Conditional Expectation and Martingales

Conditional expectation

E [X ∣ G]

is the best prediction of a random variable

X

given the information encoded by a

σ

-algebra

G

; it is the unique (a.s.)

G

-measurable, integrable random variable whose averages over

G

-sets match those of

X

. A martingale is a stochastic process

(X_{n})

that is its own best forecast:

E [X_{n + 1} ∣ F_{n}] = X_{n}

, modeling a fair game. These objects underlie modern probability, stochastic calculus, and mathematical finance, and were forged by Kolmogorov (1933), Lévy, and especially Joseph L. Doob (1940s–50s).

Conditional expectation generalizes the elementary notion $E [X ∣ Y = y]$ to conditioning on an entire $σ$ -algebra, replacing pointwise values with a measurable projection. It matters because it is the rigorous device for "partial information": filtering, prediction, Bayesian updating, and the very definition of martingales all rest upon it. A martingale formalizes a fair game where, given the whole past, the expected future increment is zero — and remarkably, such processes are forced to converge, to satisfy sharp maximal and $L^{p}$ inequalities, and to be impossible to beat by any non-anticipating strategy (optional stopping). The surprising fact: conditional expectation is literally an orthogonal projection in $L^{2}$ — geometry hides inside probability.

The Core Insight

Information is a $σ$ -algebra. To "condition on information $G$ " is to find the unique random variable that (i) depends only on $G$ and (ii) has the same integral as $X$ on every event you can resolve with $G$ . A martingale is the special case where today's value already is the best forecast of tomorrow's — so accumulated forecast errors form an orthogonal, mean-zero sequence, and the laws of large numbers and convergence theorems descend from that structure.

Formal Statement

Conditional Expectation given a σ-algebra

Let $(Ω, F, P)$ be a probability space, $G \subseteq F$ a sub- $σ$ -algebra, and $X \in L^{1} (Ω, F, P)$ (i.e. $E ∣ X ∣ < \infty$ ). A conditional expectation of $X$ given $G$ is any random variable $Y = E [X ∣ G]$ satisfying:

$Y$ is $G$ -measurable;
$Y$ is integrable;
The partial averaging (defining) identity holds:

\int_{G} Y d P = \int_{G} X d P for all G \in G .

⟳Property (3) is the definition; existence/uniqueness (below) make

Y

well-defined up to a

P

-null set. Symbols:

G

ranges over events resolvable by

G

; integrals are Lebesgue integrals against

P

Such $Y$ exists and is unique up to $P$ -a.s. equality. When $G = σ (Z)$ for a random variable $Z$ , we write $E [X ∣ Z]$ ; there is a Borel function $g$ with $E [X ∣ Z] = g (Z)$ a.s.

Existence and Uniqueness (Radon–Nikodym / Hilbert projection)

For every $X \in L^{1} (P)$ and sub- $σ$ -algebra $G$ , $E [X ∣ G]$ exists and is unique up to $P$ -null sets. For $X \in L^{2} (P)$ , $E [\cdot ∣ G]$ is the orthogonal projection of $L^{2} (F)$ onto the closed subspace $L^{2} (G)$ :

E [X ∣ G] = Π_{L^{2} (G)} (X) = ar g Y \in L^{2} (G) min E [(X - Y)^{2}] .

⟳

L^{2} (G)

is a closed subspace of the Hilbert space

L^{2} (F)

; the projection theorem gives a unique minimizer

Y

with

X - Y ⊥ L^{2} (G)

, i.e.

E [(X - Y) 1_{G}] = 0

, which is exactly [1]. Extend from

L^{2}

L^{1}

by approximation/monotone limits.

Filtration, Adapted Process, Martingale

A filtration is an increasing family $(F_{n})_{n \geq 0}$ of sub- $σ$ -algebras, $F_{0} \subseteq F_{1} \subseteq \dots \subseteq F$ , modeling information accumulating over time. A process $(X_{n})$ is adapted if each $X_{n}$ is $F_{n}$ -measurable. An adapted, integrable process is a martingale if

E [X_{n + 1} ∣ F_{n}] = X_{n} a.s. for all n \geq 0.

