BMO and the John-Nirenberg Inequality

Intuition Beginner

Imagine measuring the price of one product across thousands of shops. Two kinds of "spread" are possible. One is that prices everywhere stay within a fixed band — never below a floor, never above a ceiling. The other, looser kind says nothing about a ceiling: prices can climb arbitrarily high in a few places, but if you zoom into any neighbourhood and compare each shop to the local average, the typical gap from that local average stays modest. The second condition controls how much things wiggle locally, not how big they get globally.

Functions of bounded mean oscillation are exactly the second kind. For each box you compute the average of the function over that box, then measure how far the function sits from its own average, on average, inside the box. If that averaged gap is uniformly small across every box and every size, the function has bounded mean oscillation. A bounded function automatically qualifies, because it cannot stray far from any average. But the reverse fails: a function can have bounded mean oscillation while shooting off to infinity.

The headline example is the logarithm of distance to a point. Near that point the function plunges to negative infinity, so it is nowhere bounded. Yet over any box the logarithm's gap from its box-average stays controlled, because the logarithm grows so slowly that rescaling a box barely changes its shape. This single example shows the new class is genuinely bigger than the bounded functions, and it is the reason the class matters: it is the natural home for outputs of operations that just barely fail to keep bounded inputs bounded.

The one-sentence takeaway: bounded mean oscillation asks each box's typical deviation from its own average to be small, which lets functions blow up slowly like a logarithm while still being tightly controlled at every scale.

Visual Beginner

Picture the graph of a function over a single box, with a horizontal line drawn at the height equal to the function's average over that box. The mean oscillation is the average vertical distance between the graph and that horizontal line. A nearly flat graph hugs its average line, so its mean oscillation is tiny. A graph that swings wildly above and below the line has large mean oscillation. The defining demand is that no box, of any size or position, produces a large average gap.

The lower panel carries the punchline. Even though the logarithm dives to negative infinity at the marked point, when you draw any box around the region and compare the curve to its box-average, the typical gap stays the same modest size no matter how small the box gets. Boundedness of the function is a statement about the curve's height; bounded mean oscillation is a statement about the curve's spread around its sliding average. The second can hold while the first fails.

Worked example Beginner

We measure the mean oscillation of a simple step function over a box and read off the number by hand.

Step 1. Work on the line and let $f$ equal $0$ on the left half of the interval from $0$ to $2$, and equal $4$ on the right half, that is $f=0$ on $[0,1)$ and $f=4$ on $[1,2]$. Take the box to be the whole interval $Q=[0,2]$, of length $2$.

Step 2. Compute the average of $f$ over $Q$. The function is $0$ over a length $1$ and $4$ over a length $1$, so the total area is $0\times 1+4\times 1=4$, and the average is the area divided by the length, $4$ divided by $2$, which is $2$.

Step 3. Compute how far $f$ sits from this average of $2$. On the left half $f=0$, so the gap is $|0-2|=2$. On the right half $f=4$, so the gap is $|4-2|=2$. The gap is $2$ everywhere.

Step 4. Average the gap over the box. Since the gap is the constant $2$ across the whole interval, its average is just $2$. So the mean oscillation of $f$ over this box is $2$.

What this tells us: the mean oscillation measures the typical distance from the local average, and here it came out to exactly half the jump height, because the function spends equal time on each side of its average. A function of bounded mean oscillation is one where this number stays below a fixed ceiling for every box at once; raising the jump height would raise the oscillation, while shrinking the box around a single level would lower it toward zero.

Check your understanding Beginner

Formal definition Intermediate+

Throughout, $Q$ ranges over cubes in ${\mathbb{R}}^n$ with sides parallel to the axes (equivalently, over balls; the two families give comparable seminorms), $|Q|$ is the Lebesgue measure of $Q$, and $f_Q={\text{fint}}_{Q}f=\frac{1}{|Q|}{\int }_{Q}f$ is the average of $f$ over $Q$. All functions are real- or complex-valued and locally integrable.

