Understand any dense paper
deeply enough to explain it

Drop in the paper. Instead of summarizing it, Ritsu does what a good advisor does — shows what it assumed, rebuilds the step it skipped, and asks where the claim breaks — until you can explain it yourself.

Get started for free

It was never written for you

You're not failing the paper. It was written for someone else.

A paper isn't trying to teach you. It's trying to survive peer review — and those produce very different documents.

The methods section is compressed because the reviewer already knows the method. The notation goes unexplained because the reviewer already knows the notation. The step between equation 4 and equation 5 is missing because, to the reviewer, it's obvious.

You are not the reviewer. You've read it three times and still couldn't say, out loud, what they actually did or why it works. That isn't a failure of your intelligence — it's the document working exactly as designed, on someone it was never written for. A paper is written to be defended, not to be understood.

There's only one way in: stop trying to read it. Start interrogating it.

On the Convergence of Stochastic Gradient Methods

K. Ito · R. Chen · A. Novak — Dept. of Statistics

Preprint · v3 · under review

Abstract

We establish a non-asymptotic $O(1/\sqrt{T})$ convergence rate for stochastic gradient descent under the classical Robbins–Monro step-size conditions, matching the known lower bound up to constants.

3 · Main Result

Assuming the standard smoothness and bounded-variance conditions, the update $w_{t+1} = w_t - \eta_t \nabla f(w_t)$ attains the rate $O(1/\sqrt{T})$ ; it follows that $\mathbb{E}\,[f(\bar w_T)] - f^{*} \le \tfrac{G^2 + \sigma^2}{2\sqrt{T}}$ whenever the step sizes satisfy $\textstyle\sum_t \eta_t^2 < \infty$ — the argument mirroring the classical Robbins–Monro analysis.

3.1 · Proof Sketch

Expanding the smoothness inequality and taking expectations, $\mathbb{E}\,\|w_{t+1} - w^{*}\|^{2} \le (1 - \mu\eta_t)\,\mathbb{E}\,\|w_t - w^{*}\|^{2} + \eta_t^{2}\sigma^{2}$ ; telescoping over t and dividing by the step-size sum gives the claim. The bound fails when the noise variance is unbounded — see Remark 2.

Figure 2: Suboptimality

f(\bar w_T) - f^{*}

vs. iterations under decaying step sizes.

preprint — page 33

reading it again still isn't having it

Everything it assumed you already knew.

The workflow

A good advisor doesn't summarize the paper. They interrogate it.

On the Convergence of Stochastic Gradient Methods

K. Ito · R. Chen · A. Novak — Dept. of Statistics

Preprint · v3 · under review

Abstract

We establish a non-asymptotic $O(1/\sqrt{T})$ convergence rate for stochastic gradient descent under the classical Robbins–Monro step-size conditions, matching the known lower bound up to constants.

3 · Main Result

Assuming the standard smoothness and bounded-variance conditions, the update $w_{t+1} = w_t - \eta_t \nabla f(w_t)$ attains the rate $O(1/\sqrt{T})$ ; it follows that $\mathbb{E}\,[f(\bar w_T)] - f^{*} \le \tfrac{G^2 + \sigma^2}{2\sqrt{T}}$ whenever the step sizes satisfy $\textstyle\sum_t \eta_t^2 < \infty$ — the argument mirroring the classical Robbins–Monro analysis.

3.1 · Proof Sketch

Expanding the smoothness inequality and taking expectations, $\mathbb{E}\,\|w_{t+1} - w^{*}\|^{2} \le (1 - \mu\eta_t)\,\mathbb{E}\,\|w_t - w^{*}\|^{2} + \eta_t^{2}\sigma^{2}$ ; telescoping over t and dividing by the step-size sum gives the claim. The bound fails when the noise variance is unbounded — see Remark 2.

preprint — page 33

the reasoning, restored01 / 06

sgd-convergence.pdfthe file, not the link

01 · Drop in the PDF

01

Drop in the PDF

arXiv PDF, a journal PDF, a scan. Not the link — the file. Ritsu keeps the equations as math and shows every figure alongside the text.

