Looking at the breakthroughs achieved by AI models in mathematics in recent times, I gave it a try a few weeks ago to see if as a non-mathematician, I could use GPT to do some hard mathematics.
Using a free account of GPT, I posted a hard question in the chat and asked GPT to solve it. I reached a solution a few hours later which I have shared below, along with my experience itself of working with AI, which was perhaps more enlightening for me than the actual mathematical results.
Experience of doing Maths with AI
- Given a hard question that GPT cannot quickly solve, it thinks for some time and then posts some partial success or ideas, followed by suggested next step that it thinks it should follow. Once I give it a reply, again it comes back with next partial success or ideas, followed by suggested next step. The model keeps doing this in a loop, and after each iteration it checks in with me to make sure I'm ok with the direction it is heading in.
- In the beginning, I understood most of what it said, and was able to give meaningful opinions on next step. As it went deeper and deeper into advanced mathematics, I was only able to understand broadly what it said, and soon there were sections where the maths involved concepts I had not even heard of. In such cases, I pretended to nod along and just kept telling it to proceed with its suggested next step.
- As an approach, when GPT is stuck, it first tries to understand the shape of underlying reality, with the belief that the knowledge gained will help find the next steps. More importantly, it explains to the reader as well why it thinks a particular exploration will result in useful insights for the next step. If the gained insights are not as per expectations, it explains how it is interpreting this new insight and how it is re-calibrating its approach accordingly. This was a surprise for me. I had assumed that working with AI would result in an opaque discovery of final results, without any new learnings that explain how to get there, or that can be used in other hard problems. This will likely be true if you only read the final result, its theoroems, and proofs, but not if you also read AI's intermediate explorations and insights.
Perhaps AI companies that publish mathematical breakthroughs should also share these intermediate reasonings to showcase how AI can collaborate with mathematical community instead of it only being a result generator. Below I have shared only the final theorems and proofs, but let me know if you are interested in the entire chat.
Problem given to GPT
I first gave GPT a solved problem to see how it reacts.
The problem was to find necessary and sufficient conditions so that an infinitely nested radical \( P=\sqrt{a_1+\sqrt{a_2+\sqrt{a_3+\cdots}}} \) converges to a real number.
GPT correctly provided the solution along with the original author's name.
Then, I gave the hard problem of generalising the result.
I asked GPT to look at the nested radical
\( \sqrt{A_1 + B_1\sqrt{A_2 + B_2\sqrt{A_3 + B_3\sqrt{\cdots}}}} \)
where \(B_n\) can be mixture of positive and negative numbers, and asked it to find necessary and sufficient condition that this radical converges to a real number.
As far as I know, this is not a solved problem, but I could be completely wrong because I am not connected to the world of mathematical research.
AI's approach
GPT decided that we should first focus on the scenario where all \(Bn\) are negative and try and solve it. Once done, we can decide to come back for exploring mixed positive and negative signs. So, we focused on radicals with all negative signs, and GPT successfully solved this in a few hours.
I asked it to give a summary for this result (omitting all the dead ends and any intermediate insights that were not ultimately needed by the proof), and below is the summary generated by GPT. Perhaps next week, I'll ask GPT to continue its exploration into radicals with mixed signs.
See you then!
Summary by GPT: Convergence of Infinitely Nested Radicals
PART I. Nested radicals with all positive signs
Consider \[ P=\sqrt{a_1+\sqrt{a_2+\sqrt{a_3+\cdots}}}, \qquad a_n\geq0. \] Let \(P_{n,N}\) denote the finite radical beginning at position \(n\) and ending at position \(N\): \[ P_{n,N}=\sqrt{a_n+P_{n+1,N}}, \qquad P_{N+1,N}=0. \] The question is when \(P_{1,N}\) converges to a finite limit as \(N\to\infty\).
The infinite positive nested radical converges to a finite real number if and only if \[\boxed{ \sup_{n\geq1}a_n^{1/2^n}<\infty. } \tag{1} \] Thus the terms \(a_n\) must grow no faster than a constant raised to the power \(2^n\).
Proof
From the finite radical recursion, \[P_{n+1,N}\leq P_{n,N}^2\] Repeated application gives \[P_{n,N}\leq P_{1,N}^{2^{n-1}}\] Since \(a_n\leq P_{n,N}^2\), \[a_n^{1/2^n}\leq P_{1,N}\] If the infinite radical converges to a finite value \(P\), letting \(N\to\infty\) gives \[a_n^{1/2^n}\leq P\] Hence the sequence in (1) is bounded.
Sufficiency.
Suppose there is an \(M\geq1\) such that \[a_n\leq M^{2^n}\] for every \(n\). Normalize the finite radicals by \[Y_{n,N}=\frac{P_{n,N}}{M^{2^{n-1}}}\] Then \[Y_{n,N}^2 = \frac{a_n}{M^{2^n}}+Y_{n+1,N} \leq1+Y_{n+1,N} \] Let \[\varphi=\frac{1+\sqrt5}{2},\] so that \(\varphi^2=1+\varphi\). Starting from the terminal value \(Y_{N+1,N}=0\), backward induction gives \[Y_{n,N}\leq\varphi\] In particular, \[P_{1,N}\leq M\varphi\] Finally, \(P_{1,N}\) is nondecreasing as \(N\) increases, because all terms are nonnegative and every radical transformation is increasing. It is also bounded above.
Therefore it converges.