⟳This is the "fair game" axiom. Replacing "

=

" with "

\geq

" gives a submartingale (favorable game); "

\leq

" gives a supermartingale. Taking expectations in [3] gives constant mean

E [X_{n}] = E [X_{0}]

Equivalently, the increments $d_{n} = X_{n} - X_{n - 1}$ satisfy $E [d_{n} ∣ F_{n - 1}] = 0$ — they form a martingale difference sequence.

Fundamental Properties of Conditional Expectation

Let $X, Y \in L^{1}$ , $a, b \in R$ , and $H \subseteq G \subseteq F$ .

E [a X + bY ∣ G] = a E [X ∣ G] + b E [Y ∣ G] (linearity) .

⟳Both sides are

G

-measurable and satisfy [1] by linearity of the integral; uniqueness forces equality.

E [E [X ∣ G] ∣ H] = E [X ∣ H] (tower / smoothing) .

⟳The RHS is

H

-measurable; for

H \in H \subseteq G

\int_{H} E [X ∣ G] = \int_{H} X

by [1] (since

H \in G

), matching

\int_{H} E [X ∣ H]

E [Z X ∣ G] = Z E [X ∣ G] if Z is G -measurable, Z X \in L^{1} (taking out what is known) .

⟳Verify [1] for

Z = 1_{G_{0}}

G_{0} \in G

, then extend by linearity, monotone convergence, and approximation to general bounded/integrable

G

-measurable

Z

φ (E [X ∣ G]) \leq E [φ (X) ∣ G] for convex φ (conditional Jensen) .

⟳Write

φ (x) = sup_{n} (a_{n} x + b_{n})

over supporting lines; apply linearity and monotonicity to

a_{n} X + b_{n} \leq φ (X)

, then take the sup inside via

G

-measurability.

Why It Holds / Derivation

:::proof[Existence of $E [X ∣ G]$ via Radon–Nikodym] Step 1 (reduce to $X \geq 0$ ). Write $X = X^{+} - X^{-}$ ; treat each part, then subtract using linearity. Assume $X \geq 0$ , $E [X] < \infty$ .

Step 2 (build a measure). Define

ν (G) = \int_{G} X d P

for

G \in G

. Then

ν

is a finite measure on

(Ω, G)

, and

ν (G) = 0

whenever

P (G) = 0

, so

ν ≪ P ∣_{G}

(absolute continuity). Step 3 (Radon–Nikodym). By the Radon–Nikodym theorem there is a

G

-measurable density

Y = \frac{d ν}{d P ∣ _{G}} \geq 0

with

ν (G) = \int_{G} Y d P for all G \in G .

⟳This is exactly [1] with

ν (G) = \int_{G} X d P

. The density is

G

-measurable because

ν, P ∣_{G}

live on

G

Step 4 (integrability and uniqueness). Taking

G = Ω

gives

E [Y] = E [X] < \infty

, so

Y \in L^{1}

. If

Y_{1}, Y_{2}

both satisfy [1], then

\int_{G} (Y_{1} - Y_{2}) d P = 0

for all

G \in G

; choosing

G = {Y_{1} > Y_{2}} \in G

forces

P (Y_{1} > Y_{2}) = 0

, and symmetrically, so

Y_{1} = Y_{2}

a.s.

■

Doob's Martingale Convergence Theorem (upcrossing argument)

Claim. If

(X_{n})

is a martingale (or supermartingale) with

sup_{n} E [X_{n}^{-}] < \infty

(equivalently

sup_{n} E ∣ X_{n} ∣ < \infty

), then

X_{n} \to X_{\infty}

a.s. with

X_{\infty} \in L^{1}

. Step 1 (upcrossings). For reals

a < b

, let

U_{n} [a, b]

be the number of times

(X_{k})_{k \leq n}

crosses upward from below

a

to above

b

. Doob's upcrossing inequality states

(b - a) E [U_{n} [a, b]] \leq E [(X_{n} - a)^{-}] \leq ∣ a ∣ + E ∣ X_{n} ∣.

