Epistemological Perspectives on Population Parameters | 知识论视角下的总体参数

📚 Epistemological Perspectives on Population Parameters | 知识论视角下的总体参数

In the study of statistics, population parameters such as the mean μ, variance σ², or proportion p are often presented as fixed, unknown numbers that we estimate from samples. But what exactly is a parameter? Is it a real feature of the world, or a useful fiction we construct? This article explores population parameters through the lens of epistemology, the branch of philosophy concerned with knowledge, truth, and belief.

在统计学学习中，总体参数如均值 μ、方差 σ² 或比例 p 常常被呈现为固定但未知的数，我们通过样本来估计它们。但参数究竟是什么？它是世界的真实特征，还是我们构造出来的有用虚构？本文从知识论——研究知识、真理与信念的哲学分支——出发，探讨总体参数的本质。

1. What Is a Population Parameter? | 什么是总体参数？

A population parameter is a numerical summary of a population, such as its mean, median, standard deviation, or correlation coefficient. In mathematical notation, we commonly use Greek letters to denote parameters: μ for the population mean, σ for the population standard deviation, and ρ for the population correlation. These values are treated as constants for a given population at a given time, but they are typically inaccessible to direct observation because the population is infinite or too large to measure fully.

总体参数是描述总体特征的数值概括，例如总体均值、中位数、标准差或相关系数。在数学表示中，我们通常用希腊字母表示参数：μ 表示总体均值，σ 表示总体标准差，ρ 表示总体相关系数。这些值被视为特定时间特定总体下的常数，但由于总体是无限的或过于庞大而无法完全测量，它们通常无法直接观测。

For example, the mean height of all human beings alive today is a population parameter. It exists, but we cannot measure every person. Instead, we sample and infer. This inference is not merely a mathematical exercise; it raises deep questions about how we can claim to know something we have not directly observed.

例如，今天所有活着的人类的平均身高是一个总体参数。它确实存在，但我们无法测量每个人。相反，我们通过样本进行推断。这种推断不仅仅是数学练习；它引发了我们如何声称知道未曾直接观测的事物的深刻问题。

2. Parameters as Idealisations | 参数作为理想化概念

Statisticians often define parameters as the limit of a sample statistic as the sample size approaches infinity. For instance, the population mean μ is defined as the expected value of a random variable X under a probabilistic model. This definition is mathematical, not empirical. It relies on the idea of a hypothetical infinite sequence of observations. In this sense, parameters are not things we find; they are constructs that make our models coherent.

统计学家通常将参数定义为当样本量趋于无穷时样本统计量的极限。例如，总体均值 μ 是随机变量 X 在概率模型下的期望值。这个定义是数学的，而非经验的。它依赖于一个假设的无限观测序列。在这个意义上，参数不是我们发现的事物；而是使我们的模型自洽的构造。

Consider the simple model X ~ N(μ, σ²). Here, μ and σ² are parameters that index a family of distributions. They are not directly observable; they are abstract entities that give the model its shape. Epistemologically, this raises a question: do we know μ because it is in the world, or because we put it there through our modelling choices?

考虑简单模型 X ~ N(μ, σ²)。这里 μ 和 σ² 是索引一族分布的参数。它们不可直接观测；它们是赋予模型形状的抽象实体。从知识论来看，这引发一个问题：我们是因为 μ 在世界中而知道它，还是因为我们通过建模选择将它放在那里才认识它？

Realist view: parameters exist independently of observers.
Instrumentalist view: parameters are useful tools for predictions, not claims about reality.

实在论观点：参数独立于观察者而存在。
工具主义观点：参数是用于预测的有用工具，而非关于实在的断言。

3. The Problem of Induction | 归纳问题

When we estimate a population parameter from a sample, we perform induction: we reason from particular observations to general conclusions. David Hume famously argued that induction cannot be rationally justified, because it assumes that the future will resemble the past. In statistics, this appears as the assumption that a sample is representative of the population. How do we know the sample is representative? We cannot know with certainty; we can only assign probabilities.

当我们从样本估计总体参数时，我们进行了归纳：从特定观测推理到一般性结论。大卫·休谟曾著名地论证归纳无法被理性辩护，因为它假设未来会与过去相似。在统计学中，这表现为样本能够代表总体的假设。我们如何知道样本具有代表性？我们无法确定地知道；我们只能赋予概率。

This is not a practical problem but a philosophical one. In practice, random sampling helps, but randomness itself is a model. Even a perfectly random sample from a finite population cannot guarantee that the sample estimate is exactly equal to the parameter. The gap between sample statistic and population parameter is permanent and irreducible.

