Machine Learning

Bias-Variance Decomposition

Bias-variance decomposition separates generalization error into target mismatch, training-set instability, and irreducible noise.

status: publishedimportance: criticaldifficulty 3/5math: undergraduateread: 19mlive demo

Back to Machine Learning Next: Classification Metrics, Thresholds, and Calibration

Concept Structure

Bias-Variance Decomposition

01Intuition

Start with the picture, metaphor, or geometric mechanism.

02Math

Make the objects explicit and connect them with notation.

03Code

Mirror the equations with runnable implementation details.

04Interactive Demo

Manipulate the mechanism and watch the idea respond.

1prerequisites

0next concepts

1related links

Learning map

Bias-Variance Decomposition

BeforeLinear Regression & Least SquaresNow4/4 sections readyTryManipulate one control and predict the visible change.NextClassification Metrics, Thresholds, and Calibration

Object flow

4/4 sections readyAsk about this Research room

ConceptBias-Variance DecompositionMachine Learning EquationBias-Variance Decomposition equation 1Exact equation object CodeBias-Variance Decomposition code witness 1Exact code witness DemoBias-Variance Decomposition interactive demoVisualization object ClaimAt a fixed input under squared error, expected prediction error decom...Exact claim check SourceAn Introduction to Statistical LearningExact source object

ConceptBias-Variance DecompositionMachine Learning

2 sources attachedLocal snapshot ready

concept:machine-learning/bias-variance-decomposition

Codewitness nearby Predictbefore reveal Roomobject handoff

Conceptual Bridge

What should feel connected as you move through this page.

Carry inLinear Regression & Least Squares

Bring the mental model from Linear Regression & Least Squares; this page will reuse it instead of restarting from zero.

Work hereBias-Variance Decomposition

Bias-variance decomposition separates generalization error into target mismatch, training-set instability, and irreducible noise.

Carry outClassification Metrics, Thresholds, and Calibration

The next edge should feel earned: use the demo prediction here before following Classification Metrics, Thresholds, and Calibration.

Test the linkManipulate one control and predict the visible change.Then continue to Classification Metrics, Thresholds, and Calibration

01IntuitionStart with the picture, metaphor, or geometric mechanism.02MathMake the objects explicit and connect them with notation.03CodeMirror the equations with runnable implementation details.04Interactive DemoManipulate the mechanism and watch the idea respond.

Intuition

Build the mental picture first so the rest of the page has something to attach to.

Section prompt

You are here because "overfitting" is too vague by itself. A model can fail because it is systematically wrong, because it changes too much when the training data changes, or because the target is noisy even in the best possible world.

Before this, know least-squares regression. By the end, you should be able to name the randomness source in the decomposition, predict when a model is bias-dominated or variance-dominated, and run a small resampling witness.

Bias-variance is about imagined repeats of the training process. Hold one test input $x_0$ fixed. Now imagine drawing many different training sets from the same data-generating process, fitting a model on each one, and asking how the predictions at $x_0$ vary.

Three things can hurt expected squared error:

Bias is systematic miss. If the average fitted model is far from the true function, the model class is not flexible enough or the features are wrong.

Variance is training-set sensitivity. If different training sets produce very different predictions at the same $x_0$ , the model is flexible enough to chase sample noise.

Noise is irreducible target randomness. If the observed label is $f^\star(x_0)+\epsilon$ , no model can predict the fresh $\epsilon$ before it happens.

The common trap is to treat bias and variance as properties of one fitted curve. They are properties of the learning procedure under repeated samples.

Math

Translate the story into symbols, assumptions, and a derivation you can inspect.

Section prompt

Equation 1Y=f^\star(x_0)+\epsilon,Equation 2\mathbb E_{D,\epsilon}\left[(Y-\hat f_D(x_0))^2\right].