Definition (bounded mean oscillation). For $f\in L_{\mathrm{loc}}^1({\mathbb{R}}^n)$ the mean oscillation seminorm is $$\Vert f{\Vert }_{\ast }=\underset{Q}{\sup }\frac{1}{|Q|}{\int }_Q|f(x)-f_Q|dx,$$ the supremum taken over all cubes $Q$. The space of bounded mean oscillation is $$\mathrm{BMO}({\mathbb{R}}^n)=\{f\in L_{\mathrm{loc}}^1({\mathbb{R}}^n):\Vert f{\Vert }_{\ast }<\infty \}.$$ Since $\Vert f{\Vert }_{\ast }=0$ if and only if $f$ is almost everywhere constant, $\Vert \cdot {\Vert }_{\ast }$ is a seminorm whose null space is the constants; $\mathrm{BMO}$ is a Banach space once functions are identified modulo additive constants. The competitor centring ${\inf }_c{\text{fint}}_Q|f-c|$ is comparable: ${\inf }_c{\text{fint}}_Q|f-c|\le {\text{fint}}_Q|f-f_Q|\le 2{\inf }_c{\text{fint}}_Q|f-c|$, so replacing $f_Q$ by the best constant changes $\Vert f{\Vert }_{\ast }$ by at most a factor $2$.

Definition (sharp maximal function). The Fefferman-Stein sharp maximal function of $f$ is $$M^{\#}f(x)=\underset{Q\ni x}{\sup }\frac{1}{|Q|}{\int }_Q|f(y)-f_Q|dy,$$ the supremum over cubes containing $x$. By definition $\Vert f{\Vert }_{\ast }=\Vert M^{\#}f{\Vert }_{L^{\infty }}$, so $f\in \mathrm{BMO}$ exactly when $M^{\#}f\in L^{\infty }$, the BMO seminorm being the $L^{\infty }$-norm of the pointwise oscillation. One has $M^{\#}f\le 2Mf$ pointwise, where $M$ is the Hardy-Littlewood maximal operator 02.19.01.

Non-examples and the canonical example. The inclusion $L^{\infty }\subseteq \mathrm{BMO}$ holds with $\Vert f{\Vert }_{\ast }\le 2\Vert f{\Vert }_{\infty }$, and it is strict: the function $f(x)=\log |x|$ on ${\mathbb{R}}^n$ is unbounded yet lies in $\mathrm{BMO}$, with $\Vert \log |x|{\Vert }_{\ast }$ bounded by a dimensional constant. The function $\mathrm{sgn}(\log |x|)$, by contrast, is bounded — hence in $\mathrm{BMO}$ — even though it is the composition of a bounded function with a $\mathrm{BMO}$ function, illustrating that $\mathrm{BMO}$ is not preserved by all bounded nonlinearities. The point of the class is captured by the slogan: $\mathrm{BMO}$ is the smallest rearrangement-stable space containing $L^{\infty }$ on which Calderón-Zygmund operators remain bounded.

Counterexamples to common slips Intermediate+

• BMO is a seminorm space, not a norm space, until you quotient by constants. Adding a constant to $f$ leaves $\Vert f{\Vert }_{\ast }$ unchanged because $f-f_Q$ is insensitive to constants. Treating $\Vert \cdot {\Vert }_{\ast }$ as a norm on functions (rather than on cosets) gives a non-Hausdorff "space"; the genuine Banach space is $L_{\mathrm{loc}}^1/\mathbb{C}$ with $\Vert \cdot {\Vert }_{\ast }$.

• $\log |x|\in \mathrm{BMO}$ but $|\log |x||\notin \mathrm{BMO}$ is false — both are in BMO, yet $\mathrm{sgn}(x)\log |x|$ can leave it. The subtle fact is that BMO is not a lattice: $f\in \mathrm{BMO}$ does not force $|f|\in \mathrm{BMO}$ in general, though for $\log |x|$ it happens to hold. Pointwise nonlinear operations are not BMO-bounded in general.

• The supremum is over all cubes, not a fixed scale. A function may have small oscillation over large cubes yet large oscillation over small ones, or vice versa. Checking one scale never certifies BMO; the defining supremum couples all positions and all sizes.

• Cubes versus balls and centred versus best-constant change only the constant. Each admissible variant of the definition — cubes or balls, $f_Q$ or the infimising constant, $L^1$ or (after John-Nirenberg) $L^p$ averaging — yields a seminorm comparable to $\Vert \cdot {\Vert }_{\ast }$ up to dimensional factors. The space $\mathrm{BMO}$ is the same; only the numerical seminorm shifts.

