Exercise Type 3: Gaussian Posterior & Evidence (Gaussian-Gaussian Conjugacy)

What the exam asks: You are given a Gaussian likelihood (data model with unknown mean μ) and a Gaussian prior on μ. You must compute the posterior distribution over μ after seeing data, the evidence (marginal likelihood), and/or combine predictions from multiple models using Bayesian model averaging.


Part 0: What Do All These Symbols Mean?

The Key Notation

Symbol How to Read It What It Means
$\mu$ Greek letter "mu" The unknown mean — the parameter we're trying to estimate
$x$ lowercase x The observation — the data point we actually saw
$\sigma^2$ "sigma squared" A variance — how spread out a Gaussian distribution is
$\sigma_0^2$ "sigma naught squared" The prior variance — how uncertain we are about μ BEFORE seeing data
$\sigma^2_{post}$ "sigma post squared" The posterior variance — how uncertain we are about μ AFTER seeing data
$\mu_0$ "mu naught" The prior mean — what we believed μ was before seeing data
$\mu_{post}$ "mu post" The posterior mean — our updated estimate of μ after seeing data
$\sigma_\epsilon^2$ "sigma epsilon squared" The observation noise variance — how noisy our measurements are
$\mathcal{N}(x|\mu, \sigma^2)$ "Gaussian of x given mu and sigma squared" The Gaussian (Normal) distribution for x, with mean μ and variance $\sigma^2$
$\frac{1}{\sigma^2}$ "one over sigma squared" The precision — the inverse of variance. High precision = low uncertainty

The Gaussian Distribution Formula

How to read this: "The probability density of x given mean μ and variance $\sigma^2$." The exponential term gives the bell curve shape — highest when $x = \mu$, dropping off as x moves away.

In the exam you don't need to expand this — you just need to identify the mean and variance inside the $\mathcal{N}(\cdot|\cdot, \cdot)$ notation.


Part 1: The Core Concepts — No Math

What's Going On? (Plain English)

Imagine you're trying to figure out the true temperature μ of an oven. You have:

  1. A prior belief: "I think the oven is around 200°C, give or take 10°C." This is your Gaussian prior: $\mathcal{N}(\mu|200, 10^2)$.
  2. A measurement: You stick a thermometer in and it reads 210°C. But the thermometer is noisy. This is your likelihood: $\mathcal{N}(x=210|\mu, \sigma^2_{noise})$.
  3. An updated belief: After seeing the measurement, you revise your estimate. This is your posterior.

Bayesian inference with Gaussians is just the math of combining these two sources of information.

The Intuition Behind the Posterior

The posterior mean is a weighted average of the prior mean and the observation:

Who gets more weight? The one that's more certain (higher precision): - If the prior is very certain (small $\sigma_0^2$) → posterior stays close to $\mu_0$ - If the observation is very certain (small $\sigma^2$) → posterior moves toward $x$

The posterior variance is always smaller than both the prior and the likelihood variance — because you now have TWO sources of information.

The Intuition Behind the Evidence

The evidence $p(x)$ answers: "Before I saw the data, how likely was this particular observation under my model?"

For a Gaussian model with unknown mean, the evidence is a Gaussian centered at the prior mean with variance = prior variance + likelihood variance.

Why the summed variance? Because there are TWO sources of uncertainty: 1. We don't know μ (prior uncertainty = $\sigma_0^2$) 2. Even if we knew μ, the observation is noisy (likelihood uncertainty = $\sigma^2$)

These add up. The evidence is broader than the likelihood alone.


