## Add and subtract something nearby that is simpler

### Quick description

Suppose one is trying to estimate an expression involving a complicated function (e.g. something like ). But, one knows or believes that is somehow "close" (at least "on average") to a simpler quantity . Then it can often become advantageous to add and subtract from , i.e. to substitute In many cases the term involving is the "main term", and one can take advantage of the simpler structure of to continue estimating this portion of the expression. Meanwhile, the term is often an "error term"; either is already small, in which case one can hope that the total contribution of this term to the expression one wants to estimate is already small, or exhibits some sort of cancellation which will also make the final contribution to the original expression small (e.g. by the trick "use integration by parts to exploit cancellation").

For more complicated expressions (e.g. bilinear, multilinear, or nonlinear expressions) it is often useful to give the error term its own name, e.g. . Then , and any multilinear expression involving one or more copies of will split into a "main term" involving all 's, plus lots of "error terms" involving one or more 's. Often, one treats the error terms by relatively crude upper bound estimates, but works carefully to estimate the main term as accurately as possible.

Typical examples of choices of include

• the value of at a point nearby to • the average value of on some suitable set ;

• the conditional expectation of with respect to some -algebra .

• Some sort of regularization, discretization, or other approximation to (e.g. one could convolve with an approximation to the identity).

### Prerequisites Incomplete This article is incomplete. More examples wanted.

### Example 1

A classic " " example: show that if a sequence of continuous functions converges uniformly to a limit , then is also continuous.

To prove this, pick an and . The task is to show that if is sufficiently close to , then .

But by hypothesis, we expect to be close to , and to be close to , for large. Adding and subtracting these terms, and using the triangle inequality, we are led to the bound Because of the uniform convergence, we know that for large enough (independent of or - this is important!), we can ensure that and . Once we pick such an , we can then use the fact that is continuous to conclude that for close enough to , and the claim follows.

### Example 2

Let for some , and let be a sequence of approximations to the identity (thus , and for all . Show that the convolutions converge in to .

(Supply proof here)

### Example 3

(Calderon-Zygmund theory)

### Example 4

(Roth's argument for three-term APs)

### General discussion

In some cases, particularly those involving integration by parts or substitution, one wishes to use multiplying and dividing instead of adding and subtracting. For instance, given an integral involving an expression , one may wish to multiply and divide by in order to set up either an integration by parts, or a substitution .

A more advanced version of this technique is generic chaining.

"either is already small, in which case one can hope that the total contribution of this term is already small" sounds a bit strange to me! :D