Key theorem with proof Intermediate+

Theorem (John-Nirenberg inequality; John-Nirenberg 1961 Comm. Pure Appl. Math. 14, 415). There are dimensional constants $C_1,C_2>0$ such that for every $f\in \mathrm{BMO}({\mathbb{R}}^n)$, every cube $Q$, and every $\lambda >0$, $$|\{x\in Q:|f(x)-f_Q|>\lambda \}|\le C_1|Q|\exp (-\frac{C_2\lambda }{\Vert f{\Vert }_{\ast }}).$$ One may take $C_1=e$ and $C_2=(2^{n}e{)}^{-1}$ (with the dyadic-cube normalization below).

Proof. By homogeneity (replace $f$ by $f/\Vert f{\Vert }_{\ast }$) and translation/dilation invariance of both sides, it suffices to assume $\Vert f{\Vert }_{\ast }=1$ and prove $$\Phi (\lambda ):=\underset{Q}{\sup }\frac{1}{|Q|}|\{x\in Q:|f-f_Q|>\lambda \}|\le C_1{e}^{-C_2\lambda }.$$

Step 1 (one Calderón-Zygmund step). Fix a cube $Q$ and set $g=(f-f_Q){\chi }_Q$, so ${\text{fint}}_Q|g|={\text{fint}}_Q|f-f_Q|\le \Vert f{\Vert }_{\ast }=1$. Apply the dyadic Calderón-Zygmund decomposition 02.19.02 to $|g|$ on $Q$ at height $\alpha =2^n$: subdividing $Q$ dyadically and stopping at the maximal subcubes $\{Q_j\}$ where ${\text{fint}}_{Q_j}|g|>2^n$, one obtains pairwise disjoint dyadic subcubes with $$2^n<{\text{fint}}_{Q_j}|g|\le 2^{2n},\sum _j|Q_j|\le 2^{-n}{\int }_Q|g|\le 2^{-n}|Q|,$$ the first from the unselected dyadic parent and the second from the stopping inequality $|Q_j|<2^{-n}{\int }_{Q_j}|g|$ summed over the disjoint family. Off ${\bigcup }_j{Q}_j$ the Lebesgue differentiation theorem 02.19.01 gives $|g|\le 2^n$ a.e.

Step 2 (the averages barely move). On each selected cube $Q_j$, $$|f_{Q_j}-f_Q|=|{\text{fint}}_{Q_j}(f-f_Q)|\le {\text{fint}}_{Q_j}|g|\le 2^{2n}.$$ Thus passing from $Q$ to a child $Q_j$ shifts the centring constant by at most $2^{2n}$. Consequently, for $x\in Q_j$, $$|f(x)-f_Q|\le |f(x)-f_{Q_j}|+|f_{Q_j}-f_Q|\le |f(x)-f_{Q_j}|+2^{2n}.$$

Step 3 (the recursion). The set where $|f-f_Q|>2^{2n}$ inside $Q$ is, up to a null set, contained in ${\bigcup }_j{Q}_j$: off the selected cubes $|g|=|f-f_Q|\le 2^n\le 2^{2n}$ a.e. Hence for any $\lambda >0$, $$|\{x\in Q:|f-f_Q|>\lambda +2^{2n}\}|\le \sum _j|\{x\in Q_j:|f-f_{Q_j}|>\lambda \}|\le \sum _j\Phi (\lambda )|Q_j|\le 2^{-n}\Phi (\lambda )|Q|,$$ using Step 2 to absorb the constant shift, the definition of $\Phi$ on each $Q_j$, and the measure bound ${\sum }_j|Q_j|\le 2^{-n}|Q|$. Taking the supremum over $Q$ yields the functional inequality $$\Phi (\lambda +b)\le 2^{-n}\Phi (\lambda ),b:=2^{2n}.$$

