Sudoku is a logic puzzle on a 9×9 grid of 81 cells, split by thicker lines into nine 3×3 boxes, in which you place the digits 1 through 9 so that every row, every column and every box contains each digit exactly once. Nothing about that rule set involves arithmetic — you never add, subtract or multiply anything — which is why it belongs under logic and deduction rather than mathematics, and why the same grid supports puzzles from trivially easy to genuinely brutal without changing a single rule. The rest of this page is about how those puzzles actually work, where the difficulty comes from, and how to get better at them without wasting months on the wrong method.
What the rules actually constrain, and why that matters
Each of the 81 cells sits at the intersection of exactly three constraints: its row, its column, and its 3×3 box. That triple membership is the whole engine of the puzzle. A digit placed in a cell eliminates that digit as a candidate from eight other cells in the row, eight in the column, and four in the box (the box overlap accounts for the rest), which is why a single correct placement can cascade into several forced ones.
The practical consequence is that you should almost never be guessing. In a well-formed puzzle there is a logical path from the givens to the full grid, and every step along that path is a deduction you can name: this cell must be a 7 because the other eight digits already appear in its row, column or box; or this cell cannot be a 4 because every position for a 4 in this box has been ruled out. If you find yourself placing a digit because it "feels right," you have left the puzzle and started a search tree, and you will pay for it later when the contradiction surfaces twenty cells away and you cannot remember which assumption caused it.
The trade-off is speed. Scanning for forced placements is slower per cell than intuition, and beginners who trust pattern recognition often finish easier puzzles faster than disciplined solvers. That advantage evaporates somewhere around the middle of the difficulty curve, where the patterns stop being visual and start requiring you to hold candidate lists in your head. The earlier you build the candidate-list habit, the less painful that transition is.
One limit worth stating plainly: the constraint structure tells you nothing about whether a puzzle has a unique solution. A grid can satisfy every row, column and box rule and still admit multiple valid completions. Reputable publishers guarantee uniqueness; a puzzle pulled at random from the internet may not, and if you find two valid completions you have not made a mistake — the puzzle was simply under-constrained.
How difficulty is actually manufactured
Difficulty in Sudoku is not a function of how many digits are given. A puzzle with a generous handful of givens can be harder than one with many more, because what matters is which solving techniques the givens force you to use and in what order. A grid that falls entirely to naked singles and hidden singles — a digit that is the only possible value for a cell, or the only cell in a unit that can hold a digit — is easy regardless of how sparse it looks. A grid that requires you to spot an X-Wing or a swordfish before anything unlocks is hard regardless of how full it appears.
This is why the difficulty labels on published puzzles are so inconsistent. A newspaper's "hard" and a specialist book's "hard" are calibrated against different solver populations, and neither is a measurement. The only reliable calibration is your own: if you can finish a puzzle without erasing a candidate you wrote down, it was easy for you, and that is the number that matters.
There is a real cost to chasing harder puzzles too early. Techniques like forcing chains and uniqueness arguments require you to hold a large amount of grid state in working memory, and if your single-and-pair scanning is not yet automatic, the harder technique will keep failing not because you do not understand it but because you keep losing track of the board. Most solvers who stall do so for this reason rather than for lack of technique knowledge.
If you want a concrete sequence, the conventional ladder runs: naked and hidden singles, then locked candidates (pointing and claiming), then naked and hidden pairs and triples, then X-Wings and other single-digit patterns, then the more elaborate fish and chain techniques. Move up only when the current rung is boring rather than difficult.
Pencil marks, and when to stop using them
Writing candidate digits in the corner of each cell is the single highest-return habit in Sudoku, and it is also the habit most likely to become a crutch. The case for it is straightforward: a puzzle that requires anything beyond singles cannot be solved reliably by holding candidates in your head, because the number of simultaneous constraints exceeds what most people can track without error. Pencil marks externalise that state so your attention can go to the reasoning.
The case against is that full candidate notation on all 81 cells is a lot of writing, and much of it is wasted. A better practice is selective notation: mark candidates only in cells that have been narrowed to two or three possibilities, and only in the units you are currently working. This keeps the grid readable and stops you from drowning in tiny numbers.
The transition point away from notation is real but later than most people think. Once you can see a naked pair without writing it down, stop writing pairs. Once you can scan a row for a single missing digit without counting on your fingers, stop counting. The general rule is to remove the scaffold only for the step you have already automated, never for the step above it.
Digital solving changes the calculus. In an app, candidates cost a tap rather than a pencil stroke, so there is almost no reason not to fill them — but the apps that auto-fill candidates for you are doing the most valuable part of the work, and puzzles solved that way build noticeably less skill. If you are using an app to improve, turn auto-candidates off.
Paper, app or print: what each costs you
Paper is the best medium for learning and the worst for hard puzzles. The cost of an error is an eraser smudge rather than an undo, which forces a discipline that transfers to everything else, and the physical grid supports the kind of peripheral scanning that screens interfere with. But once you are writing three or four candidates per cell, paper runs out of space, and a single mistaken elimination is expensive to trace.