Fix an input $x_0$ . Suppose the data-generating process is

Y=f^\star(x_0)+\epsilon,

where $f^\star(x_0)$ is the true conditional mean and $\epsilon$ has mean $0$ and variance $\sigma^2$ . Let $\hat f_D(x_0)$ be the prediction made by a model trained on dataset $D$ . The randomness in $\hat f_D$ comes from the draw of $D$ .

The expected squared prediction error at $x_0$ is

\mathbb E_{D,\epsilon}\left[(Y-\hat f_D(x_0))^2\right].

Add and subtract the average fitted prediction $\bar f(x_0)=\mathbb E_D[\hat f_D(x_0)]$ :

Y-\hat f_D(x_0) = \epsilon + \left(f^\star(x_0)-\bar f(x_0)\right) + \left(\bar f(x_0)-\hat f_D(x_0)\right).

When $\epsilon$ has mean $0$ and is independent of the training-set draw, the cross terms vanish after expectation. The decomposition becomes

\mathbb E_{D,\epsilon}\left[(Y-\hat f_D(x_0))^2\right] = \sigma^2 + \left(f^\star(x_0)-\bar f(x_0)\right)^2 + \mathbb E_D\left[\left(\hat f_D(x_0)-\bar f(x_0)\right)^2\right].

These are:

\text{noise}+\text{bias}^2+\text{variance}.

The decomposition is local to an input $x_0$ . A model can be low-bias in one region and high-bias elsewhere. Whole-dataset generalization error averages this picture over the input distribution.

The formula also assumes the fresh test noise has mean $0$ , has variance $\sigma^2$ , and is independent of the training-set draw. If those assumptions change, the clean three-term story can pick up extra terms.

Code

Keep the implementation aligned with the notation so the algorithm is legible.

Section prompt

Code witness 1import numpy as np rng = np.random.default_rng(7) probe = 1.4 noise_sd = 0.25 def f_star(x):...python

import numpy as np

rng = np.random.default_rng(7)
probe = 1.4
noise_sd = 0.25

def f_star(x):
    return np.sin(1.4 * x)

def one_fit(degree, n=18):
    x = rng.uniform(-3, 3, size=n)
    y = f_star(x) + rng.normal(0, noise_sd, size=n)
    coeff = np.polyfit(x, y, deg=degree)
    return np.polyval(coeff, probe)

for degree in [1, 3, 9]:
    preds = np.array([one_fit(degree) for _ in range(300)])
    mean_pred = preds.mean()
    bias2 = (mean_pred - f_star(probe)) ** 2
    variance = preds.var()
    expected_error = noise_sd ** 2 + bias2 + variance
    print("degree", degree)
    print("  bias^2:", round(bias2, 4))
    print("  variance:", round(variance, 4))
    print("  noise:", round(noise_sd ** 2, 4))
    print("  total:", round(expected_error, 4))

The code mirrors the math: each call to one_fit draws a new training set $D$ , preds estimates the distribution of $\hat f_D(x_0)$ , and the mean of those predictions estimates $\bar f(x_0)`. These are Monte Carlo estimates, not exact values. High-degree raw polynomial fits can also add numerical conditioning artifacts; that is a separate problem from statistical variance.

Interactive Demo

Use direct manipulation to connect the explanation to a moving system.

Section prompt

Choose model flexibility, training-set size, noise, and a probe point. Before revealing the resampled fits, predict which term dominates the local error at the probe: bias, variance, noise, or a rough balance.

The reveal overlays many fitted curves from different training sets. If the average curve misses the true function at the probe, bias dominates. If the curves disagree with each other at the probe, variance dominates. If both are small relative to the noise slider, the remaining error is irreducible noise. The answer is local to the probe point and estimated from a finite set of resamples.

The noise readout is the variance $\sigma^2$ , not the noise standard deviation. The fitted curves are a finite Monte Carlo witness, and the polynomial solver uses a tiny stabilizer so numerical blow-ups do not masquerade as statistical variance.

Live Concept Demo

Explore Bias-Variance Decomposition

The stage is code-native and interactive. Use it to test the explanation against the mechanism.

difficulty 3/5undergraduatecode-aligned

Demo Prediction Checkpoint

Manipulate one control and predict the visible change.