Step 4 (iteration to exponential decay). Since $\Phi \le 1$ always (it is the relative measure of a subset of $Q$), iterating $\Phi (\lambda +kb)\le 2^{-nk}$ for every integer $k\ge 0$. For arbitrary $\lambda \ge 0$ write $\lambda =kb+s$ with $k=\lfloor \lambda /b\rfloor$ and $0\le s<b$; monotonicity of $\Phi$ gives $$\Phi (\lambda )\le \Phi (kb)\le 2^{-nk}=2^{-n\lfloor \lambda /b\rfloor }\le 2^n{2}^{-n\lambda /b}=2^n\exp (-\frac{n\log 2}{2^{2n}}\lambda ).$$ This is the claimed bound with $C_1=2^n$ and $C_2=n{2}^{-2n}\log 2>0$, after restoring $\Vert f{\Vert }_{\ast }$ by the initial normalization. $\square$

Bridge. The John-Nirenberg inequality builds toward the identification of $\mathrm{BMO}$ as the correct $p\to \infty$ endpoint of the $L^p$ scale, and it appears again in the Fefferman-Stein sharp-maximal theory and the $H^1$-$\mathrm{BMO}$ duality that anchor the rest of this chapter. The foundational reason the proof works is that the Calderón-Zygmund decomposition 02.19.02 is self-similar: one stopping-time step at height $2^n$ reproduces the same distributional question one scale down with a fixed measure contraction $2^{-n}$ and a fixed constant shift $2^{2n}$, and this is exactly the structure that turns a single linear oscillation bound into geometric — hence exponential — decay. The central insight is that bounded mean oscillation, an $L^1$ averaged condition, self-improves to exponential integrability: putting these together, $\Vert f{\Vert }_{\ast }<\infty$ forces ${\text{fint}}_Q{e}^{c|f-f_Q|}\le C$ for small $c$, so BMO functions are locally in every $L^p$ and in the Orlicz space $e^L$. The bridge is that this is dual to the atomic $H^1$ estimate, the $e^L$-integrability of $\mathrm{BMO}$ generalises the plain $L^{\infty }$ bound, and the recursion here is exactly the iterated stopping-time selection flagged in the Calderón-Zygmund decomposition's own forward map.

Exercises Intermediate+

Advanced results Master

Theorem 1 (John-Nirenberg, exponential form and self-improvement; John-Nirenberg 1961 Comm. Pure Appl. Math. 14, 415). A locally integrable $f$ lies in $\mathrm{BMO}$ if and only if there exist constants $c,C>0$ with ${\text{fint}}_Q\exp (c|f-f_Q|/\Vert f{\Vert }_{\ast })\le C$ for all cubes $Q$. Consequently the seminorms $\Vert f{\Vert }_{\ast ,p}$ for $1\le p<\infty$ are all comparable, and the BMO condition stated with $L^1$-oscillation self-improves to $L^p$- and to exponential-oscillation with no enlargement of the space. The constant $c$ is the reciprocal of the John-Nirenberg decay rate $C_2$ ^{[John-Nirenberg 1961]}.

Theorem 2 (Fefferman-Stein sharp maximal theorem; Fefferman-Stein 1972 Acta Math. 129, 137). For $1<p<\infty$ and $f$ in a suitable dense class, $\Vert f{\Vert }_{L^p}\approx \Vert M^{\#}f{\Vert }_{L^p}$ with constants depending only on $n,p$. The sharp maximal function thereby characterizes $L^p$ through pure mean-oscillation data, and $\mathrm{BMO}=\{M^{\#}f\in L^{\infty }\}$ is the $p\to \infty$ endpoint of the family $\{M^{\#}f\in L^p\}$. This is the precise sense in which $\mathrm{BMO}$ replaces $L^{\infty }$ at the top of the scale ^{[Fefferman-Stein 1972]}.

Theorem 3 (Calderón-Zygmund operators on the endpoints). Every Calderón-Zygmund operator $T$ maps $L^{\infty }\to \mathrm{BMO}$ boundedly, and by the duality of Theorem 4 its (formal) adjoint maps $H^1\to L^1$ boundedly. The pair $(L^{\infty }\to \mathrm{BMO},H^1\to L^1)$ is the correct endpoint replacement for the failed $L^{\infty }\to L^{\infty }$ and $L^1\to L^1$ bounds, and interpolating either with the $L^2$ bound recovers the full $L^p$ range $1<p<\infty$ ^{[Stein 1993]}.