Apps solve the space problem and introduce a worse one: they make it easy to check, hint, and undo, and every one of those is a small loan against your own reasoning. An app that tells you when a digit is wrong removes the feedback loop that teaches you to be careful. The best digital setup is a bare grid with candidate entry and nothing else — no error checking, no hints, no timer pressure unless you want it.
Print sits between the two and is underrated for hard puzzles. A large-print grid with generous cell size gives you room for candidates and the tactile feedback of paper, and printing a batch of puzzles at a known difficulty is the cheapest way to work through a progression systematically. The cost is that you cannot easily redo a puzzle you botched, which is a genuine loss — re-solving a puzzle you failed, once you know the technique you missed, is one of the fastest ways to consolidate it.
None of this matters as much as consistency. A puzzle a day on paper beats an occasional marathon in an app, because the scanning patterns you are building are perceptual and they decay without use.
What the evidence on puzzles and cognition does and does not show
The research on number puzzles and brain function in older adults is more modest than the marketing around it. A large study in this area assessed 14 cognitive measures in adults aged 50 to 93, which tells you the researchers were looking broadly — not just at "memory" but at a spread of specific abilities — and that the population was old enough for cognitive change to be a live question. That design is a strength: broad measurement in a relevant age band is exactly what you would want before drawing conclusions.
What such studies generally cannot establish is causation in the direction people hope for. People who are already cognitively sharper may be more likely to take up and persist with puzzles, which produces a correlation without puzzles doing the work. And improvement on the puzzle task itself is the least interesting outcome — getting better at Sudoku is guaranteed by practice and says little about anything else.
The honest position is that puzzle practice reliably improves puzzle performance and probably supports the specific habits it trains: sustained attention, working memory for candidate state, and systematic search. Whether it transfers to memory in daily life is not settled by the evidence available, and anyone telling you a daily grid is protective against cognitive decline is going beyond what the research supports. Do it because you enjoy it and because the attention it trains is useful; treat cognitive benefits as a plausible bonus rather than a reason.
There is a real cost to the framing either way. If you solve because you believe it is medicine, a plateau becomes a failure and you quit. If you solve because the deduction is satisfying, a plateau is just the point where you need a new technique.
Getting unstuck without looking up the answer
When a grid stalls, the productive move is almost always to re-scan systematically rather than to stare harder at the same region. Work digit by digit: take the 1 and sweep every row, column and box asking where a 1 can still go, then move to the 2. This catches hidden singles that unit-by-unit scanning misses, because your attention is organised around the digit rather than the region. Most stalls at the easy and medium level dissolve on the second pass.
If the digit sweep fails, switch to units: for each box, list which digits are still missing and check each against the rows and columns crossing it. Then do the same for rows and columns. The two scans are complementary — digit-first finds placements, unit-first finds eliminations — and alternating between them is more effective than repeating either.
Only after both scans come up empty should you reach for a named technique, and when you do, apply it to one candidate at a time and write down what you eliminated. The failure mode at this stage is a chain of eliminations you cannot reconstruct, followed by an incorrect placement and a slow, demoralising unwinding. If you cannot say out loud why a candidate is gone, you do not yet know that it is.
Looking up the answer is a legitimate last resort but a poor learning tool. Looking up the technique — reading what an X-Wing is, then finding one in your stalled grid — is the opposite: it is the fastest way to convert a puzzle you cannot finish into a skill you keep. The distinction is whether you are retrieving a fact or acquiring a method.
Finally, keep a small record of which puzzles beat you and why. Not the grid, just the technique: "stalled on locked candidates," "missed a hidden pair in column six." After a few weeks the pattern is obvious, and it usually points at one weak rung on the ladder rather than at general difficulty. That is a far more useful signal than any publisher's difficulty label.
Why the 3×3 box is the constraint that does the heavy lifting
Of the three units a cell belongs to, the box is the one solvers underuse. Rows and columns are easy to scan because they run edge to edge; the box forces you to look at a cluster and ask which of the digits 1 through 9 are still missing from it, then test each against the rows and columns that cross it. That crossing is where the eliminations come from — a digit that is already committed somewhere in a crossing row cannot sit in the box's cells along that row, and three or four of those together often leave a single square. Solvers who plateau at medium difficulty almost always have strong row and column habits and a weak box habit.
The diagonal variant changes which techniques pay off
Sudoku X adds two constraints: both main diagonals must also contain 1 through 9 exactly once. The grid is still 81 cells in nine 3×3 boxes, so nothing you know stops working — but the five cells at the centre of the grid now sit under five simultaneous constraints instead of three, and they become the most informative squares on the board. Scan them first. The cost is that single-digit patterns like X-Wings lose some of their reach, because a digit's placement now has to survive a diagonal it may not intersect, and puzzles that look sparse can be unexpectedly stubborn.





