Two ways to prove equality: Tao's two gates and Noether's inner ground

Two pieces of mathematical advice can sound like opposites.

Terence Tao recommends:

“Split up equalities into inequalities.”

Hermann Weyl reports Emmy Noether objecting that such a proof can miss:

“the inner ground for their equality.”

Which is right?

Both, but they answer different questions.

• Tao’s move is a reliable way to establish an equality.
• The Noether remark asks for a mechanism that explains why the equality is not accidental.
• Sometimes the two-direction proof already exposes that mechanism.
• Sometimes a valid proof certifies the result while hiding the structure that makes it general.

This tutorial begins with small examples, then reaches two famous structural ideas: the Cantor–Schröder–Bernstein theorem and a one-coordinate instance of Noether’s theorem.

This page continues What is 2?, where several proof languages recover the same reversible-distinction shape behind $1+1=2$.

1. Equality has at least three jobs

An equality proof can be judged along three separate axes.

A proof can succeed completely at the first job without succeeding at the other two.

That is not a defect in validity. It is a limit on explanatory reach.

2. Tao’s method: the gate with two locks

Imagine an equality gate with two independent locks.

first lock:   a ≤ b

second lock:  b ≤ a

both open:    a = b

The rule behind the gate is antisymmetry.

In a partially ordered set,

\[a=b \quad\Longleftrightarrow\quad \bigl(a\le b\ \land\ b\le a\bigr).\]

Read exactly:

“$a$ equals $b$ if and only if $a$ is less than or equal to $b$, and $b$ is less than or equal to $a$.”

The reverse direction is the useful one:

\[a\le b \quad\text{and}\quad b\le a \quad\Longrightarrow\quad a=b.\]

This is not a shortcut or approximation. When the order is antisymmetric, the two-lock certificate proves equality.

Rule card

The gate story is faithful only when all three rules hold:

1. The relation $\le$ is reflexive.
2. The relation $\le$ is transitive.
3. The relation $\le$ is antisymmetric.

Together these rules define a partial order.

Break test: a preorder is not enough

Define a relation on words by length:

\[u\preceq v \quad\Longleftrightarrow\quad |u|\le |v|.\]

Then

\[\texttt{cat}\preceq\texttt{dog} \quad\text{and}\quad \texttt{dog}\preceq\texttt{cat},\]

but the two words are not equal.

The relation is a preorder. It records an equivalence in length, not identity of words.

Two opposite comparisons imply equality only when the declared comparison has the required antisymmetry.

3. Noether’s question: what drives both sides?

The second lens asks for a common mechanism.

Typical answers have one of these forms:

same canonical form
same counted objects
same bijection
same invariant
same symmetry
same remainder forced to zero

One general pattern is normalization:

\[A=N(A)=N(B)=B.\]

Read exactly:

“$A$ equals the normal form of $A$, which equals the normal form of $B$, which equals $B$.”

Here the equality signs mean equality in the declared semantics, not character-for-character identity. The chain is valid only if every normalization step preserves the meaning being compared.

For a general decision procedure, more is needed:

• every allowed input must reach a normal form;
• the rewrite system must not give contradictory final forms;
• equality of normal forms must reflect equality in the intended model.

The picture is a shared blueprint press:

expression A
    └──▶ checked normalizer
             └──▶ blueprint C

expression B
    └──▶ checked normalizer
             └──▶ blueprint C

The blueprint is not evidence by appearance alone. Each arrow needs a receipt saying that the transformation preserved the relevant meaning.

4. Small example: two shuffled bags

Consider two finite lists:

\[[3,1,2,1] \qquad\text{and}\qquad [1,3,1,2].\]

They are not equal as sequences because their positions differ.

Treat them instead as multisets, where order is irrelevant but multiplicity matters. Sorting gives

\[\operatorname{sort}([3,1,2,1]) = [1,1,2,3],\]

and

\[\operatorname{sort}([1,3,1,2]) = [1,1,2,3].\]

The common sorted list is a canonical witness of multiset equality.

Why this carries more structure than checking every count separately

A two-direction multiplicity proof could show:

• the first bag contains no more copies of any value than the second;
• the second bag contains no more copies of any value than the first.

That is valid.

The sorted normal form additionally supplies:

• one replayable certificate;
• an immediate equality test for any other finite bag;
• a canonical display of every multiplicity;
• a way to locate the first disagreement when equality fails.