On the Convergence of Stochastic Gradient Methods

K. Ito · R. Chen · A. Novak — Dept. of Statistics

Preprint · v3 · under review

Abstract

We establish a non-asymptotic $O(1/\sqrt{T})$ convergence rate for stochastic gradient descent under the classical Robbins–Monro step-size conditions, matching the known lower bound up to constants.

3 · Main Result

Assuming the standard smoothness and bounded-variance conditions, the update $w_{t+1} = w_t - \eta_t \nabla f(w_t)$ attains the rate $O(1/\sqrt{T})$ ; it follows that $\mathbb{E}\,[f(\bar w_T)] - f^{*} \le \tfrac{G^2 + \sigma^2}{2\sqrt{T}}$ whenever the step sizes satisfy $\textstyle\sum_t \eta_t^2 < \infty$ — the argument mirroring the classical Robbins–Monro analysis.

preprint — page 33

the reasoning, restored01 / 06

sgd-convergence.pdfthe file, not the link

02

/prereq — what it assumed you'd read

Almost every “this paper is impenetrable” moment is an unstated dependency — it assumes you've read the one it's answering. Ritsu names it, in learning order, and the paper goes from impossible to merely hard.

On the Convergence of Stochastic Gradient Methods

K. Ito · R. Chen · A. Novak — Dept. of Statistics

Preprint · v3 · under review

Abstract

We establish a non-asymptotic $O(1/\sqrt{T})$ convergence rate for stochastic gradient descent under the classical Robbins–Monro step-size conditions, matching the known lower bound up to constants.

3 · Main Result

Assuming the standard smoothness and bounded-variance conditions, the update $w_{t+1} = w_t - \eta_t \nabla f(w_t)$ attains the rate $O(1/\sqrt{T})$ ; it follows that $\mathbb{E}\,[f(\bar w_T)] - f^{*} \le \tfrac{G^2 + \sigma^2}{2\sqrt{T}}$ whenever the step sizes satisfy $\textstyle\sum_t \eta_t^2 < \infty$ — the argument mirroring the classical Robbins–Monro analysis.

preprint — page 33

the reasoning, restored02 / 06

/prereqRobbins–Monro (1951) — assumed read, in learning order

03

/derive — the step between equation 4 and 5

The step is missing because the reviewer didn't need it. You do. Ritsu rebuilds it — then hands you the pivotal line to produce yourself, because that's the one you'll need in a viva.

On the Convergence of Stochastic Gradient Methods

K. Ito · R. Chen · A. Novak — Dept. of Statistics

Preprint · v3 · under review

Abstract

We establish a non-asymptotic $O(1/\sqrt{T})$ convergence rate for stochastic gradient descent under the classical Robbins–Monro step-size conditions, matching the known lower bound up to constants.

3 · Main Result

Assuming the standard smoothness and bounded-variance conditions, the update $w_{t+1} = w_t - \eta_t \nabla f(w_t)$ attains the rate $O(1/\sqrt{T})$ ; it follows that $\mathbb{E}\,[f(\bar w_T)] - f^{*} \le \tfrac{G^2 + \sigma^2}{2\sqrt{T}}$ whenever the step sizes satisfy $\textstyle\sum_t \eta_t^2 < \infty$ — the argument mirroring the classical Robbins–Monro analysis.

preprint — page 33

the reasoning, restored03 / 06

\mathbb{E}\,[f(w_t)] - f^{*} \le \; ?

your step — try it

04

/counter-example — where the claim breaks

You don't understand a result by restating it. You understand it by finding its edge — the case where it fails, the assumption carrying all the weight. Ritsu makes “what if n isn't large?” one word.