\(\square\)
The critical growth scale is \(2^n\). For example, \[a_n=C^{2^n}\] produces a convergent positive nested radical for every fixed \(C>0\).
Conversely, sufficiently rapid growth beyond this scale forces divergence.
We also considered the generalized positive form \[P=\sqrt{a_1+b_1\sqrt{a_2+b_2\sqrt{a_3+\cdots}}}, \qquad a_n,b_n>0\] This can be reduced to Herschfeld's theorem by rescaling. Define \[c_1=1,\qquad c_{n+1}=\frac{c_n^2}{b_n},\] and write \[P_n=c_nY_n\] Then \[Y_n=\sqrt{\frac{a_n}{c_n^2}+Y_{n+1}}\] Consequently:
The generalized positive nested radical converges if and only if \[ \boxed{ \sup_{n\geq1} \left(\frac{a_n}{c_n^2}\right)^{1/2^n} \lt \infty, \qquad c_{n+1}=\frac{c_n^2}{b_n},\quad c_1=1 } \tag{2} \] Thus positive multipliers do not require a fundamentally different convergence theory; they modify the effective coefficients.
PART II. Normalization of all-negative nested radicals
Consider \[R_1=\sqrt{A_1 - B_1\sqrt{A_2 - B_2\sqrt{A_3 - B_3\sqrt{\cdots}}}},\] where \(A_n, B_n>0\). Write the tail beginning at position \(n\) as \[R_n=\sqrt{A_n-B_nR_{n+1}}\] Normalize by putting \[R_n=\sqrt{A_n}\,Z_n\] Substitution gives \[Z_n = \sqrt{1 - \frac{B_n\sqrt{A_{n+1}}}{A_n}Z_{n+1} }\] Define the normalized coefficient \[ \boxed{\gamma_n=\frac{B_n\sqrt{A_{n+1}}}{A_n}} \tag{3} \] Our all-negative results apply to the admissible normalized regime \[0 \lt \gamma_n \leq 1.\] The normalized problem is therefore \[ \boxed{R=\sqrt{1 - \gamma_1\sqrt{1 - \gamma_2\sqrt{1 - \gamma_3\sqrt{\cdots}}}}} \tag{4} \] And it suffices to study the convergence of this normalized radical. For the rest of the summary, \(R\) denotes this normalized radical.
Define the individual radical transformation \[ \boxed{f_\gamma(x)=\sqrt{1-\gamma x}, \qquad 0\leq x\leq1} \tag{5} \] Unlike the positive case, every \(f_\gamma\) is decreasing.
This reversal of monotonicity is responsible for the different convergence behaviour.
For a finite truncation at depth \(N\), start with \[I_{N+1,N}=[0,1]\] Propagate backwards: \[I_{n,N}=f_{\gamma_n}(I_{n+1,N})\] Write \[I_{n,N}=[p_{n,N},u_{n,N}]\] Because \(f_{\gamma_n}\) is decreasing, \[ \boxed{\begin{aligned} p_{n,N}=\sqrt{1-\gamma_nu_{n+1,N}},\\ u_{n,N}=\sqrt{1-\gamma_np_{n+1,N}}. \end{aligned} } \tag{6} \]
The infinite normalized all-negative radical converges if and only if \[ \boxed{ \lim_{N\to\infty}(u_{1,N}-p_{1,N})=0. } \tag{7} \]
Proof
The intervals \(I_{1,N}\) are nested as \(N\) increases. Their endpoints are the two possible values obtained by terminating the finite radical with terminal tails zero and one.
Thus the even- and odd-depth truncations lie at the two endpoints of these intervals, with their order alternating.
If the widths tend to zero, both truncation subsequences approach the same limit.
\(\square\)
Let \[W_{n,N}=u_{n,N}-p_{n,N}\] Rationalizing the difference of the two square roots gives \[W_{n,N} = \frac{\gamma_n}{p_{n,N}+u_{n,N}}W_{n+1,N}\] Since the terminal width is one, \[ \boxed{W_{1,N} = \prod_{n=1}^{N} \frac{\gamma_n}{p_{n,N}+u_{n,N}}} \tag{8} \] This is an exact necessary-and-sufficient condition, though it still requires recursively calculating the endpoints.
PART III. The two-coefficient transformation and its contraction
The natural unit of analysis for negative radicals is a pair of consecutive coefficients.
Write \[a_k=\gamma_{2k-1}, \qquad b_k=\gamma_{2k}\] Define the increasing two-level transformation \[ \boxed{F_k(x)=f_{a_k}(f_{b_k}(x)) = \sqrt{1-a_k\sqrt{1-b_kx}}} \] The key property is that every such pair strictly contracts interval uncertainty in a suitable metric.
Introduce the log-odds coordinate \[ \boxed{\ell(x)=\log\frac{x}{1-x}, \qquad d(x,y)=|\ell(x)-\ell(y)|} \tag{9} \]
Lemma 1 — Pair contraction
For every \(a,b\in(0,1]\), and every interval \(I\subset(0,1)\), \[ \boxed{\operatorname{diam}_{d}\big(f_a\circ f_b(I)\big) \leq q\,\operatorname{diam}_{d}(I),} \tag{10} \] where \[ \boxed{q=\frac34+\frac{\sqrt3}{8} \lt 1} \] The numerical value is approximately \(0.966506\).