Theorem 4 (Fefferman duality $(H^1)^ = \mathrm{BMO}$; Fefferman 1971 Bull. AMS 77, 587; Fefferman-Stein 1972 Acta Math. 129, 137).* The dual of the real Hardy space $H^1({\mathbb{R}}^n)$ is $\mathrm{BMO}({\mathbb{R}}^n)$: every $\phi \in \mathrm{BMO}$ defines a bounded linear functional on $H^1$ by $⟨f,\phi ⟩=\int f\phi$ (suitably interpreted on atoms), with $\Vert \phi {\Vert }_{(H^1{)}^{\ast }}\approx \Vert \phi {\Vert }_{\ast }$, and every bounded functional on $H^1$ arises this way. The pairing is well defined because the atomic structure of $H^1$ (mean-zero, $L^2$-normalized bumps supported on cubes) is exactly matched by the John-Nirenberg integrability of BMO, so that $\int a\phi =\int a(\phi -{\phi }_Q)$ converges absolutely by Cauchy-Schwarz ^{[Fefferman-Stein 1972]}.

Theorem 5 (Carleson-measure characterization). A function $\phi \in L_{\mathrm{loc}}^2$ lies in $\mathrm{BMO}$ if and only if $|\nabla u(x,t){|}^{2}tdxdt$ is a Carleson measure on the upper half-space, where $u(x,t)=(P_t\ast \phi )(x)$ is the Poisson extension: ${\sup }_Q\frac{1}{|Q|}{\int }_0^{\ell (Q)}{\int }_Q|\nabla u{|}^{2}tdxdt<\infty$. This analytic characterization, equivalent to membership in $\mathrm{BMO}$ up to constants, is the route by which $\mathrm{BMO}$ enters the theory of the $\bar{\partial }$-equation and the corona problem, and it is the form in which Fefferman first proved the duality ^{[Fefferman 1971]}.

Theorem 6 (commutators; Coifman-Rochberg-Weiss). For a Calderón-Zygmund operator $T$ and $b\in \mathrm{BMO}$, the commutator $[b,T]f=bTf-T(bf)$ is bounded on $L^p$ for $1<p<\infty$, with $\Vert [b,T]{\Vert }_{L^p\to L^p}\le C\Vert b{\Vert }_{\ast }$; conversely, for $T$ a Riesz transform, boundedness of $[b,T]$ on some $L^p$ forces $b\in \mathrm{BMO}$. The commutator thereby gives an operator-theoretic characterization of $\mathrm{BMO}$, and the proof runs through the John-Nirenberg exponential integrability to control the product $b\cdot Tf$ ^{[Stein 1993]}.

Synthesis. The space $\mathrm{BMO}$ is the foundational reason that the failure of Calderón-Zygmund operators to preserve $L^{\infty }$ and $L^1$ is a feature: $\mathrm{BMO}$ and its predual $H^1$ are the correct endpoint spaces, and this is exactly what the four characterizations — mean-oscillation, sharp maximal, Carleson measure, commutator — assert from four directions. The central insight is the self-improvement engine of John-Nirenberg: a single $L^1$-averaged oscillation bound, fed through the self-similar stopping-time recursion of 02.19.02, upgrades to exponential integrability, and this is exactly what makes the pairing $\int a\phi$ converge against the $L^2$-normalized atoms of $H^1$. Putting these together, the picture generalises in three directions: vertically, $\mathrm{BMO}$ caps the $L^p$ scale as the $p\to \infty$ endpoint via the Fefferman-Stein sharp maximal theorem (Theorem 2), bracketing the whole scale $1\le p\le \infty$ by the dual pair $(H^1,\mathrm{BMO})$; horizontally, the Carleson-measure characterization (Theorem 5) ports the cube-oscillation picture to the half-space and is dual to the tent-space theory; structurally, the commutator characterization (Theorem 6) reveals $\mathrm{BMO}$ as the symbol class for which multiplication interacts boundedly with singular integration. The bridge unifying them is the John-Nirenberg inequality: each equivalence restates that bounded mean oscillation is secretly exponential control, dual atom by atom to the cancellation built into $H^1$.