Part 2: The Key Formulas (MEMORIZE)

Setup

Component Formula What You Know
Likelihood $p(x \mid \mu) = \mathcal{N}(x \mid \mu, \sigma^2)$ $\sigma^2$ is given
Prior $p(\mu) = \mathcal{N}(\mu \mid \mu_0, \sigma_0^2)$ $\mu_0$ and $\sigma_0^2$ are given

Posterior: $p(\mu \mid x) = \mathcal{N}(\mu \mid \mu_{post}, \sigma^2_{post})$

Formula Expression How to Remember
Posterior precision $\dfrac{1}{\sigma^2_{post}} = \dfrac{1}{\sigma^2_0} + \dfrac{1}{\sigma^2}$ Precision = prior precision + likelihood precision
Posterior variance $\sigma^2_{post} = \dfrac{1}{\frac{1}{\sigma^2_0} + \frac{1}{\sigma^2}} = \dfrac{\sigma^2_0 \cdot \sigma^2}{\sigma^2_0 + \sigma^2}$ Inverse of the sum of precisions
Posterior mean $\mu_{post} = \sigma^2_{post} \left( \dfrac{\mu_0}{\sigma^2_0} + \dfrac{x}{\sigma^2} \right)$ Precision-weighted average

The Precision-Weighted Average (Rewritten)

In plain English: "Each source contributes in proportion to its precision. More precise sources get more weight."

Shortcut when both variances = 1: $\mu_{post} = (\mu_0 + x) / 2$ (simple average).

Evidence: $p(x)$

Mean = prior mean. Variance = prior variance + likelihood variance.


Part 3: Tricks & Shortcuts

TRICK 1: Posterior Precision = Sum of Precisions

Precision = $1/\sigma^2$. Just add them: $1/\sigma^2_{post} = 1/\sigma^2_0 + 1/\sigma^2$. Then flip: $\sigma^2_{post} = 1 / (\text{sum})$.

TRICK 2: Posterior Mean = Precision-Weighted Average

Each source contributes: (its mean) × (its precision). Then divide by total precision.

If both variances equal: posterior mean = simple average $(\mu_0 + x) / 2$.

TRICK 3: Evidence Variance = Sum of Variances

Evidence = $\mathcal{N}(x|\mu_0, \sigma^2_0 + \sigma^2)$. Mean stays at prior mean. Variances add.

TRICK 4: Posterior Variance Always Decreases

$\sigma^2_{post}$ is always smaller than both $\sigma^2_0$ and $\sigma^2$. More information → more certainty.


Part 4: FULL Walkthrough of Real Exam Questions

EXAM QUESTION 1 (2023, Question 1a-b)

Model $m_1$: $p(x|\mu, m_1) = \mathcal{N}(x|\mu, 1)$, $p(\mu|m_1) = \mathcal{N}(\mu|0, 1)$

1a: We observe $x=1$. Determine $p(\mu|x=1, m_1)$.

Options: - (a) $\mathcal{N}(\mu|0, 0.5)$ - (b) $\mathcal{N}(\mu|1, 2)$ - (c) $\mathcal{N}(\mu|0.5, 0.5)$ - (d) $\mathcal{N}(\mu|0.5, 1)$

STEP-BY-STEP SOLUTION

Step 1: Compute posterior variance

Step 2: Compute posterior mean

Answer: (c) $\mathcal{N}(\mu|0.5, 0.5)$


Question 1b: Evidence

Determine $p(x=1|m_1)$.

Options: - (a) $\mathcal{N}(1|0, 2)$ - (b) $2/\sqrt{2\pi}$ - (c) $\mathcal{N}(0|1, 1)$ - (d) $1/\sqrt{2\pi}$

STEP-BY-STEP SOLUTION

Answer: (a) $\mathcal{N}(1|0, 2)$


Tricks & Shortcuts

TRICK 1: Posterior precision = sum of precisions - Precision = 1/variance - Posterior precision = prior precision + likelihood precision - Posterior variance = 1/(sum of precisions)

TRICK 2: Posterior mean = precision-weighted average - More precise source gets more weight - If both have same variance: mean = (μ₀ + x) / 2

TRICK 3: Evidence = Gaussian with summed variances - $p(x) = \mathcal{N}(x|\mu_0, \sigma^2_0 + \sigma^2)$ - Mean = prior mean, variance = prior variance + likelihood variance

TRICK 4: When both variances = 1 - Posterior variance = 1/2 - Posterior mean = (prior_mean + observation) / 2


Part 5: Practice Exercises

Exercise 1 (2023, Question 1a)

Model $m_1$: $p(x|\mu, m_1) = \mathcal{N}(x|\mu, 1)$, $p(\mu|m_1) = \mathcal{N}(\mu|0, 1)$

We observe $x=1$. Determine $p(\mu|x=1, m_1)$.