Proof
Take \(0 \lt x \lt y \lt 1\), and put \[s=f_\gamma(x),\qquad r=f_\gamma(y)\] Then \(s>r\).
Define \[P=\log\frac yx,\qquad Q=\log\frac{1-x}{1-y},\] and \[T=\log\frac sr+\log\frac{1+s}{1+r}\] Using \[1-f_\gamma(x)=\frac{\gamma x}{1+f_\gamma(x)},\] we find that the input and output log-odds widths are respectively \[P+Q,\qquad P+T\] Also, \[Q\geq2\log\frac sr\] and \[T\leq2\log\frac sr\] Hence \(T\leq Q\), proving that a single transformation never increases log-odds width.
Now suppose the input interval lies in \([1/2,1)\). Then \(P\leq Q\).
If \(\gamma\geq1/2\), then \(s\leq\sqrt3/2\). Using \[\log\frac{1+s}{1+r} \leq(s-r)\leq s\log\frac sr,\] we get \[T\leq \frac{1+\sqrt3/2}{2}Q\] If \(\gamma \lt 1/2\), then \(s,r>1/\sqrt2\), which gives \[T\leq\frac1{\sqrt2}Q\] Both estimates imply \[T\leq q_0 Q, \qquad q_0=\frac12+\frac{\sqrt3}{4}\] Therefore, \[\frac{P+T}{P+Q} \leq \frac{1+q_0}{2} = \frac34+\frac{\sqrt3}{8}=q\] A single transformation thus contracts log-odds width by \(q\) whenever its input interval lies above \(1/2\).
Finally consider \(f_a\circ f_b\).
- If the input interval lies below \(1/2\), the inner transformation maps it above \(1/2\), so the outer transformation contracts.
- If the input lies above \(1/2\), the inner transformation contracts, and the outer one is non-expanding.
- If the interval crosses \(1/2\), split it there and apply the preceding two arguments to the two pieces.
The pair therefore contracts by \(q \lt 1\) in every case.
\(\square\)
This result is fundamental: every two consecutive negative-radical levels reduce log-odds uncertainty by a fixed proportion.
The remaining challenge is to ensure that the uncertainty is initially bounded in this metric.
PART IV. Periodic all-negative coefficients
The contraction theorem allows us to classify every periodic coefficient sequence.
7. Period 1 — Constant coefficient
Suppose \[\gamma_n=a \qquad\text{for every }n\]
Theorem 3 — Period 1
\[ \boxed{ \begin{cases} 0 \lt a \lt 1 &: \text{the radical converges},\\ a=1 &: \text{the radical diverges}. \end{cases} } \] When \(0 \lt a \lt 1\), its value is \[ \boxed{R=\frac{\sqrt{a^2+4}-a}{2}} \tag{11} \]
Proof
Since the coefficients are constant, the limiting value satisfies \[R=\sqrt{1-aR}\] Solving the quadratic equation gives (11). If \(a=1\), the finite truncations alternate between one and zero, so the radical diverges.
\(\square\)
For example, at \(a=1/2\), \[R=\frac{\sqrt{17}-1}{4}\approx0.780776\]
Suppose \[(\gamma_n)=(a,b,a,b,a,b,\ldots)\] Theorem 4 — Period 2
\[ \boxed{\text{The radical converges} \iff a \lt 1\text{ and }b \lt 1} \tag{12} \] When it converges, the two alternating tail values \(x,y\) satisfy \[ \boxed{x=\sqrt{1-ay}, \qquad y=\sqrt{1-bx}} \tag{13} \]
Proof
If both coefficients are below one, the pair transformation maps \([0,1]\) strictly into \((0,1)\). Repeated application contracts log-odds width by Lemma 1.
Suppose instead that \(a=1\). The odd-position coefficients are all one. By the boundary theorem in Section 8, convergence would require \[\sum_{k=1}^{\infty}2^{-k}\log\frac1b=+\infty\] This series is finite for every fixed \(b>0\). Hence the radical diverges.The case \(b=1\) is symmetric.
\(\square\)
An important consequence is that the two alternating coefficients need not be equal.
Suppose \[(\gamma_n)=(a,b,c,a,b,c,\ldots)\] Theorem 5 — Period 3
\[ \boxed{ \text{The radical converges} \iff \min(a,b,c) \lt 1} \tag{14} \] When it converges, the three repeating tail values satisfy \[ \boxed{ \begin{aligned} x&=\sqrt{1-ay},\\ y&=\sqrt{1-bz},\\ z&=\sqrt{1-cx}. \end{aligned} } \]
Proof
If at least one coefficient is below one, the six-level transformation maps \([0,1]\) strictly into \((0,1)\). It consists of three two-level transformations, each contracting by Lemma 1.
Repeated application therefore forces the interval width to zero.
If all three coefficients equal one, the finite truncations alternate between zero and one, so the radical diverges.
\(\square\)
The period-3 result generalizes directly.
Theorem 6 — Arbitrary odd periods
Suppose the coefficients have an odd period \(p\). Then \[ \boxed{ \text{The radical converges} \iff \exists j\in\{1,\ldots,p\}:\gamma_j \lt 1} \tag{15} \] Thus one subcritical coefficient anywhere in an odd period is sufficient.
Proof
Because \(p\) is odd, the second occurrence of every coefficient is shifted to the opposite positional parity.
Therefore, if at least one coefficient in the period is below one, the two-period block contains a subcritical coefficient in both parity classes.