Commit to what Bias-Variance Decomposition should make visible before reading the result.

After The First Pass

Turn the concept into an inspected object.

Once the invariant is visible in the intuition, math, code, and demo, use these panels to inspect the mechanism visually, check source support, practice the idea, and attach a grounded research question.

Mechanism Storyboard

See the idea move before the page explains it

Bias-variance decomposition separates generalization error into target mismatch, training-set instability, and irreducible noise.

Prediction open01 / Intuition

Prediction lens

Start with the picture, metaphor, or geometric mechanism.

Commit first

Before reading further, choose the kind of change Bias-Variance Decomposition should make visible.

Visual Inquiry

Make the image answer a mathematical question

Bias-variance decomposition separates generalization error into target mismatch, training-set instability, and irreducible noise.

4/4 stages readyLive demo connected

Visual cueWhich visible object should carry the first intuition?

Inspection depth2/4

Prediction

Which visible object should carry the first intuition?

Commit first

Pick the cue that should make Bias-Variance Decomposition easier to reason about before the page gives the answer.

Source Grounding

Canonical references for the mechanism on this page.

book · 2023An Introduction to Statistical LearningJames, Witten, Hastie, Tibshirani, and Taylor

Claim source for expected test MSE at a fixed x0 and repeated-training-set bias-variance decomposition.

Open source

bookThe Elements of Statistical LearningHastie, Tibshirani, and Friedman

Curriculum source for statistical-learning decomposition and resampling intuition.

Open source

Claim Review

Bias-variance decomposition separates generalization error into target mismatch, training-set instability, and irreducible noise.

Status1 substantive review recorded

Claims without a substantive review badge still need exact source-support review.

Sources2 references

isl-bias-variance, esl-bias-variance

Witnesses4 local objects

Use equation, code, and demo objects to check whether the source support is operational.

Substantively reviewedAt a fixed input under squared error, expected prediction error decomposes into irreducible noise, squared bias, and variance over training-set draws.Claim metadata: source checked

ISLR directly supports the page's local fixed-x0 decomposition, repeated-training-set expectation, variance/bias terminology, irreducible-error term, and model-flexibility tradeoff framing.

Sources: An Introduction to Statistical LearningThis source check covers squared-error regression at a fixed input under the stated noise/training-draw assumptions; finite-resample demos are Monte Carlo estimates and raw high-degree polynomial fits can add numerical artifacts.A bounded review summary is present; still check caveats and exact source scope.

Checked ISLR section 2.2.2: it states expected test MSE for a given x0 decomposes into variance of fhat(x0), squared bias, and Var(epsilon), and defines the expectation by repeatedly estimating f on many training sets and testing at x0. The page preserves fixed-input locality, training-set randomness, fresh-noise, Monte Carlo, and conditioning caveats.

Reviewer: codex+gpt-pro-prior; reviewed 2026-06-28

source-span-isl-bias-variance math-object-1 math-object-2 code-witness-1 interactive-demo

Source support candidates

book 2023An Introduction to Statistical Learning

Claim source for expected test MSE at a fixed x0 and repeated-training-set bias-variance decomposition.

bookThe Elements of Statistical Learning

Curriculum source for statistical-learning decomposition and resampling intuition.

Mechanism witnesses

Equation 1

Y=f^\star(x_0)+\epsilon,

Equation 2

\mathbb E_{D,\epsilon}\left[(Y-\hat f_D(x_0))^2\right].

Code witness 1import numpy as np rng = np.random.default_rng(7) probe = 1.4 noise_sd = 0.25 def f_star(x):...Demo stateLive mechanism probe

Practice Loop

Try the idea before it explains itself

Bias-variance decomposition separates generalization error into target mismatch, training-set instability, and irreducible noise.

Readiness0/3 checks ready

Predict

Before touching the demo, predict one visible change that should happen in Bias-Variance Decomposition.

Hint 1

Reveal when your model needs a nudge.

