This is a study guide I made very early this summer, following the lecture notes of the \(T(1)\) theorem that Tao produced for his 247B course. I explain my motivation for creating these notes below.
The \(T(1)\) theorem of G. David and J.-L. Journé famously gives a necessary and sufficient condition for a distribution-valued operator (equivalently, a continuous bilinear form on Schwartz functions1) to be the restriction of a Calderón-Zygmund operator.2 This was an instigating event that marked the beginnings of the third-generation theory.
The classical statement of David & Journé (1984) deals specifically with distribution-valued operators, and so requires inputs which are test functions: the weak boundedness condition is stated in terms of bounded families in \(C^{\infty}_c\), and the desired objects \(T(1)\), \(T^*(1)\) are constructed in a weak sense, by duality acting on elements \(g \in C^{\infty}_c \cap H^1\), \(H^1\) here denoting the real-variable Hardy space.
Then, after a quick bit of calculation, one sees that if \(T\) is, in fact, given by a Calderón-Zygmund operator, by the standard method for how we make sense of \(T\) on \(L^{\infty}\) functions (i.e., first obtaining boundedness for \(T\) and \(T^*\) on all spaces \(L^p\), \(1 < p < \infty\); then obtaining boundedness at the endpoint, \(T, T^* : H^1 \to L^1\), and finally using duality to get \(T^t, (T^*)^t : L^{\infty} \to \text{BMO}\), and noting that \((T^*)^t\) coincides with \(T\) on the class of functions \(f \in L^{\infty}\) vanishing outside a compact set), we see that the objects \(T(1)\), \(T^*(1)\) so obtained in David & Journé in fact coincide with how we might have defined \(T(1)\), \(T^*(1)\) from earlier in the theory.
But this formulation is somewhat technical, and Tao chooses to focus on the case of \(T\) represented by a continuous, bounded kernel, vanishing for widely-separated points. This allows one to apply Schur’s test, and get that \(T\) is a priori bounded on \(L^2\) (indeed, all \(L^p\), \(1 \leq p \leq \infty\)), but the key is to derive estimates independent of the \(B\) from \(\|K\|_{L^{\infty}(\mathbb{R}^d \times \mathbb{R}^d)} \leq B\) and the separation value \(R = \sup_{K(x, y) \neq 0} |x \ – \, y|\).
These assumptions allow him to proceed to defining \(T(1)\), \(T^*(1)\) directly, without the need to consider the \(H^1\)-\(\text{BMO}\) duality in defining the action of \(T\) and \(T^*\) on \(L^{\infty}\) functions (which, though rather elegant once recognized, seems rather unmotivated as a matter of first definitions and impressions if one has not seen it before).3
The proof proceeds by showing that each of \(T(1)\), \(T^*(1)\), and the weak boundedness property suffice to control three types of interactions in the bilinear form \(\langle T(f), g \rangle\): the low-high, high-low, and high-high frequency pairings that result from taking a bilinear form of two Littlewood-Paley decompositions (one of the main function \(f\), and the other of the dualizing function \(g\)).
The high-low and low-high terms are each split into two contributions: an error, which one must carefully and tediously estimate to see that it gives a bounded contribution; and a main term, which actually has the structure of a Carleson measure, thanks to the \(\text{BMO}\) term (which allows us to apply our standard \(L^2\) boundedness machinery).
The high-high terms, meanwhile, are estimated by a careful, escalating sequence of lemmas, which one eventually deploys with a bit of delicate arguing to show that the integral kernel of the interaction is Schur’s test–able.
Most of Tao’s exposition is very clear, but also at a very high level. The heuristics he uses are explained wonderfully, but the careful details of the estimates (which are, after all, what analysis is about) frequently get brushed under the rug; as someone whose intuition for these estimates is not too strong, I thought I’d work out every single claimed estimate in the lecture notes, in complete detail.
For some of the main points where I tried to expand further on Tao’s arguments:
- Lemma 1.4, the estimate on Schwartz functions of mean-zero, requires a bit more writing to develop, but otherwise isn’t too bad; it does require a bit of careful casework and care for the intersections of the various balls and annuli, however.
- Lemma 1.5 was one of the most tricky (i.e., difficult) bits in the notes, as Tao proceeds from an estimate on a certain function, when integrated over small balls of fixed radius but varying center, to a full estimate on the integral of that function multiplied by a certain Schwartz function. His remark — that the second function “is a Schwartz function of height \(J^d\) adapted to the ball \(B(0, J^{- 1})\), which one can cover with \(O_d((K / J)^d)\) balls of radius \(K^{- 1}\),”4 where \(J\), \(K\) are dyadic integers, \(J \leq K\) — is tricky to implement, as the Schwartz function, while “essentially” supported on that ball and decaying rapidly outside of it (as a heuristic), is not truly compactly-supported. There are, after all, Schwartz tails we must consider. As a consequence, to prove this rigorously, we do not simply have a single contribution; it becomes a chore to carefully estimate this integral, decomposing Euclidean space into collections of dyadic annuli, grouping them into good and bad regions, and subdividing the latter class further, finally putting lots of pieces together and seeing which are small, and where.5
- Lemma 1.6 was also exceptionally tedious, mostly because many inequalities and a great deal of casework must be shown; moreover, some basic inequalities on the kernel’s outer-part required significant computation and case-checking. The main task was figuring out how to obtain uniform bounds on products of inverse-power weights with significant differences between their centers and/or width scales. This kind of thing, with the precise scaling behavior and dependence, was hard to see intuitively, and I was only satisfied after I had shown the result in each case, in great detail. Maybe some other people can see what the “right” estimate should be, as a matter of course — but I’m not at that level yet.
