Weak Topology and Functional Testing

Weak Convergence Means Passing Every Functional Test

Norm convergence asks whether the distance between vectors tends to zero. Weak convergence asks instead whether every bounded linear functional reports a convergent scalar value. This change of viewpoint is useful in $L^p$ spaces because their bounded linear functionals, for finite $p$, can be represented by integration against functions in a conjugate $L^q$ space. The word “every” is essential: passing one integral test, or even many selected tests, need not establish weak convergence.

Let $B$ be a real Banach space, with norm $\|\cdot\|_B$. Its continuous dual $B^*$ is the space of bounded linear functionals $\Lambda:B\to\mathbb R$, equipped with the operator norm $\|\Lambda\|=\sup_{\|x\|_B\le1}|\Lambda(x)|$. A sequence $(x_n)$ in $B$ converges weakly to $x\in B$ when $\Lambda(x_n)\to\Lambda(x)$ for every $\Lambda\in B^*$. We write $x_n\rightharpoonup x$.

Definition: The weak topology on $B$ is the coarsest topology for which every $\Lambda\in B^*$ is continuous. Equivalently, a basic neighborhood of $x\in B$ is specified by finitely many functionals $\Lambda_1,\ldots,\Lambda_k\in B^*$ and positive tolerances $\varepsilon_1,\ldots,\varepsilon_k$: $$ \{y\in B:\ |\Lambda_j(y-x)|<\varepsilon_j\text{ for }j=1,\ldots,k\}. $$ A sequence converges in this topology exactly when it passes every functional test: $\Lambda(x_n)\to\Lambda(x)$ for all $\Lambda\in B^*$.

The neighborhood description follows because finite intersections of inverse images of open intervals under the functionals form a basis for the coarsest topology making them continuous. For the sequence statement, convergence in the weak topology implies convergence after applying each continuous $\Lambda$. Conversely, suppose every scalar sequence $\Lambda(x_n)$ converges to $\Lambda(x)$. Given a basic neighborhood as above, for each $j$ there is an index $N_j$ such that $n\ge N_j$ implies $|\Lambda_j(x_n-x)|<\varepsilon_j$. Taking $N=\max_{1\le j\le k}N_j$ puts every $x_n$, $n\ge N$, in that neighborhood. Thus $x_n$ converges weakly to $x$.

The $L^p$-$L^q$ Testing Characterization

Let $(X,\mathcal M,\mu)$ be a sigma-finite measure space. For $1\le p<\infty$, $L^p(\mu)$ consists of equivalence classes, under equality $\mu$-a.e., of measurable real-valued functions with finite norm $\|f\|_p=(\int_X|f|^p\,d\mu)^{1/p}$. Set $q=\infty$ when $p=1$; when $1<p<\infty$, define $q$ by $1/p+1/q=1$. Thus $L^q(\mu)$ is the dual testing space, with the usual essential-supremum norm when $q=\infty$.

Theorem (Weak convergence in $L^p$ by integral testing): Let $(X,\mathcal M,\mu)$ be sigma-finite, let $1\le p<\infty$, and let $q$ be the conjugate exponent just defined. For a sequence $(f_n)$ and $f$ in the real Banach space $L^p(\mu)$, $$ f_n\rightharpoonup f \quad\Longleftrightarrow\quad \int_X f_n g\,d\mu\longrightarrow\int_X f g\,d\mu \quad\text{for every }g\in L^q(\mu). $$ The integrals are finite, and the assertion concerns equivalence classes, so changing representatives on null sets does not change either side.

Proof. First, for each $g\in L^q(\mu)$, define $\Lambda_g(h)=\int_X hg\,d\mu$ for $h\in L^p(\mu)$. Hölder’s inequality gives $|\Lambda_g(h)|\le\|h\|_p\|g\|_q$ when $1<p<\infty$; for $p=1$, the corresponding bound is $|\Lambda_g(h)|\le\|h\|_1\|g\|_\infty$. Hence the integral exists absolutely and $\Lambda_g$ is a bounded linear functional. If $f_n\rightharpoonup f$, applying the definition of weak convergence to $\Lambda_g$ gives the asserted integral convergence for every $g\in L^q(\mu)$.

For the converse, use the standard $L^p$ duality theorem: on a sigma-finite measure space, every bounded linear functional on $L^p(\mu)$, for $1\le p<\infty$, is represented by integration against some $g\in L^q(\mu)$; that is, it has the form $h\mapsto\int_X hg\,d\mu$. This is the duality result for $1<p<\infty$ and for $p=1$ established in the earlier material on $L^p$ duality. Suppose the integral convergence in the theorem holds for every $g\in L^q(\mu)$. Given any $\Lambda\in(L^p(\mu))^*$, choose its representing $g$. Then $$ \Lambda(f_n)=\int_X f_ng\,d\mu\longrightarrow\int_X fg\,d\mu=\Lambda(f). $$ This is precisely weak convergence by definition. Both implications are proved. $\square$

The theorem identifies the tests, but it does not say that weak convergence can be decided from a single $g$. Nor does it identify weak convergence with pointwise convergence, convergence in measure, or norm convergence. Those are different notions; the criterion here is convergence of the pairings for every member of the dual testing space.

Worked Example: Computing an $L^p$-$L^q$ Pairing

On $([0,1],\mathcal L,m)$, let $f=2\chi_{[0,1/2]}$ and $g=3\chi_{[0,1/2]}$. Compute their pairing and verify the Hölder bound for every $1\le p<\infty$.