An even-length composition with a subcritical coefficient in each parity class maps \([0,1]\) strictly into \((0,1)\).
It consists of \(p\) two-level transformations, so its log-odds contraction factor is at most \(q^p \lt 1\).
Repeated two-period blocks therefore force the interval width to zero.
If all coefficients are one, the radical diverges.
\(\square\)
For even periods, the answer is different.
Theorem 7 — Arbitrary even periods
Suppose the coefficients have an even period \(p\).
Then \[ \boxed{ \begin{aligned} \text{R converges} \iff{} &\text{at least one odd-position coefficient}\\ &\text{in the period is below one, and}\\ &\text{at least one even-position coefficient}\\ &\text{in the period is below one.} \end{aligned} } \tag{16} \]
Proof
Its composition maps \([0,1]\) strictly into \((0,1)\) precisely when there is at least one coefficient below one at each positional parity.
If both are present, the period consists of \(p/2\) two-level transformations. Its log-odds contraction factor is at most \[ q^{p/2} \lt 1 \] Repeated application therefore gives convergence.
If one parity consists entirely of ones, the boundary theorem in Section 8 requires a geometrically weighted logarithmic series over the other parity to diverge.
For a periodic sequence with strictly positive coefficients, that series is finite. Hence the radical diverges.
\(\square\)
For the period-four pattern \[(a,b,c,d,a,b,c,d,\ldots),\] convergence occurs precisely when \[ \boxed{\min(a,c) \lt 1 \quad\text{and}\quad \min(b,d) \lt 1} \tag{17} \] For example, \[(a,1,1,b,a,1,1,b,\ldots)\] converges whenever \(a,b \lt 1\).
\[ \begin{array} \text{Period} & \text{Necessary-and-sufficient condition} \\ \text{1} & \text{Constant coefficient below one} \\ \text{2} & \text{Both coefficients below one} \\ \text{3} & \text{At least one coefficient below one} \\ \text{Any odd period} & \text{At least one coefficient below one} \\ \text{Any even period} & \text{At least one coefficient below one} \\ & \text{on each positional parity} \\ \end{array} \] The fundamental distinction is that odd periods interchange the two positional parities after one period, whereas even periods preserve them.
PART V. Exact results for non-periodic boundary sequences
The next two results are particularly useful because they give simple series criteria without requiring periodicity.
12. All odd-position coefficients equal one
Suppose \[ \boxed{\gamma_{2k-1}=1,\qquad \gamma_{2k}=b_k} \tag{18} \] Theorem 8 — Odd positions at the boundary
The infinite radical converges if and only if \[ \boxed{ \sum_{k=1}^{\infty} 2^{-k}\log\frac1{b_k}=+\infty } \tag{19} \] When it converges, its value is zero.
Proof
The radical therefore converges precisely when the upper endpoint tends to zero. Equivalently, the corresponding even-position tail must tend to one.
Write that even-position tail as \(Y_k\). It satisfies \[Y_k=\sqrt{1-b_k\sqrt{1-Y_{k+1}}}\] Define \[t_k=-\log(1-Y_k)\] Using \[1-\sqrt{1-x} = \frac{x}{1+\sqrt{1-x}},\] we obtain \[ \boxed{t_k= \frac12t_{k+1} +\log\frac1{b_k} +\log(1+Y_k)} \tag{20} \] The final term lies between zero and \(\log2\).
Iterating the recurrence over a finite truncation shows that \(t_1\) differs from the geometrically weighted sum \[\sum_{k=1}^{N}2^{-(k-1)}\log\frac1{b_k}\] by a bounded quantity independent of \(N\).
Thus \(t_1\to+\infty\) precisely when the series in (19) diverges.
Since \(t_1\to+\infty\) is equivalent to \(Y_1\to1\), this proves the theorem.
\(\square\)
The reverse arrangement is \[ \boxed{\gamma_{2k-1}=a_k,\qquad \gamma_{2k}=1} \tag{21} \] Theorem 9 — Even positions at the boundary
The infinite radical converges if and only if \[ \boxed{\sum_{k=1}^{\infty} 2^{-k}\log\frac1{a_k}=+\infty} \tag{22} \] When it converges, its value is one.
Proof
Convergence requires the lower endpoint to tend to one.
The corresponding odd-position tail satisfies \[X_k=\sqrt{1-a_k\sqrt{1-X_{k+1}}}\] Defining \(t_k=-\log(1-X_k)\) gives exactly the recurrence (20), with \(a_k\) replacing \(b_k\).
The same argument proves condition (22).
\(\square\)
These boundary theorems also cover sequences with an arbitrary finite prefix, since a finite prefix of strictly monotone transformations does not change whether an interval has collapsed to a single point.
PART VI. The uniform two-coefficient approximation
This is the technical foundation of the general theorem.
For a pair \(a,b\), define \[A=\log\frac1{1-a}, \qquad B=\log\frac1{1-b}, \qquad G=\log\frac1b\] Also define \[L=G-A,\qquad U=B,\] and \[p=-\log(1-\sqrt{1-a})\] When \(a=1\), we have \(p=0\) and \(L=-\infty\). The two-level transformation is \[F(x)=\sqrt{1-a\sqrt{1-bx}}\] In log-odds coordinates, it becomes \[ \boxed{h(z)=\ell(F(\sigma(z)))} \tag{23} \] Define the clipping function \[\operatorname{clip}(z;L,U) = \min\{\max(z,L),U\},\] with the natural limiting conventions for infinite endpoints.