On the Convergence of Stochastic Gradient Methods

K. Ito · R. Chen · A. Novak — Dept. of Statistics

Preprint · v3 · under review

Abstract

We establish a non-asymptotic $O(1/\sqrt{T})$ convergence rate for stochastic gradient descent under the classical Robbins–Monro step-size conditions, matching the known lower bound up to constants.

3 · Main Result

Assuming the standard smoothness and bounded-variance conditions, the update $w_{t+1} = w_t - \eta_t \nabla f(w_t)$ attains the rate $O(1/\sqrt{T})$ ; it follows that $\mathbb{E}\,[f(\bar w_T)] - f^{*} \le \tfrac{G^2 + \sigma^2}{2\sqrt{T}}$ whenever the step sizes satisfy $\textstyle\sum_t \eta_t^2 < \infty$ — the argument mirroring the classical Robbins–Monro analysis.

preprint — page 33

the reasoning, restored04 / 06

/counter-examplewhat if

\textstyle\sum_t \eta_t^2

diverges? — the assumption carrying all the weight

05

/askme — say the contribution back

The real test: explain what this paper contributes, in one paragraph, in your own words. Ritsu tells you where your explanation broke — the way your advisor would.

On the Convergence of Stochastic Gradient Methods

K. Ito · R. Chen · A. Novak — Dept. of Statistics

Preprint · v3 · under review

Abstract

We establish a non-asymptotic $O(1/\sqrt{T})$ convergence rate for stochastic gradient descent under the classical Robbins–Monro step-size conditions, matching the known lower bound up to constants.

3 · Main Result

Assuming the standard smoothness and bounded-variance conditions, the update $w_{t+1} = w_t - \eta_t \nabla f(w_t)$ attains the rate $O(1/\sqrt{T})$ ; it follows that $\mathbb{E}\,[f(\bar w_T)] - f^{*} \le \tfrac{G^2 + \sigma^2}{2\sqrt{T}}$ whenever the step sizes satisfy $\textstyle\sum_t \eta_t^2 < \infty$ — the argument mirroring the classical Robbins–Monro analysis.

preprint — page 33

the reasoning, restored05 / 06

/askmeexplain the contribution — one paragraph, your own words

06

Keep it past the forgetting curve

One click puts each concept on a spaced-review schedule, so the paper you cracked today is still yours at your next lab meeting.

On the Convergence of Stochastic Gradient Methods

K. Ito · R. Chen · A. Novak — Dept. of Statistics

Preprint · v3 · under review

Abstract

We establish a non-asymptotic $O(1/\sqrt{T})$ convergence rate for stochastic gradient descent under the classical Robbins–Monro step-size conditions, matching the known lower bound up to constants.

3 · Main Result

Assuming the standard smoothness and bounded-variance conditions, the update $w_{t+1} = w_t - \eta_t \nabla f(w_t)$ attains the rate $O(1/\sqrt{T})$ ; it follows that $\mathbb{E}\,[f(\bar w_T)] - f^{*} \le \tfrac{G^2 + \sigma^2}{2\sqrt{T}}$ whenever the step sizes satisfy $\textstyle\sum_t \eta_t^2 < \infty$ — the argument mirroring the classical Robbins–Monro analysis.

preprint — page 33

the reasoning, restored06 / 06

review dueone click — back before your next lab meeting

Pedagogy, packaged

Every command packs a proven learning method.

Take one idea from the paper. Six commands take it from “I followed the words” to “I could defend it” — each one asking more of you than the last.

The convergence bound· one idea out of your paper

/explain01

Understand

The core idea first, then the layers — built from the paper you just dropped in.

/quiz02

Recall

Retrieval practice: the most proven way to make it stick. Pull it back out, cold.

/flashcard03

Drill

The parts that have to be automatic, turned into cards you'll actually run.

/askme04

Explain back

The Feynman test. Say it in your own words — that's where the gaps show.

/exercise05

Apply

A real problem on the real concept. Knowing it and using it aren't the same.

/write06

Create

Write it out in full. The highest bar there is: if you can write it, it's yours.

