A function is not differentiable at a point when its graph has no single, well-defined tangent line there. On a graph, you spot these points as sharp corners, cusps, vertical tangents, jump discontinuities, or removable holes. At any such x-value, the left-hand and right-hand slopes disagree, grow without bound, or fail to exist, so the derivative cannot be assigned a finite value.

Reading differentiability off a graph is a skill of pattern recognition backed by a precise limit definition. Most short explainers list the shapes and stop. This guide goes further: it explains why each shape breaks the derivative, quantifies how far apart the one-sided slopes can be, and separates the failure modes that students routinely confuse. You will also get thresholds for classroom accuracy and a realistic account of what a graph can and cannot tell you.

The secant-to-tangent mechanism behind differentiability

The derivative is a limit of secant slopes. For a point x = a, you compute [f(a + h) − f(a)] / h and let h shrink toward zero. Differentiability requires that this single limit exist as a finite real number. When you take the limit from the left (h < 0) and from the right (h > 0) separately, you get the left-hand and right-hand derivatives. The derivative exists only when both one-sided derivatives exist and are equal.

That equality condition is the engine behind every visual failure. A sharp corner is exactly a point where the two one-sided limits produce different finite numbers. A cusp is a point where they diverge to +∞ and −∞. A vertical tangent is a point where both diverge to the same infinity, so no finite slope exists. A jump discontinuity forces the numerator to approach a nonzero gap while h approaches zero, which drives the quotient to infinity.

A plain-language version: the derivative asks "what single slope best matches the curve right here?" If the left side and right side of the point demand different answers, there is no single answer to give.

The graph is a compressed picture of that limit process. A corner drawn on paper is the visible signature of unequal one-sided slopes. Once you internalize the limit definition, the shapes stop being a memorized list and become consequences you can predict.

Quantified context: how big the slope gap can get

The slope gap at a corner is not a technicality; it can be arbitrarily large. Consider f(x) = |x| at x = 0. The left-hand derivative is −1 and the right-hand derivative is +1, a gap of exactly 2. Now consider f(x) = |2x|. The gap doubles to 4. Scale the factor to k and the gap is 2k for any positive k. No finite ceiling exists on how steeply the two sides can disagree.

Steepness itself is not the problem. A line with slope 1,000 is perfectly differentiable everywhere. The failure at a corner comes from disagreement between sides, not from magnitude. This distinction matters because students often misread a steep graph as "not differentiable." A slope of 1,000 versus a slope of −1,000 at the same point is a genuine failure; a constant 1,000 on both sides is fine.

For vertical tangents, the slope magnitude grows without bound. The classic case is f(x) = x^(1/3) at x = 0, where the secant slopes approach +∞ from both sides. Larger and larger finite numbers never settle on a finite limit, so the derivative is undefined even though the function is continuous and the tangent line is visibly vertical.

Continuity is necessary but not sufficient. Differentiability implies continuity, so every differentiable point is continuous, but the converse fails at corners, cusps, and vertical tangents. In typical calculus curricula, roughly the first two to three weeks of derivative instruction are devoted to these exceptions, which signals how central they are to the concept.

A comparison table of non-differentiable graph features

The table below organizes the main visual signatures by their underlying limit behavior. Use it as a diagnostic key when scanning a graph.

Graph feature Continuity at the point One-sided slope behavior Derivative status Canonical example
Sharp corner Continuous Two different finite slopes Undefined f(x) = |x| at x = 0
Cusp Continuous +∞ on one side, −∞ on the other Undefined f(x) = x^(2/3) at x = 0
Vertical tangent Continuous Both sides approach +∞ or both −∞ Undefined (infinite slope) f(x) = x^(1/3) at x = 0
Jump discontinuity Discontinuous Quotient blows up to ±∞ Undefined Step function at the jump
Removable hole Discontinuous as written Both sides approach one finite value Undefined at the hole, definable by filling it (x² − 1)/(x − 1) at x = 1
Oscillation Continuous No limit at all Undefined x·sin(1/x) at x = 0
Smooth curve Continuous One shared finite slope Defined f(x) = x² everywhere

The table reveals a useful hierarchy. The first three rows are continuous failures; the next two involve discontinuity; the oscillation row is a stranger case where the function wiggles infinitely often near the point. Recognizing which row a feature belongs to tells you the mechanism before you compute anything.

Edge cases that graphs hide or distort

Graphs are finite-resolution pictures, and some failures are invisible at normal zoom. The function x·sin(1/x) oscillates with increasing frequency as x approaches 0, and its secant slopes have no limit. A plotted graph may look like a blur near the origin, and a coarse plot can even suggest the curve flattens out. The derivative still fails to exist because the limit never settles.

Another edge case is the function that is continuous everywhere but differentiable nowhere. The Weierstrass function is the standard example: it is a sum of infinitely many scaled cosine waves and has a corner-like irregularity at every point. Plotted at ordinary resolution, it looks like fractal noise. No finite sampling of a graph can certify differentiability; the graph can only suggest it or reveal a clear failure.