Lemma 2 — Clipped-affine approximation
There is an absolute constant \[C_0=4\log2\] such that the exact pair transformation can be approximated uniformly by \[ \boxed{\tau(z)=p-\frac G2+\frac12\operatorname{clip}(z;L,U)} \tag{24} \] If \(L>U\), the window is empty and \(\tau\) is instead the constant \(h(-\infty)\).
For every \(a,b\in(0,1]\), \[ \boxed{\sup_{z\in\mathbb R}|h(z)-\tau(z)| \leq C_0} \tag{25} \] The crucial feature is that the constant does not depend on the coefficients.
Proof
Using \[1-y=\frac{b\sigma(z)}{1+y},\] we obtain \[ \boxed{\frac q\delta = (1-e^{-A}) \frac{e^{z-L}}{(1+e^z)(1+y)}} \tag{29} \] Define \[C_L(z)= \begin{cases} \min\{\max(z-L,0),-L\},&L \lt 0,\\ 0,&L\geq0. \end{cases} \] A direct division into the ranges \(z\leq L\), \(L\leq z\leq0\), and \(z\geq0\) gives \[ \boxed{\left| \log\left(1+\frac q\delta\right)-C_L(z) \right| \leq\log8} \tag{30} \]
For example, when \(L \lt 0\) and \(a\geq1/2\), equation (29) shows that the ratio \(q/\delta\) is between \(e^{z-L}/8\) and \(e^{z-L}\) in the middle range, and between \(e^{-L}/8\) and \(e^{-L}\) in the upper range. When \(a \lt 1/2\), we have \(A \lt \log2\), so the required error is directly bounded.
Combining (28) and (30), \[\left| K(z)-K(-\infty)-\frac12C_L(z) \right| \leq\frac52\log2\] For \(L\leq0\), the identity \[C_L(z)+\operatorname{clip}(z;0,B) = \operatorname{clip}(z;L,B)-L\] gives the approximation (24), using \[K(-\infty)-\log a=p-\frac A2\] If \(0 \lt L\leq B\), the replacement changes the clipped expression by at most \(L\). But \[L \lt G,\qquad L\leq B,\] and \[\min(B,G)\leq\log2\] Thus the extra error is at most \(\frac12\log2\).
If \(L>B\), then the window is empty. In this case \(B \lt L \lt G\) and \(B \lt \log2\); equation (29) shows that the variation of \(h\) from \(h(-\infty)\) is at most \(2\log2\).
Finally, if \(a=1\), then \(L=-\infty\). Direct substitution into (26) gives the required approximation with error at most \(2.5\log2\).
All cases are therefore covered by \(C_0=4\log2\).
\(\square\)
Every pair behaves approximately like a clipped affine map:
- Below its active window, the map is approximately constant.
- Inside the window, it has slope \(1/2\).
- Above the window, it is approximately constant again.
The approximation error is uniformly bounded, even when either coefficient approaches zero or one.
PART VII. Discounted interval intersections
For the coefficient pair at position \(k\), write \[ \boxed{ \begin{aligned} a_k&=\gamma_{2k-1},\\ b_k&=\gamma_{2k},\\ A_k&=\log\frac1{1-a_k},\\ B_k&=\log\frac1{1-b_k},\\ G_k&=\log\frac1{b_k},\\ L_k&=G_k-A_k,\\ U_k&=B_k. \end{aligned} } \] Define \[ \boxed{c_k=-\log(1-\sqrt{1-a_k})-\frac12G_k} \tag{31} \] The approximate pair map is \[\tau_k(z)=c_k+\frac12\operatorname{clip}(z;L_k,U_k)\] when \(L_k\leq U_k\). For an empty window, use the constant approximation from Lemma 2.
Every nonconstant \(\tau_k\) is \(1/2\)-Lipschitz.
Consequently, after composing any number of pairs, the pointwise error between the exact and approximate compositions is bounded by \[C_0+\frac{C_0}{2}+\frac{C_0}{4}+\cdots \leq2C_0 \] The corresponding interval diameters therefore differ by at most \[4C_0=16\log2\]
Define the discounted cumulative centre \[S_0=0,\qquad S_k=\sum_{r=1}^{k}2^{1-r}c_r\] For each pair, define the interval \[\boxed{K_k=[X_k^-,X_k^+],}\] where \[ \boxed{ \begin{aligned} X_k^-&=S_k+2^{-k}L_k,\\ X_k^+&=S_k+2^{-k}U_k. \end{aligned} } \tag{32} \]
If \(L_k>U_k\), the interval is empty.
Lemma 3 — Interval-intersection identity
The range of the clipped composition \[\tau_1\circ\tau_2\circ\cdots\circ\tau_N\] is exactly \[ \boxed{\bigcap_{k=1}^{N}K_k.} \tag{33} \]
Explanation
The identity also holds at the window boundaries because clipping maps the exterior of a window to its endpoints.
\(\square\)
For every \(k\), define \[ \boxed{\Delta_k = \left[\inf_{j\geq k}X_j^+-\sup_{j\geq k}X_j^-\right]_+} \tag{34} \] This is the length of the intersection of all discounted windows from \(k\) onward.
Let \(d_k\) be the limiting log-odds diameter of the tail beginning at pair \(k\).
Provided both positional parities contain infinitely many coefficients strictly below one, the limiting tail intervals are interior to \((0,1)\), so these diameters are finite.