Full proof set Master

Proposition 1 ($\Vert \cdot {\Vert }_{\ast }$ is a seminorm with null space the constants). $\Vert \cdot {\Vert }_{\ast }$ is subadditive and absolutely homogeneous, and $\Vert f{\Vert }_{\ast }=0$ if and only if $f$ is a.e. equal to a constant.

Proof. Absolute homogeneity is clear from $|(cf)-(cf{)}_Q|=|c||f-f_Q|$. For subadditivity, $(f+g{)}_Q=f_Q+g_Q$, so $|(f+g)-(f+g{)}_Q|\le |f-f_Q|+|g-g_Q|$ pointwise; averaging and taking the supremum over $Q$ gives $\Vert f+g{\Vert }_{\ast }\le \Vert f{\Vert }_{\ast }+\Vert g{\Vert }_{\ast }$. If $f\equiv c$ then $f_Q=c$ and the oscillation vanishes, so $\Vert f{\Vert }_{\ast }=0$. Conversely if $\Vert f{\Vert }_{\ast }=0$ then ${\text{fint}}_Q|f-f_Q|=0$ for every cube, so $f=f_Q$ a.e. on each $Q$; since the constants $f_Q$ must agree on overlapping cubes, $f$ is a.e. a single constant. $\square$

Proposition 2 ($\mathrm{BMO}$ is complete modulo constants). The quotient $\mathrm{BMO}/\mathbb{C}$ with seminorm $\Vert \cdot {\Vert }_{\ast }$ is a Banach space.

Proof. Let $(f_k)$ be Cauchy in $\Vert \cdot {\Vert }_{\ast }$; normalize representatives so that $(f_k{)}_{Q_0}=0$ for a fixed unit cube $Q_0$. For any fixed cube $Q$, the averages ${\text{fint}}_Q|f_j-f_k|\le {\text{fint}}_Q|(f_j-f_k)-(f_j-f_k{)}_Q|+|(f_j-f_k{)}_Q|$; the first term is $\le \Vert f_j-f_k{\Vert }_{\ast }$, and chaining $Q$ to $Q_0$ through finitely many overlapping cubes bounds $|(f_j-f_k{)}_Q|$ by $C_Q\Vert f_j-f_k{\Vert }_{\ast }$. Hence $(f_k)$ is Cauchy in $L^1(Q)$ for every $Q$, so it converges in $L_{\mathrm{loc}}^1$ to some $f$. Lower semicontinuity of ${\text{fint}}_Q|\cdot -(\cdot {)}_Q|$ under $L^1(Q)$-convergence gives $\Vert f-f_k{\Vert }_{\ast }\le {\mathrm{liminf}}_m\Vert f_m-f_k{\Vert }_{\ast }\to 0$, so $f\in \mathrm{BMO}$ and $f_k\to f$. $\square$

Proposition 3 (one Calderón-Zygmund step recursion, restated). With $\Phi (\lambda )={\sup }_Q\frac{1}{|Q|}|\{x\in Q:|f-f_Q|>\lambda \}|$ and $\Vert f{\Vert }_{\ast }=1$, one has $\Phi (\lambda +2^{2n})\le 2^{-n}\Phi (\lambda )$ for all $\lambda \ge 0$.

Proof. This is Steps 1-3 of the Key Theorem. The decomposition of $g=(f-f_Q){\chi }_Q$ at height $2^n$ produces disjoint dyadic subcubes $\{Q_j\}$ with ${\sum }_j|Q_j|\le 2^{-n}|Q|$ and $|f_{Q_j}-f_Q|\le 2^{2n}$, off which $|f-f_Q|\le 2^n$ a.e. Hence $\{x\in Q:|f-f_Q|>\lambda +2^{2n}\}\subseteq {\bigcup }_j\{x\in Q_j:|f-f_{Q_j}|>\lambda \}$ up to null sets, and summing the per-cube bound $\Phi (\lambda )|Q_j|$ yields the contraction. $\square$

Proposition 4 (exponential decay from the recursion). $\Phi (\lambda )\le 2^n{e}^{-C_2\lambda }$ with $C_2=n{2}^{-2n}\log 2$.