The product is $fg=6$ on $[0,1/2]$ and $0$ on its complement, so $$ \int_{[0,1]}fg\,dm=6\,m([0,1/2])=6\cdot\frac12=3. $$ For $1<p<\infty$, $q$ satisfies $1/p+1/q=1$, and direct integration gives $$ \|f\|_p=2\left(\frac12\right)^{1/p}, \qquad \|g\|_q=3\left(\frac12\right)^{1/q}. $$ Their product is $6(1/2)^{1/p+1/q}=6(1/2)=3$, equal to the pairing. At $p=1$, $\|f\|_1=2(1/2)=1$ and $\|g\|_\infty=3$, so $\|f\|_1\|g\|_\infty=3$, again equal to the pairing. Check: in each case the computed integral is $3$, and it meets the relevant Hölder bound with equality.

Worked Example: Norm Convergence Gives Every Integral Test

On $([0,1],\mathcal L,m)$, fix $1\le p<\infty$ and define $f_n=(1+1/n)\chi_{[0,1]}$ and $f=\chi_{[0,1]}$. Verify weak convergence by the $L^q$ tests.

The difference is $f_n-f=(1/n)\chi_{[0,1]}$. Since $m([0,1])=1$, $$ \|f_n-f\|_p=\left(\int_{[0,1]}\left|\frac1n\right|^p\,dm\right)^{1/p} =\left(\frac1{n^p}\cdot1\right)^{1/p}=\frac1n\longrightarrow0. $$ For $1<p<\infty$ and any $g\in L^q([0,1])$, Hölder’s inequality yields $$ \left|\int_{[0,1]}(f_n-f)g\,dm\right| \le\|f_n-f\|_p\|g\|_q=\frac{\|g\|_q}{n}\longrightarrow0. $$ For $p=1$ the same conclusion follows from $\left|\int(f_n-f)g\,dm\right|\le\|f_n-f\|_1\|g\|_\infty=\|g\|_\infty/n$. Thus every $L^q$ test converges, and the characterization gives $f_n\rightharpoonup f$. Check: using $g=\chi_{[0,1]}$, the pairing difference is $\int(f_n-f)g\,dm=1/n$, which tends to zero as predicted.

A Common Mistake: Testing Only the Constant Function

Worked Example: One Successful Test Is Not Enough

On $([0,1],\mathcal L,m)$, define $u=\chi_{[0,1/2]}-\chi_{(1/2,1]}$ and take the constant sequence $f_n=u$. The two intervals have equal measure $1/2$, and the value at the single endpoint does not affect the integrals.

A tempting but incorrect argument is: “$\int f_n\,dm=0$ for every $n$, so $f_n\rightharpoonup0$.” This checks only the test $g=\chi_{[0,1]}$, because $$ \int_{[0,1]}f_n\,dm=\frac12-\frac12=0. $$ But $u\in L^q([0,1])$ for every $1\le q\le\infty$, and testing with $g=u$ gives $$ \int_{[0,1]}f_nu\,dm =\int_{[0,1]}u^2\,dm =\frac12+\frac12=1 $$ for every $n$. The required limit for weak convergence to $0$ would be $\int 0\cdot u\,dm=0$, so this test fails. Correction: weak convergence to $0$ requires convergence to zero against every $g\in L^q$, not merely against the constant function. Here the sequence is constant and nonzero, and the test $u$ detects it.

This pitfall reflects the structure of the weak topology. A basic neighborhood imposes only finitely many tests at a time, but convergence must eventually satisfy every such neighborhood, and the definition quantifies over the entire dual. Choosing tests suited to a particular problem can be useful; it does not replace the universal condition unless a separate argument shows that the chosen tests suffice.

Takeaway: In a Banach space, weak convergence means convergence under every bounded linear functional. For $L^p(\mu)$ with sigma-finite $(X,\mathcal M,\mu)$ and $1\le p<\infty$, the duality theorem turns those functionals into exactly the integral tests $h\mapsto\int_X hg\,d\mu$ with $g\in L^q(\mu)$.

Check Your Understanding

For each question, identify the space and the full family of tests required by the weak-convergence definition.

State the definition of weak convergence for a sequence in a real Banach space $B$.
For $p=1$, what is the conjugate testing space, and which norm is used on it?
On a sigma-finite measure space, explain why each $g\in L^q(\mu)$ defines a bounded functional on $L^p(\mu)$.
Why does convergence of $\int_X f_ng\,d\mu$ for one fixed $g$ not establish weak convergence?
If $\|f_n-f\|_p\to0$, how does Hölder’s inequality show convergence of the pairings for every $g\in L^q(\mu)$, including the case $p=1$?

Weak Topology and Functional Testing

What You'll Learn

Weak Convergence Means Passing Every Functional Test

The \(L^p\)-\(L^q\) Testing Characterization

Worked Example: Computing an \(L^p\)-\(L^q\) Pairing

Worked Example: Norm Convergence Gives Every Integral Test

A Common Mistake: Testing Only the Constant Function

Worked Example: One Successful Test Is Not Enough

Check Your Understanding

Weak Topology and Functional Testing

What You'll Learn

Weak Convergence Means Passing Every Functional Test

The \(L^p\)-\(L^q\) Testing Characterization

Worked Example: Computing an \(L^p\)-\(L^q\) Pairing

Worked Example: Norm Convergence Gives Every Integral Test

A Common Mistake: Testing Only the Constant Function

Worked Example: One Successful Test Is Not Enough

Check Your Understanding

Previous

Next