A subtler trap is a point that looks like a corner but is actually smooth at higher zoom. Some curves have very high curvature over a tiny interval, which reads as a kink on a standard plot. Zooming in resolves a smooth curve with a well-defined tangent. The professional caveat is blunt: a printed graph is evidence, not proof. Only the limit computation is decisive.

Piecewise functions add another layer. A piecewise graph can be continuous at the join yet have mismatched slopes, producing a corner exactly at the boundary. The function f(x) = x for x ≤ 0 and f(x) = 2x for x > 0 is continuous at 0 but has left-hand slope 1 and right-hand slope 2. The corner is real, and the derivative at 0 is undefined even though both pieces are individually smooth.

One more edge case deserves mention: endpoints of a closed interval. At an endpoint, only a one-sided derivative is available, so differentiability in the two-sided sense is not defined there. Many textbooks handle this by defining a one-sided derivative at endpoints, which is a convention rather than a contradiction.

Misconceptions that trip up graph readers

The most common misconception is that continuity guarantees differentiability. It does not. The absolute value function is continuous at 0 and fails to be differentiable there. The correct relationship runs one way only: differentiability implies continuity, never the reverse.

A second misconception is that "the graph has no break, so it must be smooth." Smoothness is stronger than continuity. A curve can be unbroken and still carry corners or cusps. The visual test is not "is there a gap?" but "does a single tangent line fit the point from both sides?"

A third misconception confuses vertical asymptotes with vertical tangents. A vertical asymptote occurs where the function itself diverges, so the point is not even in the domain. A vertical tangent occurs where the function is finite but the slope diverges. The two look similar on a stretched plot but are different phenomena with different limit behavior.

A fourth error is assuming a cusp and a corner are the same. Both are non-differentiable continuous points, but the cusp has infinite slopes of opposite sign, while the corner has finite slopes that simply disagree. The corrective lens is the one-sided derivative: compute both sides and compare magnitudes as well as signs.

A fifth misconception is that a removable hole makes the function differentiable once you "fill it in." Filling the hole changes the function's definition at that point, which may restore continuity, but differentiability at the filled point still requires checking the one-sided slopes of the new definition. Continuity alone is not enough.

Red flags to watch for when reading a graph: A sharp direction change with no rounding, which signals a corner. Two branches meeting in a spike, which signals a cusp. A curve that flattens into a vertical line, which signals a vertical tangent. A jump between two levels, which signals a discontinuity and automatic non-differentiability. A tiny oscillation or blur near a point, which may hide an oscillation failure.

Practical actions for reading differentiability from a graph

Start by locating every candidate point: corners, cusps, vertical tangents, jumps, holes, and anything that looks blurred at normal zoom. Mark each one before doing any algebra. This scan takes under a minute and prevents the common error of missing a subtle feature.

Next, test continuity at each candidate. If the graph has a gap, jump, or hole, the function is discontinuous and therefore not differentiable there. This single check eliminates the discontinuity rows of the table immediately.

For continuous candidates, estimate the one-sided slopes. Pick a point just left and just right of the feature, draw the secant lines, and compare. If the two slopes differ by a visible margin, you have a corner. If one side points steeply up and the other steeply down, you have a cusp. If both sides steepen toward vertical, you have a vertical tangent.

Then confirm with the limit definition. Compute [f(a + h) − f(a)] / h from both sides and compare. This is the only rigorous step; the graph work is a guide, not a verdict. For piecewise functions, evaluate each piece's derivative at the join and compare the results.

Finally, zoom aggressively on any feature that looked smooth at first pass. A plot with resolution near the feature's scale can expose a hidden kink. If the zoom reveals a smooth curve, you have likely found finite curvature rather than a true corner.

Limitations of graph-based differentiability judgments

A graph can show you where a function clearly fails, but it cannot prove where it succeeds. Finite pixel resolution, axis scaling, and plotting artifacts all distort small-scale behavior. A feature that looks differentiable on a 500-pixel plot may reveal a corner at 10,000-pixel resolution, and a genuine oscillation failure may be invisible at any practical resolution.

The limit definition remains the authority. Treat graphs as diagnostic aids that generate hypotheses, then verify with one-sided derivatives. This division of labor is exactly how working mathematicians use graphs: for intuition and error-catching, not for proof.

For students preparing for exams, the practical threshold is consistency. Use the graph scan to identify candidates, then always confirm with the limit. A correct answer names the mechanism, not just the shape. If you can say "the left-hand slope is −1 and the right-hand slope is +1, so the limit fails," you have demonstrated understanding rather than memorization.

One honest caveat about the graph itself: a plotted curve is a sample of points, and differentiability is a property of the function at a point, not of the ink on the page. The ink is a map; the limit is the territory.