这不是实践问题，而是哲学问题。在实践中，随机抽样有帮助，但随机性本身是一种模型。即使从有限总体中完美随机抽样，也无法保证样本估计恰好等于参数。样本统计量与总体参数之间的鸿沟是永久且不可约的。

P(μ̂ = μ) = 0 for a continuous parameter, even with a large but finite sample.

对于连续参数，即使样本很大但有限，P(μ̂ = μ) = 0。

4. Frequentist vs Bayesian Epistemology | 频率学派与贝叶斯学派的知识论

Two major schools of statistical inference offer different epistemological stances toward parameters. The frequentist school treats parameters as fixed constants. Probability statements refer to the long-run frequency of procedures, not to the parameter itself. A 95% confidence interval, for instance, means that if we repeated the sampling procedure many times, 95% of the resulting intervals would contain the true parameter. For a given interval, the parameter either lies inside or outside; there is no probability attached to that specific statement.

两大统计推断学派对参数提供不同的知识论立场。频率学派将参数视为固定常数。概率陈述指的是程序在长期运行中的频率，而非参数本身。例如，95% 置信区间意味着如果我们多次重复抽样过程，95% 的区间将包含真实参数。对于给定区间，参数要么在其中，要么不在；该特定陈述没有附上概率。

The Bayesian school, in contrast, treats parameters as random variables. Probability represents subjective degrees of belief, and we update our belief about a parameter using Bayes’ theorem after observing data. The prior distribution encodes our uncertainty before seeing data, and the posterior distribution encodes it after data. For a Bayesian, a 95% credible interval means that there is a 95% probability that the parameter lies in that interval given the data and the prior.

贝叶斯学派则相反，将参数视为随机变量。概率代表主观信念程度，我们在观测数据后使用贝叶斯定理更新对参数的信念。先验分布编码了在观测数据之前的不确定性，后验分布编码了在数据之后的不确定性。对贝叶斯主义者而言，95% 可信区间意味着在给定数据和先验的情况下，参数落在该区间内的概率为 95%。

Feature	Frequentist	Bayesian
Parameter status	Fixed unknown constant	Random variable
Probability meaning	Long-run frequency	Degree of belief
Key tool	Confidence interval	Credible interval

特征	频率学派	贝叶斯学派
参数地位	固定未知常数	随机变量
概率含义	长期频率	信念程度
关键工具	置信区间	可信区间

5. Confidence Intervals: Epistemic Ambiguity | 置信区间：知识论上的模糊性

Confidence intervals are a staple of IB mathematics, yet their interpretation is notoriously counterintuitive. Students often want to say, “There is a 95% chance that the true mean lies in this interval.” The frequentist statistician must correct them: “The interval was produced by a procedure that succeeds 95% of the time.” The parameter is not random; the interval is random before sampling, and fixed after sampling.

置信区间是 IB 数学的核心内容，但其解释却以违背直觉而闻名。学生总想说：”真实均值落在这个区间内有 95% 的可能性。”频率学派统计学家必须纠正他们：”这个区间是由一个在 95% 情况下成功的过程产生的。”参数不是随机的；区间在抽样前是随机的，抽样后是固定的。

This distinction has deep epistemological consequences. It shows that a confidence interval is not a direct statement about the parameter, but about the reliability of a method. It is a form of procedural knowledge, not propositional knowledge about the world. When we say we are “confident”, we are expressing trust in a method, not a degree of certainty about a number.

这种区分具有深刻的知识论后果。它表明置信区间不是关于参数的直接陈述，而是关于方法可靠性的陈述。这是一种程序性知识，而非关于世界的命题性知识。当我们说我们”有信心”时，我们表达的是对方法的信任，而非对某个数字的确定程度。

A 95% confidence interval: P(interval contains θ) = 0.95 before sampling.

95% 置信区间：抽样前 P(区间包含 θ) = 0.95。

6. Prior Knowledge and Subjectivity | 先验知识与主观性

In Bayesian inference, the prior distribution is a formal way to incorporate existing knowledge about a parameter. But where does this prior come from? It can come from previous studies, expert opinion, or, in the absence of any information, a non-informative prior such as a uniform distribution. The choice of prior can dramatically affect the posterior, especially with small samples.

在贝叶斯推断中，先验分布是将关于参数的已有知识进行形式化的方法。但这个先验从何而来？它可以来自先前研究、专家意见，或者在没有任何信息的情况下，使用均匀分布等无信息先验。先验的选择会极大影响后验，尤其是在小样本情况下。

This introduces an inevitable element of subjectivity. Is this a weakness? Not necessarily. The Bayesian framework makes its assumptions explicit, whereas frequentist methods often hide assumptions about randomness and representativeness. In knowledge theory, making one’s assumptions transparent can be seen as an epistemic virtue, even if it means acknowledging that different analysts may reach different conclusions from the same data.