Options: - (a) $\mathcal{N}(\mu|0, 0.5)$ - (b) $\mathcal{N}(\mu|1, 2)$ - (c) $\mathcal{N}(\mu|0.5, 0.5)$ - (d) $\mathcal{N}(\mu|0.5, 1)$


Exercise 2 (2023, Question 1b)

Same model. Determine $p(x=1|m_1)$.

Options: - (a) $\mathcal{N}(1|0, 2)$ - (b) $2/\sqrt{2\pi}$ - (c) $\mathcal{N}(0|1, 1)$ - (d) $1/\sqrt{2\pi}$


Exercise 3 (2023, Question 1c)

Model $m_1$: $p(x|\mu, m_1) = \mathcal{N}(x|\mu, 1)$, $p(\mu|m_1) = \mathcal{N}(\mu|0, 1)$ Model $m_2$: $p(x|m_2) = \mathcal{N}(x|1, 1)$ $p(m_1) = 2/3$, $p(m_2) = 1/3$

Determine $p(x=2)$ by Bayesian model averaging.

Options: - (a) $\frac{2}{3\sqrt{2\pi}} + \frac{1}{3}\mathcal{N}(2|0, 1)$ - (b) $\frac{1}{3}\mathcal{N}(2|1, 2) + \frac{1}{3\sqrt{2\pi}}$ - (c) $\frac{2}{3}\mathcal{N}(2|0, 2) + \frac{1}{3}\mathcal{N}(2|1, 1)$ - (d) $\frac{1}{3\sqrt{2\pi}} + \frac{1}{3}\mathcal{N}(2|1, 1)$

STEP-BY-STEP SOLUTION

Step 1: Write the Bayesian model averaging formula

We need two ingredients: the evidence $p(x=2|m_1)$ and the evidence $p(x=2|m_2)$, then weight by their prior model probabilities.


Step 2: Compute the evidence for model $m_1$

Model $m_1$ has unknown parameter μ. The likelihood is $\mathcal{N}(x|\mu, 1)$ and the prior on μ is $\mathcal{N}(\mu|0, 1)$.

Key formula — Gaussian evidence:

When likelihood = $\mathcal{N}(x|\mu, \sigma^2)$ and prior = $\mathcal{N}(\mu|\mu_0, \sigma_0^2)$:

How to read this: The evidence is a Gaussian centered at the prior mean with variance = prior variance + likelihood variance.

From $m_1$: - $\mu_0 = 0$ (prior mean) - $\sigma_0^2 = 1$ (prior variance) - $\sigma^2 = 1$ (likelihood variance)

Why is the variance 2? Because we don't know μ — we averaged over all possible μ values (the prior), which adds uncertainty on top of the observation noise.


Step 3: Compute the evidence for model $m_2$

Model $m_2$ has NO unknown parameter — it directly specifies:

So we just evaluate at $x=2$:

Key difference from $m_1$: $m_2$ is fully specified (no parameter to integrate over), so its evidence is just the Gaussian evaluated at 2. $m_1$ has unknown μ, so its evidence is broader (variance 2 instead of 1).


Step 4: Combine with model averaging


Step 5: Match the answer

(c) matches exactly. ✓

(a) Has $\frac{2}{3\sqrt{2\pi}}$ — this looks like someone tried to evaluate $\mathcal{N}(2|0, 1)$ numerically but got it wrong. Also has $\mathcal{N}(2|0, 1)$ instead of $\mathcal{N}(2|0, 2)$. ELIMINATE. (b) Has the model weights swapped ($1/3$ with the first term, $1/3\sqrt{2\pi}$ with the second). Also the second term is a number not a Gaussian. ELIMINATE. (d) Same issues — wrong weights and wrong terms. ELIMINATE.

