What is Standard Deviation?

The Story That Explains Standard Deviationवह कहानी जो Standard Deviation समझाती है一个讲清标准差的故事

You are a quality control manager at a biscuit factory. Every biscuit should weigh exactly 20 grams. At the end of the day, two machines have both made biscuits with a mean weight of 20 g. Your boss is happy. But you check the individual weights.

आप एक biscuit factory में quality control manager हैं। हर biscuit का वज़न ठीक 20 grams होना चाहिए। दिन के अंत में दोनों machines के biscuits का mean वज़न 20 g है। आपके boss ख़ुश हैं। पर आप एक-एक biscuit का वज़न जाँचते हैं।

你是一家饼干厂的质检经理。每块饼干都应该正好重 20 克。一天结束时，两台机器生产的饼干平均重量都是 20 g。 老板很满意。但你去查看了每块饼干的重量。

Machine A: mean = 20 g, standard deviation = 0.16 g — nearly perfect every time.
Machine B: mean = 20 g, standard deviation = 4.47 g — all over the place.

Machine A: mean = 20 g, standard deviation = 0.16 g — हर बार लगभग perfect.
Machine B: mean = 20 g, standard deviation = 4.47 g — पूरी तरह बिखरा हुआ।

机器 A：均值 = 20 g，标准差 = 0.16 g —— 几乎每次都完美。
机器 B：均值 = 20 g，标准差 = 4.47 g —— 忽大忽小。

What Standard Deviation Actually MeansStandard Deviation का असली मतलब标准差到底是什么意思

Standard deviation answers one simple question: "How far are the values from the mean, typically?" Small SD = values bunch together. Large SD = values spread out.

Standard deviation एक आसान सवाल का जवाब देता है: "Values आमतौर पर mean से कितनी दूर हैं?" छोटा SD = values पास-पास। बड़ा SD = values फैली हुई।

标准差回答一个简单的问题：“数值通常离均值有多远？”标准差小 = 数值挤在一起；标准差大 = 数值很分散。

Step-by-Step CalculationStep-by-Step गणना分步计算

Dataset — exam scores of 8 students: [52, 74, 68, 90, 61, 85, 72, 78]

Dataset — 8 students के exam scores: [52, 74, 68, 90, 61, 85, 72, 78]

数据集 —— 8 名学生的考试分数：[52, 74, 68, 90, 61, 85, 72, 78]

Python ImplementationPython ImplementationPython 实现

Manual calculationहाथ से गणना手动计算

scores = [52, 74, 68, 90, 61, 85, 72, 78]

n    = len(scores)
mean = sum(scores) / n

squared_diffs = [(x - mean) ** 2 for x in scores]
variance      = sum(squared_diffs) / (n - 1)
std_dev       = variance ** 0.5

print(f"Mean:     {mean}")
print(f"Variance: {variance:.2f}")
print(f"Std Dev:  {std_dev:.2f}")

Using the statistics modulestatistics module से使用 statistics 模块

import statistics

scores = [52, 74, 68, 90, 61, 85, 72, 78]

print(statistics.mean(scores))                 # mean
print(round(statistics.variance(scores), 2))   # sample variance (n - 1)
print(round(statistics.stdev(scores), 2))      # sample std dev
print(round(statistics.pstdev(scores), 2))     # population std dev (n)

Using NumPyNumPy से使用 NumPy

import numpy as np

scores = [52, 74, 68, 90, 61, 85, 72, 78]

print(np.mean(scores))
print(np.std(scores, ddof=1).round(2))   # sample std dev
print(np.std(scores).round(2))           # population std dev — NumPy's default!
print(np.var(scores, ddof=1).round(2))   # sample variance

Complete comparison — the two machinesपूरी तुलना — दोनों machines完整对比 —— 两台机器

import numpy as np

machine_a = [19.8, 20.1, 20.0, 19.9, 20.2]
machine_b = [14.0, 26.0, 18.0, 22.0, 20.0]

for name, data in [("Machine A", machine_a), ("Machine B", machine_b)]:
    mean     = np.mean(data)
    variance = np.var(data, ddof=1)
    std_dev  = np.std(data, ddof=1)
    print(f"{name}: mean={mean:.1f} g  variance={variance:.3f} g²  std_dev={std_dev:.2f} g")

The 68 – 95 – 99.7 Rule68 – 95 – 99.7 नियम68 – 95 – 99.7 法则

When data follows a normal distribution (bell curve), the standard deviation tells you what share of values falls in any range. This is the Empirical Rule.

जब data normal distribution (bell curve) पर चलता है, तो standard deviation बताता है कि किसी भी range में कितनी values आती हैं। इसे Empirical Rule कहते हैं।

当数据服从正态分布（钟形曲线）时，标准差能告诉你任意区间内有多少比例的数值。这就是经验法则（Empirical Rule）。

Real Example — Adult Heightsअसली उदाहरण — वयस्कों की ऊँचाई真实例子 — 成年人身高

Take adult heights that follow a normal distribution with mean = 175 cm and standard deviation = 7 cm (a common textbook approximation for adult men).