这引入了不可避免的主观性因素。这是弱点吗？未必。贝叶斯框架使其假设显式化，而频率学派方法通常隐藏关于随机性和代表性的假设。在知识理论中，使假设透明可被视为一种知识论美德，即使这意味着承认不同的分析者可能从相同数据中得出不同结论。

7. Parameters in Machine Learning: Knowledge as Representation | 机器学习中的参数：作为表征的知识

In modern machine learning, parameters appear in models such as linear regression y = wx + b or neural networks. These parameters are learned from data through optimisation algorithms. The word “learning” suggests that the model acquires knowledge about the world. But do we “know” the true parameters? Often, the parameters are not interpretable at all; they are simply coefficients that make predictions accurate.

在现代机器学习中，参数出现在线性回归 y = wx + b 或神经网络等模型中。这些参数通过优化算法从数据中学习。所谓”学习”意味着模型获得了关于世界的知识。但我们是否”知道”真实参数？通常，参数完全不可解释；它们只是使预测准确的系数。

This challenges a simplistic view of parameters as representations of stable truths. A machine learning model may have millions of parameters that do not correspond to any meaningful population attribute. They are functional, not referential. This aligns with instrumentalism: parameters are valuable because they help us predict, not because they mirror reality.

这挑战了将参数视为稳定真理表征的简单观点。机器学习模型可能有数百万个参数，这些参数并不对应任何有意义的总体属性。它们是功能性的，而非指称性的。这与工具主义一致：参数有价值是因为它们帮助我们预测，而非因为它们反映现实。

8. Decision-Making Under Uncertainty | 不确定性下的决策

Ultimately, the purpose of estimating population parameters is to make decisions. Should a pharmaceutical company release a drug? Should a government impose a lockdown? These decisions rely on estimates of parameters such as the true effectiveness of a treatment or the true infection fatality rate. The epistemological status of the parameter matters less than the consequences of acting on our estimates.

归根结底，估计总体参数的目的是为了做决策。制药公司是否应该发布一种药物？政府是否应该实施封锁？这些决策依赖于对真实治疗效果或真实感染病死率等参数的估计。参数的知识论地位不如根据估计采取行动的后果重要。

Here we encounter a classic tension: the desire for certainty versus the reality of uncertainty. In formal decision theory, we assign utilities to outcomes and choose the action that maximises expected utility. The expected utility depends on our posterior belief about the parameter, which already incorporates uncertainty. Thus, even if we cannot know the parameter with certainty, we can still act rationally.

在这里，我们遇到一个经典张力：对确定性的渴望与不确定性的现实。在形式化决策理论中，我们为结果赋予效用，并选择最大化期望效用的行动。期望效用取决于我们对参数的后验信念，这一信念已经包含了不确定性。因此，即使我们无法确定地知道参数，我们仍然可以理性行动。

Choose action a that maximises E[U(a, θ) | data].

选择使 E[U(a, θ) | data] 最大化的行动 a。

9. Conclusion: Knowing Within Limits | 结论：在限制中认识

Population parameters are central to statistical inference, but their epistemological status is far from simple. They may be real quantities, useful fictions, or flexible representations depending on the framework we adopt. The key lesson for IB mathematics students is that statistical knowledge is not absolute. It is conditional on models, assumptions, and sampling procedures. This does not make it useless; rather, it demands a humble and critical attitude toward numbers.

总体参数是统计推断的核心，但其知识论地位远非简单。它们可能是真实的量、有用的虚构，或者灵活的呈现，取决于我们采用的框架。对 IB 数学学生而言，关键教训是统计知识不是绝对的。它依赖于模型、假设和抽样程序。这并不会使它无用；相反，它要求我们对数字采取谦逊和批判的态度。

As we move from data to knowledge, we must remember that every estimate carries a shadow of ignorance. The discipline of statistics does not eliminate uncertainty; it measures it. And in that measurement lies both the power and the limit of human knowledge.

当我们从数据走向知识时，我们必须记住，每一个估计都带着无知的阴影。统计学科并不能消除不确定性；它只是测量不确定性。而正是在这种测量中，蕴藏着人类知识的力量与限度。

Published by TutorHao | Mathematics Revision Series | aleveler.com

更多咨询请联系16621398022（同微信）

Epistemological Perspectives on Population Parameters | 知识论视角下的总体参数