Answer: (c) $\frac{2}{3}\mathcal{N}(2|0, 2) + \frac{1}{3}\mathcal{N}(2|1, 1)$


KEY INTUITION

Bayesian model averaging = weighted average of predictions

Think of it like asking two weather forecasters: - Forecaster $m_1$ says "20°C, but I'm uncertain (wide error bar)" - Forecaster $m_2$ says "20°C, but I'm more confident (narrow error bar)" - You trust forecaster $m_1$ twice as much (2/3 vs 1/3)

Your combined forecast = $2/3 \times m_1\text{'s prediction} + 1/3 \times m_2\text{'s prediction}$

Why does $m_1$ have variance 2? Because $m_1$ has an unknown parameter μ. It's uncertain about both the noise AND where μ is. $m_2$ has no unknown parameter — it's just one Gaussian with variance 1.


Exercise 4 (Kalman-Style Posterior Variance Update)

In a Kalman-style update, the posterior variance update is:

Options: - (a) $\sigma_k^2 = (1-K_k)\cdot\sigma_{k-1}^2 + \sigma_\epsilon^2$ - (b) $\sigma_k^2 = \sigma_{k-1}^2 + K_k\cdot(\sigma_\epsilon^2-\sigma_{k-1}^2)$ - (c) $\sigma_k^2 = K_k\cdot\sigma_{k-1}^2$ - (d) $\sigma_k^2 = (1-K_k)\cdot\sigma_{k-1}^2$

STEP-BY-STEP SOLUTION

Step 1: Understand the setup

This is the Kalman filter applied to estimating a static parameter θ. At step k: - Prior (before seeing $x_k$): $p(\theta|D_{k-1}) = \mathcal{N}(\theta|\mu_{k-1}, \sigma_{k-1}^2)$ - Likelihood: $p(x_k|\theta) = \mathcal{N}(x_k|\theta, \sigma_\epsilon^2)$ - Posterior (after seeing $x_k$): $p(\theta|D_k) = \mathcal{N}(\theta|\mu_k, \sigma_k^2)$

Step 2: Use the Gaussian multiplication rule

Posterior precision = prior precision + likelihood precision:

Step 3: Solve for $\sigma_k^2$

Now factor out $\sigma_{k-1}^2$:

Step 4: Recognize the Kalman gain

The Kalman gain is:

So $1 - K_k = 1 - \frac{\sigma_{k-1}^2}{\sigma_{k-1}^2 + \sigma_\epsilon^2} = \frac{\sigma_\epsilon^2}{\sigma_{k-1}^2 + \sigma_\epsilon^2}$

Step 5: Substitute

Step 6: Match the answer

(a) Adds $+\sigma_\epsilon^2$ at the end — this would make variance increase, which is wrong. ELIMINATE. (b) A completely different form — doesn't match the derivation. ELIMINATE. (c) Uses $K_k$ instead of $(1-K_k)$ — this would give the wrong factor. ELIMINATE. (d) Matches exactly. ✓

Answer: (d) $\sigma_k^2 = (1-K_k)\cdot\sigma_{k-1}^2$

KEY INTUITION

  • Since $0 < K_k < 1$, we have $0 < (1-K_k) < 1$
  • So $\sigma_k^2 < \sigma_{k-1}^2$ — posterior variance always decreases with each new observation
  • This makes sense: more data → more certainty → lower variance
  • The Kalman gain $K_k$ controls how fast: high $K_k$ (trusting the observation more) → bigger variance reduction

Answers

Exercise 1 **Answer: (c) 𝒩(μ|0.5, 0.5)** Posterior precision = 1/1 + 1/1 = 2, so variance = 1/2 = 0.5. Mean = 0.5 × (0/1 + 1/1) = 0.5
Exercise 2 **Answer: (a) 𝒩(1|0, 2)** Evidence = 𝒩(x|μ₀, σ²₀ + σ²) = 𝒩(1|0, 1+1) = 𝒩(1|0, 2)
Exercise 3 **Answer: (c)** p(x=2|m₁) = 𝒩(2|0, 2), p(x=2|m₂) = 𝒩(2|1, 1). Weighted sum: (2/3)𝒩(2|0, 2) + (1/3)𝒩(2|1, 1)
Exercise 4 **Answer: (d)** In the standard Kalman filter, σ²_k = (1-K_k)·σ²_{k-1}. The posterior variance decreases by the Kalman gain factor.