The pair approximation and the interval-intersection identity give \[ \boxed{|d_k-2^{k-1}\Delta_k|\leq16\log2} \tag{35} \] The quantities \(\Delta_k\) are nondecreasing: removing an early window restriction cannot reduce the intersection.
We can now establish:
Theorem 10 — Exact interval-intersection criterion
If both odd- and even-position coefficient subsequences contain infinitely many terms strictly below one, then \[ \boxed{ R\text{ converges} \iff \limsup_{k\to\infty}X_k^- \geq \liminf_{k\to\infty}X_k^+} \tag{36} \]
Proof
Since the finite prefix consists of strictly monotone transformations, the root interval collapses if and only if every fixed tail interval collapses.
If the radical converges, \(d_k=0\) for every \(k\). Equation (35) then gives \[0\leq\Delta_k \leq\frac{16\log2}{2^{k-1}}\] Since \(\Delta_k\) is nondecreasing and tends to zero, every \(\Delta_k\) must be zero.
Conversely, suppose every \(\Delta_k=0\). Equation (35) implies \[d_k\leq16\log2\] for every \(k\).
By Lemma 1, each two-level transformation contracts log-odds width by \(q \lt 1\). Therefore, \[d_1\leq q^{k-1}d_k \leq16\log2\,q^{k-1}\] Letting \(k\to\infty\), we obtain \(d_1=0\). Thus the radical converges.
Finally, \[\lim_{k\to\infty}\Delta_k = \left[\liminf_{k\to\infty}X_k^+- \limsup_{k\to\infty}X_k^- \right]_+\] Hence all the \(\Delta_k\) vanish precisely when the inequality in (36) holds.
\(\square\)
This is the bridge between the finite-width problem and a coefficient-only convergence condition.
PART VIII. Reduction to the general necessary-and-sufficient condition
The interval criterion can be simplified considerably.
19. Two weighted logarithmic series
Recall \[a_k=\gamma_{2k-1},\qquad b_k=\gamma_{2k}\] Define \[ \boxed{S_o=\sum_{k=1}^{\infty}2^{-k}\log\frac1{a_k}, \qquad S_e= \sum_{k=1}^{\infty}2^{-k}\log\frac1{b_k}} \tag{37} \] These are the odd- and even-parity weighted logarithmic series.
Also define the two parity-wise deficit measures \[\boxed{\eta_o= \liminf_{k\to\infty} 2^{-k}\log\frac1{1-a_k},}\] \[ \boxed{\eta_e=\liminf_{k\to\infty} 2^{-k}\log\frac1{1-b_k}} \tag{38} \] All quantities are interpreted in the extended nonnegative real numbers. In particular, \[\log\frac1{1-a_k}=+\infty\] when \(a_k=1\).
We first work under the assumption that both parity classes contain infinitely many coefficients below one.
Define \[p_k=-\log(1-\sqrt{1-a_k}),\] and \[P_k=\sum_{r=1}^{k}2^{1-r}p_r,\qquad Q_k=\sum_{r=1}^{k}2^{-r}G_r\] Let \[R_k=P_k-Q_k\] Because \[p_k=\log\frac1{a_k}+\log(1+\sqrt{1-a_k}),\] the second term is bounded between zero and \(\log2\). Therefore, \[ \boxed{ \begin{aligned} P_k\text{ is bounded} &\iff S_o \lt \infty,\\ Q_k\text{ is bounded} &\iff S_e \lt \infty. \end{aligned} } \tag{39} \] Substituting the definitions into the discounted window endpoints gives \[ \boxed{ \begin{aligned} X_k^-&=R_k+2^{-k}G_k-2^{-k}A_k,\\ X_k^+&=R_k+2^{-k}B_k. \end{aligned} } \tag{40} \] Suppose the radical diverges.
By Theorem 10, the limiting intersection of the windows has positive length. Thus there are finite \(u \lt v\) such that, for every sufficiently large \(k\), \[X_k^-\leq u,\qquad X_k^+\geq v\] Put \[\theta=\frac{u+v}{2},\qquad \varepsilon=\frac{v-u}{2}>0\] Divide the indices into two classes.
High indices: \(R_k\geq\theta\).
The first inequality in (40) gives \[2^{-k}A_k\geq\varepsilon\] Hence \[A_k\geq\varepsilon2^k\] But \[p_k=-\log(1-e^{-A_k/2}),\] so, for sufficiently large \(k\), \[2^{1-k}p_k \leq 2^{2-k}e^{-\varepsilon2^{k-1}}\] The right-hand side is summable. Thus the increments of \(P_k\) over high indices are summable.
Low indices: \(R_k \lt \theta\).
The second inequality in (40) gives \[2^{-k}B_k>\varepsilon\] Therefore, \[2^{-k}G_k \leq 2^{1-k}e^{-\varepsilon2^k}\] Again, the right-hand side is summable. Thus the increments of \(Q_k\) over low indices are summable.
Now observe that \[ \boxed{R_k-R_{k-1} = (P_k-P_{k-1})-(Q_k-Q_{k-1})} \tag{41} \] On each consecutive run of high indices, \(R_k\) enters below \(\theta\) and leaves above it. Telescoping (41) shows that the total increment of \(Q_k\) on such a run cannot exceed the increment of \(P_k\).
Similarly, on each consecutive run of low indices, the total increment of \(P_k\) cannot exceed that of \(Q_k\).