⟳Build a predictable

0/1

"betting" strategy

C_{k}

that buys below

a

and sells above

b

; the resulting gain

(C \cdot X)_{n} = \sum_{k} C_{k} (X_{k} - X_{k - 1})

is a supermartingale with mean

\leq 0

, and each completed upcrossing contributes at least

(b - a)

Step 2 (finite upcrossings). Let

U_{\infty} [a, b] = lim_{n} U_{n} [a, b]

. Monotone convergence in [9] gives

(b - a) E [U_{\infty} [a, b]] \leq ∣ a ∣ + sup_{n} E ∣ X_{n} ∣ < \infty

, so

U_{\infty} [a, b] < \infty

a.s. Step 3 (rule out oscillation). On the event

{lim inf X_{n} < lim sup X_{n}}

there exist rationals

a < b

with

lim inf < a < b < lim sup

, forcing

U_{\infty} [a, b] = \infty

. Taking the union over countably many rational pairs,

P (lim inf X_{n} < lim sup X_{n}) = 0

, so

X_{n} \to X_{\infty} \in [- \infty, \infty]

a.s. Step 4 (integrability). Fatou's lemma:

E ∣ X_{\infty} ∣ \leq lim inf E ∣ X_{n} ∣ \leq sup_{n} E ∣ X_{n} ∣ < \infty

, so

X_{\infty}

is finite a.s. and in

L^{1}

■

Joseph Leo Doob | 1910–2004

Doob, working at the University of Illinois, transformed an informal "fair game" idea (named after a French betting strategy) into a rigorous theory between 1940 and 1953, culminating in his monograph Stochastic Processes (1953). He proved the martingale convergence theorem, the optional stopping theorem, the maximal and $L^{p}$ inequalities, and the Doob–Meyer decomposition. He insisted on $σ$ -algebras as the carrier of information and on conditional expectation as a projection — a viewpoint that made stochastic analysis possible. Paul Lévy independently developed closely related ideas; Kolmogorov's 1933 Grundbegriffe had supplied the measure-theoretic foundation via Radon–Nikodym.

Building Intuition

Think of $G$ as a blurry photograph of $Ω$ : you can only distinguish events $G$ can resolve. $E [X ∣ G]$ replaces $X$ by its average within each cell of the blur — constant on indistinguishable outcomes, but matching the true average over every region you can see (that is [1]). In $L^{2}$ this is literally dropping a perpendicular: $E [X ∣ G]$ is the closest $G$ -measurable shadow of $X$ , and the error $X - E [X ∣ G]$ is orthogonal to everything you know.

A martingale is then a process where the best guess of tomorrow, using all of today's information, equals today's value. The increments are pure "surprise": uncorrelated with the past and mean-zero. Because surprises orthogonally accumulate, the variance grows additively (like a random walk), yet the upcrossing argument shows a martingale cannot oscillate forever without unbounded energy — hence convergence. The tower property [5] is the formal statement that "coarsening a refined forecast gives the coarse forecast": you cannot gain by predicting your own prediction.

Conditional expectation as averaging over a partition; martingale as a constant-mean random walk