Boundary

Sorting does not prove equality of the original sequences. It proves equality only after the model declares order irrelevant. Changing the equivalence relation changes the theorem.

5. Small example: a set identity from both lenses

Let $A$, $B$, and $C$ be sets. Consider

\[A\setminus(B\cup C) = (A\setminus B)\cap(A\setminus C).\]

The two-gate proof

For the first inclusion, take an element on the left. It lies in $A$ and in neither $B$ nor $C$, so it lies in both $A\setminus B$ and $A\setminus C$.

For the reverse inclusion, take an element on the right. It lies in $A$, not in $B$, and not in $C$, so it is not in $B\cup C$.

Both inclusions hold, so the sets are equal.

The inner ground

Membership on both sides reduces to the same Boolean condition. Unpack it one layer at a time.

Start with

\[x\in A\setminus(B\cup C).\]

This is equivalent to all three statements

\[x\in A, \qquad x\notin B, \qquad x\notin C.\]

Those statements are equivalent to both conditions

\[x\in A\setminus B, \qquad x\in A\setminus C.\]

Therefore

\[x\in(A\setminus B)\cap(A\setminus C).\]

Read exactly:

“$x$ is in $A$ minus the union of $B$ and $C$ exactly when $x$ is in $A$, $x$ is not in $B$, and $x$ is not in $C$. Those same three conditions say that $x$ is in both $A$ minus $B$ and $A$ minus $C$. Therefore $x$ is in their intersection.”

The two inclusions and the shared predicate are not competing proofs here. The elementwise proof exposes the Boolean mechanism while it opens both gates.

6. A picture that predicts a theorem

Now consider the sum of the first $n$ positive odd integers:

\[S_n=\sum_{k=1}^{n}(2k-1).\]

The claim is

\[S_n=n^2.\]

A table of examples can confirm small cases:

$n$	Sum	Result
1	$1$	$1$
2	$1+3$	$4$
3	$1+3+5$	$9$
4	$1+3+5+7$	$16$

The table certifies only the rows checked. The square picture reveals the rule that generates every row.

The exact algebra behind the picture is

\[\begin{aligned} k^2-(k-1)^2 &=\bigl(k-(k-1)\bigr)\bigl(k+(k-1)\bigr)\\ &=1\cdot(2k-1)\\ &=2k-1. \end{aligned}\]

Read exactly:

“$k$ squared minus the square of the quantity $k-1$ equals the quantity $k-(k-1)$ multiplied by the quantity $k+(k-1)$.”

Then read the simplification:

“The first factor equals one. The second factor equals two times $k$ minus one. Their product equals two times $k$ minus one.”

Now add the successive differences:

\[\begin{aligned} \sum_{k=1}^{n}(2k-1) &=\sum_{k=1}^{n}\bigl(k^2-(k-1)^2\bigr)\\ &=n^2-0^2\\ &=n^2. \end{aligned}\]

All middle square terms cancel. This is a telescoping sum.

Generative test

The picture predicts the next layer before it is calculated.

If the old square has side $n-1$ and the new square has side $n$, the added layer needs:

• $n$ new cells along one edge;
• $n$ new cells along the other edge;
• one corner counted in both edge descriptions, so subtract it once.

Therefore the new layer contains

\[n+n-1=2n-1\]

cells.

The picture did more than decorate a known equation. Its rules generated the correct recurrence, and the algebra verified it.

7. Return to $1+1=2$

Part 2 did not stop after checking one arithmetic expression. It constructed two tagged singleton shapes:

\[\{\star_L\} \qquad\text{and}\qquad \{\star_R\},\]

then formed their disjoint union:

\[\{\star_L\}\sqcup\{\star_R\}.\]

The tags make the positions distinct. Its cardinal shape is the unique nonempty free transitive involution developed in that tutorial.

This is an inner-ground proof:

one undivided position
plus one differently tagged undivided position
gives one reversible distinction and nothing outside it.

The interactive shape-search lab supplies a second kind of evidence. Every complete route over its finite candidate universe rejects the malformed shapes and accepts the same structural certificate.

The two pieces of evidence have different scopes:

• the structural derivation proves the general characterization under its stated rules;
• the finite lab tests the implementation and demonstrates route independence inside its declared finite universe.

8. When two directions create the deeper object

The Cantor–Schröder–Bernstein theorem is a beautiful bridge between the two proof virtues.