The summability estimates above therefore force both \(P_k\) and \(Q_k\) to converge.
We have proved \[ \boxed{\text{Divergence of the radical} \Longrightarrow S_o \lt \infty\text{ and }S_e \lt \infty} \tag{42} \] This is the crucial reduction.
Suppose now that both series converge.
Then \[R_k\longrightarrow R_\infty,\] and \[2^{-k}G_k\longrightarrow0\] Using (40), \[\limsup_{k\to\infty}X_k^- = R_\infty-\eta_o,\] and \[\liminf_{k\to\infty}X_k^+ = R_\infty+\eta_e\] The interval criterion (36) therefore becomes \[R_\infty-\eta_o \geq R_\infty+\eta_e\] Since \(\eta_o,\eta_e\geq0\), this is equivalent to \[ \boxed{\eta_o=\eta_e=0} \tag{43} \]
Combining this with (42), we obtain the principal result.
Theorem 11 — General necessary-and-sufficient condition for all-negative radicals
For the normalized radical \[R= \sqrt{1-\gamma_1\sqrt{1-\gamma_2 \sqrt{1-\gamma_3\cdots}}}, \qquad 0 \lt \gamma_n\leq1,\] define \(a_k=\gamma_{2k-1}\), \(b_k=\gamma_{2k}\), and the quantities in (37)–(38). Then \[ \boxed{ \begin{aligned} R\text{ converges}\iff{}& S_o=+\infty\\ &\text{or }S_e=+\infty\\ &\text{or }(\eta_o=0\text{ and }\eta_e=0). \end{aligned} } \tag{44} \] Equivalently, \[ \boxed{ \begin{aligned} R\text{ diverges}\iff{}& S_o \lt \infty,\quad S_e \lt \infty,\\ &\text{and }(\eta_o>0\text{ or }\eta_e>0). \end{aligned} } \tag{45} \]
Boundary cases
If the odd coefficients are eventually all one, Theorem 8 gives \[R\text{ converges} \iff S_e=+\infty\] This agrees with (44), because then \(\eta_o=+\infty\).
Similarly, if the even coefficients are eventually all one, Theorem 9 gives \[R\text{ converges} \iff S_o=+\infty\] If both parity classes are eventually all one, both series are finite and both deficit measures are infinite, so the radical diverges.
Thus Theorem 11 covers every sequence in the normalized class \(0 \lt \gamma_n\leq1\), including all boundary cases.
\(\square\)
PART IX. The general theorem in a compact single-energy form
The general criterion can be compressed into one scalar quantity constructed from the log-odds of the coefficients.
22. Weighted log-odds
Define \[\lambda_n=\log\frac{\gamma_n}{1-\gamma_n}\] For any real \(x\), write \[x^+=\max(x,0), \qquad x^-=\max(-x,0)\] Define the negative energy \[ \boxed{\mathcal N=\sum_{k=1}^{\infty} 2^{-k} \left[(\lambda_{2k-1})^-+(\lambda_{2k})^-\right]} \tag{46} \] Define the positive barrier \[ \boxed{\mathcal D=\liminf_{k\to\infty}2^{-k}(\lambda_{2k-1})^+ + \liminf_{k\to\infty}2^{-k}(\lambda_{2k})^+} \tag{47} \] Finally, define \[ \boxed{\mathcal H=\mathcal N+\frac1{\mathcal D}} \tag{48} \] We adopt the conventions \[\frac10=+\infty,\qquad \frac1{+\infty}=0\] Theorem 12 — Single weighted-energy criterion
The all-negative normalized radical converges if and only if \[\boxed{\mathcal H=+\infty}\]
Proof
After multiplication by the geometric weights, these bounded differences do not affect convergence of the series or the values of the parity-wise lower limits.
Therefore, \[\mathcal N \lt \infty \iff S_o \lt \infty\text{ and }S_e \lt \infty, \] and \[\mathcal D=0 \iff\eta_o=\eta_e=0\] The result follows directly from Theorem 11.
\(\square\)
The negative energy measures the accumulated effect of small coefficients.
The positive barrier measures whether either parity remains persistently close to one on the critical exponential scale.
The classification is: \[ \begin{array} \text{Negative energy } \mathcal N\ & \text{Positive barrier } \mathcal D & \text{Result} \\ \text{Infinite} & \text{Any value} & \text{Converges} \\ \text{Finite} & \text{Zero} & \text{Converges} \\ \text{Finite} & \text{Positive} & \text{Diverges} \\ \end{array} \] The two parity-wise lower limits must remain independent. Replacing them by a single synchronized lower limit is not generally valid.
PART X. Important consequences of the arbitrary-sequence theorem
Suppose \[\gamma_n\longrightarrow1\] Then both weighted logarithmic series \(S_o,S_e\) are finite, because \[\log\frac1{\gamma_n}\longrightarrow0\]
The general theorem therefore reduces to a particularly elegant condition.
Theorem 13 — Near-critical convergence criterion
If \(\gamma_n\to1\), then \[ \boxed{ \begin{aligned} R\text{ converges}\iff{}& \liminf_{\substack{n\to\infty\\n\text{ odd}}}\frac{\log(1/(1-\gamma_n))}{2^{n/2}}=0\\[1mm] &\text{and}\\[1mm] &\liminf_{\substack{n\to\infty\\n\text{ even}}}\frac{\log(1/(1-\gamma_n))}{2^{n/2}}=0. \end{aligned} } \tag{49} \] Thus, on each positional parity, there must be arbitrarily deep coefficients whose departure from one is sufficiently large relative to their depth.