<svg viewBox="0 0 760 380" xmlns="http://www.w3.org/2000/svg">
<rect width="760" height="380" fill="#0a0e17"/>
<text x="20" y="28" fill="#e2e8f0" font-size="15" font-family="sans-serif">Conditional expectation: average X over each cell of the σ-algebra G</text>
<!-- partition cells -->
<line x1="40" y1="50" x2="40" y2="190" stroke="#94a3b8"/>
<line x1="340" y1="50" x2="340" y2="190" stroke="#94a3b8"/>
<line x1="40" y1="190" x2="340" y2="190" stroke="#94a3b8"/>
<line x1="140" y1="50" x2="140" y2="190" stroke="#94a3b8" stroke-dasharray="4 3"/>
<line x1="240" y1="50" x2="240" y2="190" stroke="#94a3b8" stroke-dasharray="4 3"/>
<text x="78"  y="205" fill="#e2e8f0" font-size="12">G1</text>
<text x="178" y="205" fill="#e2e8f0" font-size="12">G2</text>
<text x="278" y="205" fill="#e2e8f0" font-size="12">G3</text>
<!-- wiggly X -->
<path d="M40,150 C70,90 110,170 140,110 C170,70 210,160 240,130 C270,100 310,140 340,95" fill="none" stroke="#94a3b8" stroke-width="1.5"/>
<text x="345" y="98" fill="#94a3b8" font-size="12">X(ω)</text>
<!-- piecewise constant E[X|G] -->
<line x1="40" y1="125" x2="140" y2="125" stroke="#d4a853" stroke-width="3"/>
<line x1="140" y1="105" x2="240" y2="105" stroke="#d4a853" stroke-width="3"/>
<line x1="240" y1="118" x2="340" y2="118" stroke="#d4a853" stroke-width="3"/>
<text x="345" y="122" fill="#d4a853" font-size="12">E[X|G]</text>
<!-- martingale panel -->
<text x="430" y="28" fill="#e2e8f0" font-size="15" font-family="sans-serif">Martingale: E[X_{n+1}|F_n]=X_n</text>
<line x1="430" y1="50" x2="430" y2="200" stroke="#94a3b8"/>
<line x1="430" y1="125" x2="740" y2="125" stroke="#94a3b8" stroke-dasharray="3 3"/>
<text x="600" y="120" fill="#94a3b8" font-size="11">mean level E[X0]</text>
<polyline points="430,125 460,100 490,140 520,95 550,150 580,110 610,135 640,90 670,130 700,108 730,118" fill="none" stroke="#d4a853" stroke-width="2"/>
<text x="438" y="218" fill="#e2e8f0" font-size="11">n=0</text>
<text x="715" y="218" fill="#e2e8f0" font-size="11">n</text>
<text x="430" y="250" fill="#e2e8f0" font-size="12" font-family="sans-serif">Increments d_n = X_n − X_{n−1} are mean-zero given the past.</text>
<text x="430" y="270" fill="#e2e8f0" font-size="12" font-family="sans-serif">Constant expectation; no drift; bounded in L¹ ⇒ converges a.s.</text>
<text x="20" y="250" fill="#e2e8f0" font-size="12" font-family="sans-serif">Gold step matches the gray curve's average on each cell:</text>
<text x="20" y="270" fill="#e2e8f0" font-size="12" font-family="sans-serif">∫_{Gi} E[X|G] dP = ∫_{Gi} X dP  (partial averaging).</text>
</svg>

Worked Examples

Conditional expectation given a discrete partition

Given. Roll a fair die,

Ω = {1, \dots, 6}

each with

P = \frac{1}{6}

. Let

X (ω) = ω^{2}

(the squared face). Let

G = σ ({1, 2, 3}, {4, 5, 6})

— you only learn "low" vs "high". Steps.

$G$ has two atoms: $L = {1, 2, 3}$ , $H = {4, 5, 6}$ , each with $P = \frac{1}{2}$ .
On an atom $A$ , $E [X ∣ G] = \frac{1}{P ( A )} \int_{A} X d P = \frac{\sum _{ω \in A} X ( ω )}{∣ A ∣}$ .
Low: $\frac{1 + 4 + 9}{3} = \frac{14}{3}$ . High: $\frac{16 + 25 + 36}{3} = \frac{77}{3}$ .

Check [1] on $Ω$ .

E [E [X ∣ G]] = \frac{1}{2} \cdot \frac{14}{3} + \frac{1}{2} \cdot \frac{77}{3} = \frac{91}{6}

, and

E [X] = \frac{1 + 4 + 9 + 16 + 25 + 36}{6} = \frac{91}{6} .

✓

E [X ∣ G] = \frac{14}{3} 1_{{1, 2, 3}} + \frac{77}{3} 1_{{4, 5, 6}} \approx 4.67 on low, 25.67 on high (dimensionless)

Optional stopping for a gambler's fair random walk

Given.

X_{n} = \sum_{k = 1}^{n} ξ_{k}

with i.i.d.