Proof. From $\Phi \le 1$ and Proposition 3, induction gives $\Phi (k{2}^{2n})\le 2^{-nk}$. For general $\lambda$, write $k=\lfloor \lambda /2^{2n}\rfloor$ and use monotonicity: $\Phi (\lambda )\le \Phi (k{2}^{2n})\le 2^{-nk}\le 2^n{2}^{-n\lambda /2^{2n}}=2^n{e}^{-C_2\lambda }$. $\square$

Proposition 5 ($\log |x|\in \mathrm{BMO}$, sharpness). $\log |x|\in \mathrm{BMO}({\mathbb{R}}^n)$, while $|x{|}^{\epsilon }\notin \mathrm{BMO}$ for any $\epsilon >0$.

Proof. Membership is Exercise 3. For sharpness, test $f(x)=|x{|}^{\epsilon }$ on cubes $Q_R$ of side $R$ at the origin: $f_{Q_R}\sim R^{\epsilon }$ and ${\text{fint}}_{Q_R}|f-f_{Q_R}|\sim R^{\epsilon }\to \infty$ as $R\to \infty$, so $\Vert f{\Vert }_{\ast }=\infty$. The dilation $f(Rx)=R^{\epsilon }f(x)$ scales the oscillation by $R^{\epsilon }$, which is bounded over scales only when $\epsilon =0$; the logarithm sits exactly at $\epsilon =0$, where the dilation acts by an additive constant the seminorm ignores. $\square$

Proposition 6 (sharp maximal pointwise bounds). $M^{\#}f\le 2Mf$ pointwise, and $\Vert f{\Vert }_{\ast }=\Vert M^{\#}f{\Vert }_{L^{\infty }}$.

Proof. For a cube $Q\ni x$, ${\text{fint}}_Q|f-f_Q|\le {\text{fint}}_Q|f|+|f_Q|\le 2{\text{fint}}_Q|f|\le 2Mf(x)$ after replacing the cube by a comparable centred ball; taking the supremum over $Q\ni x$ gives $M^{\#}f(x)\le 2Mf(x)$. The identity $\Vert f{\Vert }_{\ast }=\Vert M^{\#}f{\Vert }_{\infty }$ is immediate: $\Vert M^{\#}f{\Vert }_{\infty }={\mathrm{esssup}}_x{\sup }_{Q\ni x}{\text{fint}}_Q|f-f_Q|={\sup }_Q{\text{fint}}_Q|f-f_Q|=\Vert f{\Vert }_{\ast }$, since every cube contains some point and the essential supremum over $x$ of a quantity depending only on the cubes through $x$ recovers the supremum over all cubes. $\square$

Connections Master

• The Calderón-Zygmund decomposition 02.19.02. The direct prerequisite and the engine of the John-Nirenberg proof. One stopping-time step at height $2^n$ on $(f-f_Q){\chi }_Q$ produces the disjoint subcubes and the measure contraction $2^{-n}$ that make the distributional recursion self-similar; iterating that single decomposition is exactly what converts bounded mean oscillation into exponential decay, so the John-Nirenberg inequality is the decomposition applied to itself across scales.

• Hardy-Littlewood maximal function and the Vitali covering lemma 02.19.01. The direct prerequisite supplying the sharp maximal function's domination $M^{\#}f\le 2Mf$, the Lebesgue differentiation theorem used to control the good part off the selected cubes, and the pointwise bound $|f|\le Mf$ a.e. that closes the Fefferman-Stein equivalence $\Vert f{\Vert }_{L^p}\approx \Vert M^{\#}f{\Vert }_{L^p}$.

• $L^p$ spaces, Hölder, Minkowski, Riesz-Fischer completeness 02.07.06. The direct prerequisite carrying the layer-cake distribution-function calculus behind the $L^p$-BMO seminorm comparison and the exponential-integrability corollary, the Cauchy-Schwarz estimate in the $L^{\infty }\to \mathrm{BMO}$ proof, and the Marcinkiewicz interpolation that turns the $L^p$-$\mathrm{BMO}$ endpoints into boundedness on the full range.

• Singular integral operators and the Hilbert transform [forward: 02.19.03]. The principal application. Calderón-Zygmund operators fail to preserve $L^{\infty }$, and $\mathrm{BMO}$ is precisely the space into which they map it; the commutator $[b,T]$ is $L^p$-bounded exactly when $b\in \mathrm{BMO}$, making BMO the natural symbol class for singular-integral multiplication.