Careful verification of these bits comprised most of the bulk of the length, and also explain why my write-up is 30 pages to Tao’s 6 pages in the relevant section of his notes.6
Psychologically, this was one of the most interesting and most difficult pieces of writing I’ve done in mathematics, so far. The gaps and omitted steps in the notes constituted a daunting number of propositions to be filled in. To be fair, I can see now in hindsight that most of these were indeed follow-your-nose type arguments7 (once you had the right decomposition, or saw what cases were outstanding, and had done all the heavy calculating, it then became possible to write-out the details thoroughly).
Still, there is something about the task of independently establishing an inequality in hard analysis, especially one you do not find obvious and have no intuition for, which only serves as a tool to be used to connect other propositions, and with no suggested plan of attack; this can strike one as a daunting task. It reminds me of a lot of the extension exercises from undergrad textbooks (“fill in this argument that was left open in the previous section”), but far, far more open-ended. There isn’t any guarantee that what you’re doing is right — that is how it feels — and sometimes your thoughts drift to: “can I even estimate this?”; “am I actually on track, or even making progress at all?”.
(Or perhaps all of this is just all in my own mind, and other people don’t really have this experience. I’m not sure; a lot of my friends are either not in analysis, or are (broadly speaking) but in very different areas from harmonic analysis.)
There were plenty of false starts in some arguments, and eventually, after I finished the write-up, I carefully read the paper, checking every single line and step, for 3 or 4 complete runs until I was generally convinced of its correctness.
Because this write-up has so much independent calculation and argument, parts of which are so detached from the original source, I still feel somewhat-uneasy with the correctness of the (that is to say, my) arguments. It’s possible there’s still some erroneous step in the expanded proof which I added and never realized. If anyone finds a serious issue, I would be deeply grateful if you would let me know.
Finally, I feel like I should note that this argument, despite the ways I found it technically-intense, also helped clarify the structure of a lot of material that I had learned this year. I initially studied Littlewood-Paley theory as a method for decomposition of functions: one divides frequency space into dyadic scales, and observes the behavior of each smooth localization, before recombining the pieces together. That is, I thought of the Littlewood-Paley decomposition as a procedure whose terms are very much adapted to the “shape” of the input, in frequency space.
But another way to see Littlewood-Paley, more broadly, is as giving the generic form of \(L^p\) (\(1 < p < \infty\)) functions; the representation \(f = \sum_N f_N\) is showing that an \(L^p\) function can be considered a collection of smooth, mean-zero pieces oscillating at widely-separated characteristic scales in frequency.
Then, as one eventually finds in Tao’s proof, the boundedness of the operator \(T\) is determined by how it maps pairs of translated inverse Littlewood-Paley smooth projections, \(\langle T(\tau_y (\mathcal{F}^{- 1} \varphi_N)), \tau_x (\mathcal{F}^{- 1} \varphi_M) \rangle\). What the preceding lemmas amount to verifying is that the hypotheses of the \(T(1)\) theorem give sufficient control to bound the full, summed bilinear form. I find this to be quite elegant; what matters is really the behavior of the operator when tested against a single family of mean-zero Schwartz functions, varying only by translations and (a countable family of) dilations; they are in a sense universal.8
More poetically, I have a mental image of all these functions, hovering and dancing across Euclidean space, ranging in shape from the very sharp and very narrow, to the very low and very wide; going from a low, slow rumble produced by the deepest bass notes and climbing up toward ever-higher frequencies. The operator transforms these tones into something new; and that collection is tested, for resonance and harmony, against all the various elements in our grand scale of tones. To me, the way this argument comes together in the end is deeply satisfying, in a way I almost can’t explain.
Work begun: July 27, 2025. Completed: July 29, 2025.
But
- Which will be the standard collection to which we refer, whenever we talk of test functions. I might get to it in a future post, but in brief, I find the “usual” test function space \(\mathcal{D}\) to be quite horrible to work with; I like to understand the topology of any spaces I handle, and the strict inductive limit of LCTVSs (even LFs) is really exceptionally unpleasant. ↩︎
- As always, we will observe the careful distinction that a Calderón-Zygmund operator is to be a linear operator defined on (a dense subset of) the \(L^2(\mathbb{R}^d)\) functions, with the property that it is \(L^2\)-bounded, and such that off the support of a compactly-supported test function, it is represented pointwise by integration against a singular kernel. ↩︎
- One should note that the focus on this special case marks a similarity to the lecture notes of V. Chousionis & X. Tolsa, on the same subject. However, while Tao and Tolsa end up approaching (roughly) the same \(T(1)\) theorem, they do so in radically-different ways; one by building up to a Littlewood-Paley argument, and another by working solely with dyadic cubes. ↩︎
- The text says “balls of radius \(K\)”, but this is certainly a typo, as the display immediately prior discusses balls of radius \(K^{- 1}\). ↩︎
- At this point, I have to admit that while the bulk of this mathematical translation took me 3 days, most of that time (the first night, which was relatively sleepless, and much of the morning after) was spent figuring out how to fix this. The calculations became increasingly complex until I was surprised to find that the decomposition-style approach actually worked to give an admissible contribution. ↩︎
- I began this project early in the summer, since I wanted to learn about \(T(1)\) theorems and thought that the short page-count of Tao’s notes (only 6 (!)) would mean that I could read and digest it quickly. As it turns out, this is very much not the case, at least if you’re like me and have an extreme (my friends like the word “pathological”) need for detail and clarity. ↩︎
- To use Professor Killip’s trademark phrase for these kinds of things. ↩︎
- I haven’t gotten around to it yet, but at some point, I should spend some time perusing my copies of Meyer & Coifman’s two volumes on wavelets, and check the strategy of their \(T(1)\) theorem proof. ↩︎
Leave a Reply