ξ_{k} = \pm 1

each prob

\frac{1}{2}

X_{0} = 0

(symmetric simple random walk; a martingale since

E [ξ_{k}] = 0

). Start with capital that lets you bet between barriers

- a

and

+ b

(

a, b > 0

integers). Let

T = in f {n : X_{n} = - a or X_{n} = b}

, the first hit of a barrier. Steps.

$T$ is a stopping time and $E [T] < \infty$ (finite expected exit time for bounded walk), with $∣ X_{n \land T} ∣ \leq max (a, b)$ , so the Optional Stopping Theorem applies:

E [X_{T}] = E [X_{0}] = 0.

⟳

(X_{n \land T})

is a bounded martingale; uniform integrability/bounded convergence gives

E [X_{T}] = lim_{n} E [X_{n \land T}] = E [X_{0}]

Let $p = P (hit b first)$ , so $1 - p = P (hit - a first)$ . Then $E [X_{T}] = p \cdot b + (1 - p) (- a) = 0$ .
Solve: $p b = (1 - p) a \Rightarrow p = \frac{a}{a + b}$ .
Expected duration via the martingale $X_{n}^{2} - n$ (since $E [X_{n}^{2} - n]$ is constant): $E [T] = E [X_{T}^{2}] = p b^{2} + (1 - p) a^{2} = \frac{a}{a + b} b^{2} + \frac{b}{a + b} a^{2} = ab .$

Numbers. With

a = 3

b = 5

P (reach + 5) = \frac{3}{8} = 0.375, E [T] = 3 \cdot 5 = 15 steps

Common Misconceptions

Conditional expectation is a random variable, not a number.

E [X ∣ G]

is a function on

Ω

G

-measurable. Only

E [X ∣ Y = y]

for a fixed value

y

is a number. Confusing the two breaks the tower property. Optional stopping needs a hypothesis.

E [X_{T}] = E [X_{0}]

can FAIL for unbounded

T

or unbounded

X_{n \land T}

. One sufficient condition:

T

bounded; or

E [T] < \infty

with bounded increments; or

(X_{n \land T})

uniformly integrable. Martingale ≠ Markov, and Markov ≠ martingale. A martingale is defined by a mean condition relative to the whole filtration; a Markov process is defined by distributional memorylessness. Each can hold without the other. $E [X ∣ G]$ depends on $P$ , not just on $G$ and $X$ . Changing the measure (e.g. a Girsanov change) changes conditional expectations and can turn a non-martingale into a martingale.

A martingale where naive optional stopping fails — the doubling strategy

Let $X_{n}$ be a fair random walk and consider the gambler who doubles the bet after each loss until the first win, stopping at $T = in f {n : first win}$ . The wealth process is a martingale and $T < \infty$ a.s., yet $X_{T} = + 1$ deterministically, so $E [X_{T}] = 1 \neq = 0 = E [X_{0}]$ . Optional stopping is violated because $(X_{n \land T})$ is not uniformly integrable: the potential losses before stopping are unbounded. This is precisely why casinos impose table limits and why "sure-win" systems require infinite capital.

Connections

Conditional expectation is the analytic backbone of Bayesian Inference (posterior means are conditional expectations) and of Filtering Theory (the Kalman filter computes $E [state ∣ observations]$ ). Through [2], it ties probability to Hilbert Space Projection and the geometry of $L^{2}$ . Martingales generalize sums of independent mean-zero variables, refining the Law of Large Numbers and the Central Limit Theorem (via the martingale CLT). The Doob–Meyer decomposition links submartingales to Stochastic Integration and the Itô Calculus, where the integral $\int H d M$ against a martingale is again a martingale. Optional stopping connects to Potential Theory (harmonic functions of Brownian motion are martingales) and to the theory of Markov Processes via the generator and Dynkin's formula.