• Hardy spaces and atomic decomposition [forward: 02.20.02]. The dual partner. The Fefferman duality $(H^1{)}^{\ast }=\mathrm{BMO}$ pairs the mean-zero $L^2$-normalized atoms of $H^1$ against BMO functions; the absolute convergence of the pairing $\int a(\phi -{\phi }_Q)$ is supplied by the John-Nirenberg integrability proved here.

• Littlewood-Paley theory and square functions [forward: 02.20.03]. The analytic face. The Carleson-measure characterization expresses $\Vert \phi {\Vert }_{\ast }$ through the Littlewood-Paley square function of the Poisson extension, embedding BMO into the tent-space and square-function framework that organises the rest of this chapter.

Historical & philosophical context Master

The space $\mathrm{BMO}$ and the inequality bearing their names were introduced by Fritz John and Louis Nirenberg in their 1961 Communications on Pure and Applied Mathematics paper On functions of bounded mean oscillation ^{[John-Nirenberg 1961]}. John had arrived at the mean-oscillation condition through his work on elasticity and quasi-isometric (bi-Lipschitz) deformations, where the logarithm of the Jacobian of such a map is the prototypical bounded-mean-oscillation function; the exponential integrability they proved was the tool needed to control such deformations. Their original argument is the Calderón-Zygmund stopping-time iteration reproduced above, run on the level sets of the oscillation.

The decisive structural role of $\mathrm{BMO}$ in harmonic analysis emerged a decade later. Charles Fefferman announced the duality $(H^1{)}^{\ast }=\mathrm{BMO}$ in his 1971 Bulletin of the American Mathematical Society note Characterizations of bounded mean oscillation ^{[Fefferman 1971]}, identifying the dual of the real Hardy space — until then an analytically awkward object — with the concrete and computable space of mean oscillation, and supplying the Carleson-measure characterization through the Poisson extension. The full theory, including the sharp maximal function and the systematic use of BMO as the $L^{\infty }$-substitute endpoint for singular integrals, was developed by Fefferman and Elias Stein in their 1972 Acta Mathematica paper $H^p$ spaces of several variables ^{[Fefferman-Stein 1972]}. The commutator characterization, completing the operator-theoretic picture, was established by Ronald Coifman, Richard Rochberg, and Guido Weiss in 1976. Within this lineage the John-Nirenberg inequality is the technical keystone: it is the self-improvement that makes mean oscillation strong enough to serve as a dual space and as an interpolation endpoint.

Bibliography Master

@article{JohnNirenberg1961,
  author  = {John, Fritz and Nirenberg, Louis},
  title   = {On functions of bounded mean oscillation},
  journal = {Communications on Pure and Applied Mathematics},
  volume  = {14},
  year    = {1961},
  pages   = {415--426}
}

@article{Fefferman1971,
  author  = {Fefferman, Charles},
  title   = {Characterizations of bounded mean oscillation},
  journal = {Bulletin of the American Mathematical Society},
  volume  = {77},
  year    = {1971},
  pages   = {587--588}
}

@article{FeffermanStein1972,
  author  = {Fefferman, Charles and Stein, Elias M.},
  title   = {$H^p$ spaces of several variables},
  journal = {Acta Mathematica},
  volume  = {129},
  year    = {1972},
  pages   = {137--193}
}

@article{CoifmanRochbergWeiss1976,
  author  = {Coifman, Ronald R. and Rochberg, Richard and Weiss, Guido},
  title   = {Factorization theorems for Hardy spaces in several variables},
  journal = {Annals of Mathematics},
  volume  = {103},
  year    = {1976},
  pages   = {611--635}
}

@book{Stein1993,
  author    = {Stein, Elias M.},
  title     = {Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory Integrals},
  publisher = {Princeton University Press},
  year      = {1993}
}

@book{Grafakos2014Modern,
  author    = {Grafakos, Loukas},
  title     = {Modern Fourier Analysis},
  edition   = {3},
  publisher = {Springer},
  year      = {2014}
}

@book{Duoandikoetxea2001,
  author    = {Duoandikoetxea, Javier},
  title     = {Fourier Analysis},
  publisher = {American Mathematical Society},
  year      = {2001}
}