Hint 2

Reveal when your model needs a nudge.

Hint 3

Reveal when your model needs a nudge.

Your answer canvas

Local checks

Claim

A concrete answer is on the canvas.

Mechanism

The answer names why the claim should hold.

Bridge

It touches the page context or a neighboring idea.

Misconception check

Object research drawerClose

ConceptBias-Variance DecompositionMachine Learning

Code witness comparisonBias-Variance Decomposition code witness 1rng = np.random.default_rng(7)Prediction before revealBias-Variance Decomposition interactive demoManipulate one control and predict the visible change.

Grounded room questionWhat is the smallest example that makes Bias-Variance Decomposition click without losing the math?Local snapshot ready

Research Room

Attach the question to an exact object

Pick the concept, equation, source, code witness, claim, misconception, or demo state before asking for help. The handoff stays grounded to that object.

Next local actionNo local draft saved yet

Open the draft below to save one note and next action in this browser.

conceptMachine Learning

Bias-Variance Decomposition

Anchored question

What is the smallest example that makes Bias-Variance Decomposition click without losing the math?

Local action draftNo local draft saved yetExpand only when ready to capture one local next action

Local action draft

This draft stays locally in this browser for concept:machine-learning/bias-variance-decomposition.

Draft noteNext action

No local draft saved.

Evidence to inspect

Source ids to inspect: isl-bias-variance, esl-bias-variance
Definition, prerequisite, and contrast concept links
The equation or code witness that makes the concept operational
One demo state that shows the invariant instead of a slogan

What would resolve this

The learner can state the mechanism in their own words
The learner can name the prerequisite that would repair confusion
The learner can predict how the mechanism changes under one perturbation

Grounded AI handoff

I am working in Continuous Function's research reading room. Object: concept - Bias-Variance Decomposition Object key: concept:machine-learning/bias-variance-decomposition Context: Machine Learning Anchor id: concept/concept-notebook/machine-learning/bias-variance-decomposition Open question: What is the smallest example that makes Bias-Variance Decomposition click without losing the math? Evidence to inspect: - Source ids to inspect: isl-bias-variance, esl-bias-variance - Definition, prerequisite, and contrast concept links - The equation or code witness that makes the concept operational - One demo state that shows the invariant instead of a slogan What would resolve this: - The learner can state the mechanism in their own words - The learner can name the prerequisite that would repair confusion - The learner can predict how the mechanism changes under one perturbation Answer as a careful research tutor: stay source-grounded, separate verified evidence from assumptions, name the relevant math objects, and end with one next action.

Open source object

concept/concept-notebook/machine-learning/bias-variance-decomposition
concept:machine-learning/bias-variance-decomposition

Learning Map

Before / Now / Try / Next

BeforeLinear Regression & Least Squares

NowIntuition → Math → Code → Demo

TryManipulate one control and predict the visible change.

NextClassification Metrics, Thresholds, and Calibration

Intuitionready
Mathready
Codeready
Interactive Demoready

Object Companion

Ask beside the selected object

Bias-variance decomposition separates generalization error into target mismatch, training-set instability, and irreducible noise.

Your question

GoalComfortStyleStuck on

Context prompt

You are my AI learning companion for Continuous Function. Current context: Machine Learning concept. Learning surface: Bias-Variance Decomposition. What this page says: Bias-variance decomposition separates generalization error into target mismatch, training-set instability, and irreducible noise. Current section: Intuition, math, code, and interactive demo. Suggested next step: Manipulate one control and predict the visible change.. Learner goal: Understand the idea. Learner comfort level: New to this. Preferred explanation style: Visual first. Task: Explain the central idea in plain language, then restate it with the exact math objects from the page. Answer in a way that helps me learn: ask one clarifying question only if needed, use intuition before notation, and end with one thing I should try on the page.

Domain

Machine Learning

machine-learninggeneralizationstatisticsmodel-selection

Prerequisites

Linear Regression & Least Squares

Gradient Descent

Within this domain