Real-World Applications

Mathematical finance (no-arbitrage pricing). The Fundamental Theorem of Asset Pricing states that absence of arbitrage is equivalent to the existence of an equivalent measure under which discounted prices are martingales; option prices are conditional expectations $E^{Q} [payoff ∣ F_{t}]$ (Black–Scholes, Harrison–Pliska).
Kalman and particle filtering. Tracking, GPS, and robotics compute the conditional expectation of a hidden state given noisy measurements — the optimal $L^{2}$ estimator by [2].
Sequential analysis & A/B testing. Wald's sequential probability ratio test and "always-valid" $p$ -values use martingale maximal inequalities (Ville's inequality) to control error under optional stopping.
Stochastic approximation & SGD. Robbins–Monro and stochastic gradient descent are analyzed via supermartingale convergence (the Robbins–Siegmund lemma) to prove a.s. convergence.
Insurance and ruin theory. Surplus processes are modeled with (super/sub)martingales; ruin probabilities follow from optional stopping and exponential martingales (Lundberg bound).
Branching processes / genetics. The normalized population size $Z_{n} / m^{n}$ is a martingale (Galton–Watson), whose limit (Kesten–Stigum) governs survival and growth.
Information theory & gambling. Kelly betting and universal portfolios exploit the martingale property of likelihood ratios and wealth under fair odds.

Historical Development

1654 — Pascal–Fermat correspondence on the "problem of points" computes fair-division expectations: the embryo of conditional expectation and fair games.
1900 — Louis Bachelier models stock prices by Brownian motion, implicitly using the martingale (fair-game) property in Théorie de la spéculation.
1909–1912 — Émile Borel and others develop measure theory; the term martingale (a doubling betting system) enters probabilistic vocabulary.
1933 — Kolmogorov's Grundbegriffe der Wahrscheinlichkeitsrechnung gives the measure-theoretic foundation and defines conditional expectation via Radon–Nikodym.
1934–1940 — Paul Lévy studies sums of dependent variables and "conditionally centered" sequences, proving early convergence results (Lévy's martingale convergence / 0–1 law).
1940 — J. L. Doob introduces and names the systematic theory of martingales, proving convergence and inequalities.
1953 — Doob's Stochastic Processes codifies optional stopping, maximal inequalities, $L^{p}$ inequalities, and regularization.
1962–1963 — Paul-André Meyer and the Strasbourg school develop the Doob–Meyer decomposition and the "general theory of processes," enabling continuous-time stochastic integration.
1979–1981 — Harrison, Kreps, and Pliska recast option pricing entirely in martingale terms, launching modern quantitative finance.

Subtleties & Limits

Regular conditional distributions need not exist in pathological (non-Polish) spaces; the standard guarantee requires the underlying space to be a standard Borel space. Constructing them in greater generality, and understanding when "disintegration" fails, remains technically delicate. Continuous-time foundations require the "usual conditions" (right-continuity and completeness of the filtration) and notions of predictability; the choice of filtration is not innocent, and questions about enlargement of filtrations (initial vs progressive) and their effect on the martingale property are an active research area (informed trading, insider information models). Branching-measure and noncommutative analogues. Free-probability martingales and quantum (von Neumann algebra) conditional expectations generalize [1]–[2] to noncommutative

L^{p}

spaces; sharp constants in noncommutative Burkholder–Davis–Gundy and Doob inequalities are still being refined. Optimal constants in classical martingale inequalities (Doob's

L^{p}

maximal inequality, Burkholder's sharp constants for martingale transforms and the UMD property) connect to deep problems in harmonic analysis (the Beurling–Ahlfors transform), where exact best constants remain partly conjectural.

Conditional Expectation and Martingales

The Core Insight

Formal Statement

Conditional Expectation given a σ-algebra

Existence and Uniqueness (Radon–Nikodym / Hilbert projection)

Filtration, Adapted Process, Martingale

Fundamental Properties of Conditional Expectation

Why It Holds / Derivation

Doob's Martingale Convergence Theorem (upcrossing argument)

Joseph Leo Doob | 1910–2004

Building Intuition

Conditional expectation as averaging over a partition; martingale as a constant-mean random walk

Worked Examples

Conditional expectation given a discrete partition

Optional stopping for a gambler's fair random walk

Common Misconceptions

A martingale where naive optional stopping fails — the doubling strategy

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also

Before this

Where this leads