Skip to main content
PathsChebyshev variance tail

Theorem trail

Chebyshev variance tail

How variance turns Markov into a two-sided deviation bound.

Record

Research Record

A compact audit view for the theorem before the full source, proof, diagnostic, and failure-mode sections.

Claim scope
theorem
Evidence level
Exact Lean wrapper
Source precision
reviewed / A
Diagnostics
1 item
Exercises
1 exercise
Failure checks
2 checks

Theorem Statement

Chebyshev uses variance information: if the spread of around its mean is small ( is small), then large deviations are unlikely. Applied to the sample mean, the variance decreases as , giving the familiar concentration.

Assumptions

  • finite variance

Proof Sketch

Apply Markov to the nonnegative random variable (X - mu)^2. The event |X - mu| >= t is the same as (X - mu)^2 >= t^2, so the numerator becomes Var(X) and the denominator becomes t^2.

Proof Obligations

Finite variance and positive threshold

assumption

Verify the variable has finite variance and the deviation threshold is positive before replacing the event |X - mu| >= t by the squared event.

Without finite variance the numerator is not a usable quantity, and a nonpositive threshold makes the displayed tail event meaningless.

Square the deviation before applying Markov

proof step

The proof works because (X - mu)^2 is nonnegative and the event |X - mu| >= t is equivalent to (X - mu)^2 >= t^2.

Applying Markov directly to X - mu would lose nonnegativity and invalidate the tail-bound argument.

Variance-only information gives polynomial decay

boundary

Chebyshev should be used as the variance-only baseline; it cannot deliver the logarithmic finite-sample rates obtained from bounded-variable concentration.

Treating Chebyshev like a sub-Gaussian inequality overstates what the assumptions justify.

Centered deviation form is required

boundary

Use the theorem on deviations from the mean, not raw tails around zero, unless the mean is known and the event has been translated correctly.

Forgetting the centering step can turn a variance bound into a statement about the wrong event.

Source Locators

High-Dimensional Statistics (2019)

proof

Chapter 2 — Basic tail and concentration bounds

wainwright-2019:ch2

kind
chapter
review
reviewed / A
extract
manual

Chapter-level canonical support for Chebyshev-style tail bounds.

High-Dimensional Statistics (2019)

proof

Section 2.1.1 — From Markov to Chernoff

wainwright-2019:sec2.1

kind
section
review
reviewed / A
extract
manual
number
2.1.1

Anchor: From Markov to Chernoff

Reviewed section-level support for deriving Chebyshev as a Markov-style variance tail bound.

Curated Trace

  1. 1

    Variance is the expectation of the squared distance from the mean.

    Checkpoint

    Confirm the second moment is finite.

  2. 2

    The squared deviation is nonnegative, so Markov applies.

    Checkpoint

    Translate |X - mu| >= t into a squared event.

  3. 3

    Substituting Var(X) and t^2 gives the polynomial tail bound.

    Checkpoint

    The bound cannot deliver log(1/delta) sample complexity by itself.

  4. 4

    Chebyshev on the sample mean gives the standard weak law route.

    Checkpoint

    Track the 1/n variance shrinkage, not just the final limit.

Failure Modes

Variance-only information gives polynomial decay

Chebyshev should be used as the variance-only baseline; it cannot deliver the logarithmic finite-sample rates obtained from bounded-variable concentration.

Risk: Treating Chebyshev like a sub-Gaussian inequality overstates what the assumptions justify.

Check: Chebyshev should be used as the variance-only baseline; it cannot deliver the logarithmic finite-sample rates obtained from bounded-variable concentration.

Centered deviation form is required

Use the theorem on deviations from the mean, not raw tails around zero, unless the mean is known and the event has been translated correctly.

Risk: Forgetting the centering step can turn a variance bound into a statement about the wrong event.

Check: Use the theorem on deviations from the mean, not raw tails around zero, unless the mean is known and the event has been translated correctly.