Suppose there is an injection

\[f\colon A\to B\]

Read exactly:

“$f$ is a function from $A$ to $B$.”

and an injection

\[g\colon B\to A.\]

Read exactly:

“$g$ is a function from $B$ to $A$.”

An injection never sends two different inputs to the same output. Informally,

\[|A|\le |B| \qquad\text{and}\qquad |B|\le |A|.\]

The theorem concludes that there is a bijection

\[h\colon A\to B.\]

A bijection is a one-to-one correspondence that reaches every element of the target.

The structural construction

Define

\[A_0=A\setminus g(B),\]

then recursively define

\[A_{n+1}=g\bigl(f(A_n)\bigr),\]

and collect all these layers:

\[A_{\ast}=\bigcup_{n=0}^{\infty}A_n.\]

Now define

\[h(a)= \begin{cases} f(a),&a\in A_{\ast},\\ g^{-1}(a),&a\notin A_{\ast}. \end{cases}\]

Read exactly:

“$h$ of $a$ equals $f$ of $a$ when $a$ is in $A_{\ast}$, and $h$ of $a$ equals the unique element $b$ for which $g$ of $b$ equals $a$ when $a$ is not in $A_{\ast}$.”

Because $g$ is injective, $g^{-1}(a)$ names at most one element. Because every element outside $A_{\ast}$ lies in the image of $g$, that second branch is defined where it is used.

The construction splices the two one-way injections into one bijection.

Why the splice is a bijection

It is injective on $A_{\ast}$ because $f$ is injective. It is injective outside $A_{\ast}$ because $g^{-1}$ is injective on the image of $g$.

The two branches cannot collide. If some $f(a)$ from the first branch equaled $g^{-1}(a’)$ from the second branch, then

\[a'=g(f(a)).\]

But $a\in A_{\ast}$ implies $g(f(a))\in A_{\ast}$ by the layer construction, contradicting $a’\notin A_{\ast}$.

It is also surjective. Take any $b\in B$.

• If $b=f(a)$ for some $a\in A_{\ast}$, then $h(a)=b$.
• Otherwise, let $a=g(b)$. That $a$ cannot lie in $A_{\ast}$. If it did, its layer would force $b$ to equal $f(a’)$ for an earlier $a’\in A_{\ast}$, contrary to the case assumption. Therefore $h(g(b))=g^{-1}(g(b))=b$.

So every $b$ is reached exactly once.

Here the Tao-style split does not hide the inner ground. The proof of antisymmetry for cardinal comparison is itself the construction of the shared structural witness.

9. A world-class inner ground: symmetry makes momentum constant

Noether’s 1918 theorem concerns invariant variational problems. A full statement requires more machinery than this beginner tutorial. A one-coordinate special case already shows the central shape.

Assumptions

Let

\[L(q,\dot q,t)\]

be a differentiable Lagrangian, and let the path $q(t)$ satisfy the Euler–Lagrange equation:

\[\frac{d}{dt} \left( \frac{\partial L}{\partial\dot q} \right) = \frac{\partial L}{\partial q}.\]

Read exactly:

“The derivative with respect to time of the partial derivative of $L$ with respect to $\dot q$ equals the partial derivative of $L$ with respect to $q$.”

Assume the Lagrangian is unchanged by translating the coordinate:

\[L(q+c,\dot q,t)=L(q,\dot q,t)\]

for every sufficiently small constant $c$ and every admissible $q$, $\dot q$, and $t$.

Read exactly:

“$L$ of $q$ plus $c$, $\dot q$, and $t$ equals $L$ of $q$, $\dot q$, and $t$.”

Differentiate this symmetry with respect to $c$ at $c=0$:

\[\frac{\partial L}{\partial q}=0.\]

Insert that fact into the Euler–Lagrange equation:

\[\frac{d}{dt} \left( \frac{\partial L}{\partial\dot q} \right) =0.\]

Therefore

\[p=\frac{\partial L}{\partial\dot q}\]

is constant along the motion.

For a free particle,

\[L=\frac12m\dot q^2,\]

so

\[p=m\dot q\]

is the familiar momentum.

Why this is an inner-ground explanation

A calculation might show that momentum at one time equals momentum at another time. The symmetry argument explains why an entire family of such equalities must hold:

1. Translation changes no law.
2. Therefore the coordinate contributes no preferred location.
3. The Euler–Lagrange equation has a zero force term in that direction.
4. Momentum in that direction is conserved.