वयस्कों की ऊँचाई लें जो mean = 175 cm और standard deviation = 7 cm वाले normal distribution पर चलती है (वयस्क पुरुषों के लिए एक आम textbook अनुमान)।

假设成年人身高服从均值 = 175 cm、标准差 = 7 cm 的正态分布（这是教科书中对成年男性常用的近似）。

from scipy.stats import norm

mean, std_dev = 175, 7   # cm

for k in (1, 2, 3):
    lo, hi = mean - k * std_dev, mean + k * std_dev
    share = norm.cdf(hi, mean, std_dev) - norm.cdf(lo, mean, std_dev)
    print(f"±{k}σ: {lo} – {hi} cm  →  {share:.2%} of men")

# Any other range works too, e.g. taller than 190 cm:
print(f"Taller than 190 cm: {norm.sf(190, mean, std_dev):.2%}")

Standard Deviation in Machine LearningMachine Learning में Standard Deviation机器学习中的标准差

Standard deviation is not just a statistics idea — it powers several machine-learning tools you will use every day.

Standard deviation सिर्फ़ statistics का विचार नहीं — यह कई machine-learning औज़ारों को चलाता है जो आप रोज़ इस्तेमाल करेंगे।

标准差不只是统计学概念 —— 它支撑着你每天都会用到的多种机器学习工具。

Z-scores and StandardScalerZ-scores और StandardScalerZ 分数与 StandardScaler

Many algorithms (SVM, KNN, neural networks, PCA) are sensitive to the scale of features. StandardScaler turns each feature into z-scores: mean 0 and standard deviation 1. A z-score says "how many SDs away from the mean" — so values in different units become comparable.

कई algorithms (SVM, KNN, neural networks, PCA) features के scale से प्रभावित होते हैं। StandardScaler हर feature को z-scores में बदलता है: mean 0 और standard deviation 1. Z-score बताता है "mean से कितने SD दूर" — इसलिए अलग units की values की तुलना हो सकती है।

很多算法（SVM、KNN、神经网络、PCA）对特征的尺度很敏感。StandardScaler 把每个特征转换成 z 分数：均值 0、标准差 1。z 分数表示“离均值有几个标准差” —— 于是不同单位的数值也能放在一起比较。

import numpy as np
from sklearn.preprocessing import StandardScaler

# Two features on very different scales: height (cm) and weight (kg)
X = np.array([[160, 55], [175, 70], [180, 80], [155, 50], [190, 95]])

scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)

print("Means used :", scaler.mean_)
print("SDs used   :", scaler.scale_.round(2))      # population SD (ddof=0)
print("Scaled data:\n", X_scaled.round(2))
print("New means  :", X_scaled.mean(axis=0).round(2) + 0.0)
print("New SDs    :", X_scaled.std(axis=0).round(2))

Both columns now have mean 0 and SD 1, so height (cm) and weight (kg) count equally in distance-based models. Note that StandardScaler divides by the population SD (12.88 for height, not the sample SD 14.40).

अब दोनों columns का mean 0 और SD 1 है, इसलिए distance-based models में height (cm) और weight (kg) बराबर गिने जाते हैं। ध्यान दें कि StandardScaler population SD से भाग देता है (height के लिए 12.88, sample SD 14.40 नहीं)।

现在两列的均值都是 0、标准差都是 1，所以在基于距离的模型里，身高（cm）和体重（kg）的分量相同。注意 StandardScaler 除的是总体标准差（身高为 12.88，而不是样本标准差 14.40）。

Anomaly Detection — and a TrapAnomaly Detection — और एक जाल异常检测 —— 以及一个陷阱

A common rule: flag any value more than 3 standard deviations from the mean. Let us test it on server response times with one obvious problem (500 ms).

एक आम नियम: mean से 3 standard deviations से ज़्यादा दूर वाली हर value को flag करें। चलिए इसे server response times पर आज़माएँ जिनमें एक साफ़ समस्या है (500 ms)।

一条常见规则：离均值超过 3 个标准差的值都标记为异常。我们在服务器响应时间上试一试，其中有一个明显的问题（500 ms）。

import numpy as np

# Server response times (ms)
response_times = np.array([120, 118, 122, 119, 121, 120, 118, 500, 123, 119])

mean    = response_times.mean()
std_dev = response_times.std(ddof=1)
print(f"Mean:    {mean:.1f} ms")
print(f"Std Dev: {std_dev:.1f} ms")

# Naive rule: anything above mean + 3 std devs is an anomaly
threshold = mean + 3 * std_dev
print(f"Threshold: {threshold:.1f} ms")
print("Anomalies detected:", response_times[response_times > threshold].tolist())

The fix: measure the centre and spread with statistics the outlier cannot move — the median and the MAD (median absolute deviation).

हल: केंद्र और फैलाव ऐसे statistics से मापें जिन्हें outlier हिला न सके — median और MAD (median absolute deviation)।

解决办法：用离群点撼动不了的统计量来衡量中心和离散程度 —— 中位数和 MAD（中位数绝对偏差）。

# Robust fix: use the median and the MAD (median absolute deviation),
# which the outlier cannot inflate.
median = np.median(response_times)
mad    = np.median(np.abs(response_times - median))
robust_sd = 1.4826 * mad            # MAD scaled to match the SD for normal data

robust_z = (response_times - median) / robust_sd
print(f"Median: {median:.1f} ms   robust SD: {robust_sd:.2f} ms")
print("Robust z of 500 ms:", round(robust_z[response_times == 500][0], 1))
print("Anomalies detected:", response_times[np.abs(robust_z) > 3].tolist())

Variance vs Standard Deviation — SummaryVariance vs Standard Deviation — सार方差 vs 标准差 — 总结

Golden RulesGolden Rules黄金法则