: For cases where $k > 2$ and $x^*_{max}$ is unknown, the authors propose an iterative approach: perform an initial design to estimate parameters and the variance function, then refine the design in subsequent steps to approach the theoretical optimum.

The work concludes that the complexity of multi-point designs and the necessity of weighting can be avoided in favor of a simpler, theoretically grounded two-point strategy, which offers superior precision and ease of implementation for standard addition analyses.