The conservation equality is no longer an isolated coincidence. It is generated by a symmetry.

Boundary

This is a scoped one-parameter, one-coordinate consequence of Noether’s first theorem, not a statement of either theorem in full generality. It assumes differentiability, the Euler–Lagrange dynamics, and the stated continuous symmetry. Arbitrary systems do not inherit this conservation law.

The historical equality remark and Noether’s theorem are also different pieces of evidence. The theorem is used here as an example of the structural mathematical style described by Weyl, not as proof that the quoted remark was its source.

10. Measuring explanatory depth without pretending it is objective

Mathematical beauty is not a numerical invariant. Explanatory reach can still be stress-tested with concrete questions.

These tests do not assign an “understanding score.” They make the word deep less vague.

The odd-square proof has broad reach because one layer rule predicts every later case. The symmetry argument has broad reach because one invariance condition generates a conservation law across an entire trajectory.

11. Neither method should be turned into dogma

A valid split proof is still a proof

If $a\le b$ and $b\le a$ are established in a partial order, rejecting the proof because it lacks a preferred explanation would confuse validity with taste.

A beautiful picture can still be wrong

A geometric arrangement proves an identity only when the correspondence preserves the counted units, has no overlaps or omissions, and covers the declared cases.

A shared normal form can smuggle in the conclusion

If the normalizer is unsound, or if its rules silently identify objects that the model keeps distinct, matching outputs prove nothing about the original objects.

The “inner ground” need not be unique

One equality may have several useful explanations:

• a bijection;
• an algebraic factorization;
• a symmetry;
• a probabilistic coupling;
• a Fourier identity;
• a categorical universal property.

Different explanations preserve different information.

12. The combined workflow

The strongest working method is not “Tao or Noether.” It is a loop.

1. Declare the model and equality notion.
2. Split the equality when the ambient order makes that useful.
3. Record the witness for each direction.
4. Search for a shared mechanism: normal form, bijection, invariant, symmetry, or zero remainder.
5. Prove that the mechanism really implies the equality.
6. Remove one hypothesis and seek a counterexample.
7. Give the derivation or certificate to an independent checker.

The order can reverse during discovery. A structural idea may arrive first, then a two-direction audit can expose a hidden gap.

The compact lesson is:

Tao’s move opens the equality gate. Noether’s question asks what built the gate, why both keys fit, and which other gates use the same mechanism.

13. Practice problems

Problem 1: two inclusions

Prove A ∩ B = B ∩ A by two inclusions. Then write the single membership condition shared by both sides.

Answer: both sides have the condition x ∈ A and x ∈ B. Commutativity of logical conjunction is the inner ground.

Problem 2: find the missing hypothesis

A relation satisfies a ≼ b and b ≼ a. Does a = b follow?

Answer: only if the relation has an appropriate antisymmetry rule. A preorder yields equivalence, not necessarily literal identity.

Problem 3: predict the next layer

A square of side 12 grows to a square of side 13. How many unit cells enter in the new L-shaped layer?

Answer: 13 + 13 - 1 = 25, which is 2(13) - 1.

Problem 4: break the symmetry

Suppose a Lagrangian depends on position through a nonconstant potential V(q). Which step in the momentum argument can fail?

Answer: translation invariance can fail, so ∂L/∂q need not be zero. The Euler–Lagrange equation can then give a changing momentum.

14. What the two lenses finally reveal

The apparent disagreement dissolves once proof and explanation are separated.

A theorem needs correctness. A great tutorial should also expose the mechanism, state its boundary, and show what the mechanism predicts next.

Sources and further reading

• Terence Tao, “245A: Problem solving strategies”, especially “Split up equalities into inequalities.” The same post also discusses counterexamples, simpler cases, and abstraction.
• Hermann Weyl, “Emmy Noether,” Scripta Mathematica 3 (1935), 201–220. The equality remark is Weyl’s report. The essay is reprinted in Levels of Infinity: Selected Writings on Mathematics and Philosophy.
• Peter Roquette, “Emmy Noether and Hermann Weyl”, for historical context on Noether’s structural and abstract style.
• Emmy Noether, “Invariant Variation Problems”, M. A. Tavel’s English translation of the 1918 paper.