/counter-example·/what-if

The two moves off the ladder — and no chat box offers “show me where this breaks” as a first-class move, because nobody thinks to type it. Ritsu makes it one word.

That's Bloom's taxonomy, climbed one command at a time. And the science isn't a slogan — it's a flag on the command:

/quiz --bloom="auto"

40+ commands, all built like this.

Knowledge Map

You always know exactly where you stand.

A paper gives you no way to check whether you got it — until a lab meeting does, in public. Ritsu scores every concept as you work, so you find your gaps at your desk instead of across a table.

Paper · On the Convergence of Stochastic Gradient MethodsScore 58%

masterednot yetreview due

01

Every concept, scored.

Not one vague sense of “I think I get it” — a mastery score per idea in the paper.

02

Dim = what will catch you out.

The concepts you haven't got yet, named — before someone else names them for you.

03

The whole paper, one screen.

Contribution, method, and the claims it rests on — mapped.

Spaced review

Understand it today. Still have it at your next lab meeting.

You forget most of what you read within days — Ebbinghaus measured it in 1885. Ritsu brings each concept back right before it fades, so the paper you cracked this week is still there when someone asks you about it.

Five minutes tonight. Three tomorrow. Two in three days.
Built from your paper — no cards to make.

When you're stuck

The moments you actually get stuck — and what to type.

Then the paper probably assumed something you haven't seen. /prereq names what it took for granted — in learning order — so you fix the missing prerequisite, not the symptom.

On the Convergence of Stochastic Gradient Methods

K. Ito · R. Chen · A. Novak — Dept. of Statistics

Preprint · v3 · under review

Abstract

We establish a non-asymptotic $O(1/\sqrt{T})$ convergence rate for stochastic gradient descent under the classical Robbins–Monro step-size conditions, matching the known lower bound up to constants.

3 · Main Result

Assuming the standard smoothness and bounded-variance conditions, the update $w_{t+1} = w_t - \eta_t \nabla f(w_t)$ attains the rate $O(1/\sqrt{T})$ ; it follows that $\mathbb{E}\,[f(\bar w_T)] - f^{*} \le \tfrac{G^2 + \sigma^2}{2\sqrt{T}}$ whenever the step sizes satisfy $\textstyle\sum_t \eta_t^2 < \infty$ — the argument mirroring the classical Robbins–Monro analysis.

3.1 · Proof Sketch

Expanding the smoothness inequality and taking expectations, $\mathbb{E}\,\|w_{t+1} - w^{*}\|^{2} \le (1 - \mu\eta_t)\,\mathbb{E}\,\|w_t - w^{*}\|^{2} + \eta_t^{2}\sigma^{2}$ ; telescoping over t and dividing by the step-size sum gives the claim. The bound fails when the noise variance is unbounded — see Remark 2.

Figure 2: Suboptimality

f(\bar w_T) - f^{*}

vs. iterations under decaying step sizes.

preprint — page 33

every piece of reasoning, restored

Stop re-reading. Start understanding.

Get started for free

Free forever · no card

Understand any dense paper deeply enough to explain it

You're not failing the paper. It was written for someone else.

A good advisor doesn't summarize the paper. They interrogate it.

Drop in the PDF

/prereq — what it assumed you'd read

/derive — the step between equation 4 and 5

/counter-example — where the claim breaks

/askme — say the contribution back

Keep it past the forgetting curve

Every command packs a proven learning method.

You always know exactly where you stand.

Every concept, scored.

Dim = what will catch you out.

The whole paper, one screen.

Understand it today. Still have it at your next lab meeting.

The moments you actually get stuck — and what to type.

“I've read section 3 four times and the notation still won't resolve.”/prereq

“It jumps from equation 4 to equation 5 and I can't see how.”/derive

“I can restate the result. I don't actually understand it.”/counter-example

“Will it handle the math — LaTeX, matrices, proofs?”yes

“Can I just paste the arXiv link?”the PDF

Stop re-reading. Start understanding.

Understand any dense paper
deeply enough to explain it