The two parities need not achieve this at the same indices.
Consider \[ \boxed{\gamma_n=1-\exp(-Mc^n),\qquad M>0,\quad c>0} \tag{50} \] Then \[\log\frac1{1-\gamma_n}=Mc^n\] For \(c>1\), applying Theorem 13 gives \[\frac{Mc^n}{2^{n/2}} = M\left(\frac c{\sqrt2}\right)^n\] For \(0 \lt c\leq1\), the general theorem gives convergence directly.
Combining the cases, \[ \boxed{ \begin{cases} 0 \lt c \lt \sqrt2 &: \text{converges},\\ c=\sqrt2 &: \text{diverges},\\ c>\sqrt2 &: \text{diverges}. \end{cases} } \tag{51} \] The critical value is exactly \(\sqrt2\).
The borderline deserves particular attention: even when \(M\) is arbitrarily small, the sequence with \(c=\sqrt2\) diverges. The uncertainty can be extremely small, but it does not disappear.
A useful corollary of Theorem 11 is the following.
Corollary 2 — Sparse subcritical coefficients
Suppose there is a fixed \(\rho \lt 1\) such that \[\gamma_n\leq\rho\] for infinitely many odd positions and infinitely many even positions.
Then the infinite radical converges.
Proof
The result follows from Theorem 11.
\(\square\)
The radical converges, despite the coefficients being equal to one at almost every position.
This is a particularly interesting non-periodic example.
Consider the sequence \[ \boxed{ \begin{aligned} \gamma_{2k-1}&=1-e^{-4^k},\\ \gamma_{2k}&=e^{-2\cdot4^k}. \end{aligned} } \tag{53} \] The odd coefficients are extraordinarily close to one, while the even coefficients are extraordinarily close to zero.
For the even-position series, \[ \begin{aligned} S_e &=\sum_{k=1}^{\infty} 2^{-k}\log\frac1{e^{-2\cdot4^k}}\\ &=\sum_{k=1}^{\infty}2^{-k}(2\cdot4^k)\\ &=+\infty. \end{aligned} \] Therefore the radical converges, irrespective of the odd-position behaviour.
This illustrates an important feature of the general theorem: a sufficiently strong accumulation of small coefficients on just one parity can guarantee convergence.
PART XI. Consolidated results
The following tables collect the main milestone theorems.
Radical with all positive signs
\[ \begin{array} \text{Coefficient pattern} & \text{Necessary-and-sufficient condition} \\ \text{Arbitrary} a_n\geq0 & \sup_n a_n^{1/2^n} \lt \infty \\ \text{Arbitrary} a_n,b_n>0 & \text{Herschfeld condition after rescaling, equation (2)} \\ \end{array} \] Radical with all negative signs
\[ \begin{array} \text{Coefficient pattern} & \text{Necessary-and-sufficient condition} \\ \text{Constant coefficient} a & a \lt 1 \\ \text{Period 2} & \text{Both coefficients below one} \\ \text{Period 3} & \text{At least one coefficient below one} \\ \text{Any odd period} & \text{At least one coefficient below one} \\ \text{Any even period} & \text{At least one coefficient below one on each parity} \\ \text{Odd coefficients eventually one} & S_e=+\infty \\ \text{Even coefficients eventually one} & S_o=+\infty \\ \text{Arbitrary coefficients} & \text{Theorem 11, equation (44)} \\ \text{Coefficients approaching one} & \text{Theorem 13, equation (49)} \\ \end{array} \]
PART XII. Overall mathematical conclusions
1. The positive and negative cases are fundamentally different
For positive nested radicals, convergence is controlled by the individual terms and their growth on the scale \(2^n\).
For negative nested radicals, the problem is controlled by two-level transformations. Their repeated action produces the critical scale \(2^{n/2}\).
2. The odd/even distinction is intrinsic
For periodic sequences:
- An odd period interchanges the two positional parities after one period. A single subcritical coefficient is enough.
- An even period preserves the positional parities. A subcritical coefficient is required on each parity.
For arbitrary sequences, the same distinction survives as two independent parity-wise lower-limit conditions.
3. The arbitrary-sequence problem has a coefficient-only solution
The main result is Theorem 11:
\[
\boxed{
\begin{aligned}
R\text{ converges}\iff{}&
\sum_{k=1}^{\infty}2^{-k}\log\frac1{\gamma_{2k-1}}=+\infty\\
&\quad\text{or}\\
&\sum_{k=1}^{\infty}2^{-k}\log\frac1{\gamma_{2k}}=+\infty\\
&\quad\text{or}\\
&\left[\liminf_{k\to\infty}2^{-k}\log\frac1{1-\gamma_{2k-1}}=0\right.\\
&\qquad\left.\text{and }\liminf_{k\to\infty}2^{-k}\log\frac1{1-\gamma_{2k}}=0\right].
\end{aligned}
}
\]
This condition does not require periodicity, monotonicity, positive density of subcritical coefficients, or convergence of the coefficient sequence.
It covers the full normalized range
\[0 \lt \gamma_n\leq1\]
The case \(\gamma_n>1\) is not covered: then real-valuedness depends on the actual subsequent tail values, and a separate analysis is needed.




